The global energy is increasingly tight, and the global temperature is gradually rising. Energy efficiency assessment and carbon emission accounting can provide theoretical tools and practical support for the formulation of energy conservation and emission reduction strategies for the food cold chain, and is also a prerequisite for the sustainable development of the food cold chain. In this paper, the relationship and differences between energy consumption and carbon emissions in the general food cold chain are first described, and the principle, advantages and disadvantages of three energy consumption conversion standards of solar emergy value, standard coal and equivalent electricity are discussed. Besides, the possibilities of applying these three energy consumption conversion standards to energy consumption analysis and energy efficiency evaluation of food cold chain are explored. Then, for a batch of fresh agricultural products, the energy consumption of six links of the food cold chain, including the first transportation, the manufacturer, the second transportation, the distribution center, the third transportation, and the retailer, are systematically and comprehensively analyzed from the product level, and the comprehensive energy consumption level of the food cold chain are obtained. On this basis, ten energy efficiency indicators from five aspects of macro energy efficiency are proposed, including micro energy efficiency, energy economy, environmental energy efficiency and comprehensive energy efficiency, and constructs the energy efficiency evaluation index system of food cold chain. At the same time, other energy efficiency evaluation indicators and methods are also summarized. In addition, the standard of carbon emission conversion of food cold chain, namely carbon dioxide equivalent is introduce, the boundary of carbon emission accounting is determined, and the carbon emission factors of China's electricity is mainly discussed. Furthermore, the origin, principle, advantages and disadvantages of the emission factor method, the life cycle assessment method, the input-output analysis method and the hybrid life cycle assessment method, and the basic process of life cycle assessment method in the calculation of food cold chain carbon footprint are also reviewed. In order to improve the energy efficiency level of the food cold chain and reduce the carbon emissions of each link of the food cold chain, energy conservation and emission reduction methods for food cold chain are proposed from five aspects: refrigerant, distribution path, energy, phase change cool storage technology and digital twin technology. Finally, the energy efficiency assessment and carbon emission accounting of the food cold chain are briefly prospected in order to provide reference for promoting the sustainable development of China's food cold chain.
The new generation of information technology has led to the rapid development of the intelligent level of the cold chain, and the precise control of the development level of the smart cold chain is the prerequisite foundation and guarantee to achieve the key breakthrough of the technical bottleneck and the strategic layout of the development direction. Based on this, an evaluation index system for China's intelligent cold chain development from the dimensions of supply capacity, storage capacity, transportation capacity, economic efficiency and informationization level was conducted. The entropy weight method combined with the technique for order preference by similarity to ideal solution (TOPSIS) was used to quantitatively evaluate the development of intelligent cold chain in 30 Chinese provinces and cities (excluding Tibet, Hong Kong, Macao and Taiwan) from 2017 to 2021. The quantitative evaluation of the level of intelligent cold chain development was conducted. The impact of the evaluation indicators on different provinces and cities was analysed by exploratory spatial data analyses (ESDA) and geographically weighted regression (GWR). The results showed that indicators such as economic development status, construction of supporting facilities and informationization level had greater weight and played a more important role in influencing the construction of intelligent cold chain. The overall level of intelligent cold chain development in China is divided into four levels, with most cities at the third and fourth levels. Beijing and the eastern coastal provinces and cities generally have a better level of intelligent cold chain development, while the southwest and northwest regions are developing slowly. In terms of overall development, the overall development of China's intelligent cold chain is relatively backward, with insufficient inter-regional synergy. The global spatial autocorrelation analysis shows that the variability in the development of China's intelligent cold chain logistics is gradually becoming greater. Through the local spatial autocorrelation analysis, it can be seen that there is a positive spatial correlation between the provinces and cities in East China, and negative spatiality in North China and South China. After geographically weighted regression analysis, it can be seen that the evaluation indicators have significant spatial and temporal heterogeneity in 2017, with the degree of influence changing with spatial location and time, and the spatial and temporal heterogeneity of the evaluation indicators is not significant in 2021. In order to improve the overall development level of China's intelligent cold chain, corresponding development countermeasures are proposed to strengthen the construction of supporting facilities and promote the transformation and upgrading of information technology. This study can provide a scientific basis for the global planning, strategic layout and overall promotion of China's intelligent cold chain.
Ensuring the stability of agricultural products logistics is the key to ensuring people's livelihood. The forecast of agricultural products logistics demand is an important guarantee for rational planning of agricultural products logistics stability. However, the forecasting of agricultural products logistics demand is actually complicated, and it will be affected by various factors in the forecasting process. Therefore, in order to ensure the accuracy of forecasting the logistics demand of agricultural products, many influencing factors need to be considered. In this study, the logistics demand of agricultural products is taken as the research object, relevant indicators from 2017 to 2021 were selected as characteristic independent variables and a neural network model for forecasting the logistics demand of agricultural products was constructed by using Informer neural network. Taking Henan province, Hubei province and Hunan province in Central China as examples, the logistics demands of agricultural products in the three provinces were predicted. At the same time, long short-term memory network (LSTM) and Transformer neural network were used to forecast the demand of agricultural products logistics in three provinces of Central China, and the prediction results of the three models were compared. The results showed that the average percentage of prediction test error based on Informer neural network model constructed in this study was 3.39%, which was lower than that of LSTM and Transformer neural network models of 4.43% and 4.35%. The predicted value of Informer neural network model for three provinces was close to the actual value. The predicted value of Henan province in 2021 was 4185.33, the actual value was 4048.10, and the error was 3.389%. The predicted value of Hubei province in 2021 was 2503.64, the actual value was 2421.78, and the error was 3.380%. The predicted value of Hunan province in 2021 was 2933.31, the actual value was 2836.86, and the error was 3.340%. Therefore, it showed that the model can accurately predict the demand of agricultural products logistics in three provinces of Central China, and can provide a basis for rational planning and policy making of agricultural products logistics. Finally, the model and parameters were used to predict the logistics demand of agricultural products in Henan, Hunan, and Hubei provinces in 2023, and the predicted value of Henan province in 2023 was 4217.13; Hubei province was 2521.47, and Hunan province was 2974.65, respectively. The predicted values for the three provinces in 2023 are higher than the predicted values in 2021. Therefore, based on the logistics and transportation supporting facilities in 2021, it is necessary to ensure logistics and transportation efficiency and strengthen logistics and transportation capacity, so as to meet the growing logistics demand in Central China.
In recent years, China's cold chain logistics industry has entered a stage of rapid development. At the same time, with the increase of greenhouse gas emissions, green and low-carbon transformation has become a new feature and direction of high-quality and healthy development of the cold chain industry to meet the future development needs of China's low-carbon economy. In view of this, in order to ensure the scientificity of China's low-carbon cold chain logistics evaluation system, in this paper, 30 indicators from the four levels of energy transformation, technological innovation, economic efficiency, and national policy based on different relevant levels were first preliminarily determined, and finally 14 indicators for building China's low-carbon cold chain logistics development evaluation system through consulting experts and the possibility of data acquisition were determined. Data from 2017 to 2021 were selected to conduct a quantitative evaluation of the development level of low-carbon cold chain logistics in China. Firstly, the entropy weight method was used to analyze the weight and obstacle degree of different indicators to explore the impact of different indicators on the development of low-carbon cold chain logistics; Secondly, a weighted decision-making matrix was constructed based on the weights of different indicators, and the technology for order preference by similarity to ideal solution (TOPSIS) evaluation model was used to evaluate the development of low-carbon cold chain logistics in China from 2017 to 2021, in order to determine the development and changes of low-carbon cold chain logistics in China. The research results showed that among the 14 different indicators of the established evaluation system for the development of low-carbon cold chain logistics in China, the growth rate of the use of green packaging materials, the number of low-carbon technical papers published, the proportion of scientific research personnel, the growth rate of cold chain logistics demand for fresh agricultural products, and the reduction rate of hydrochlorofluorocarbon refrigerants account for a relatively large proportion, ranking in the top five, respectively reaching 0.1243, 0.1074, 0.1066, 0.0982, and 0.0716, accounting for more than half of the overall proportion. It has a significant impact on the development of low-carbon cold chain logistics in China. From 2017 to 2021, the development level of China's low-carbon cold chain logistics was scored from 0.1498 to 0.2359, with a year-on-year increase of about 57.5%, indicating that China's low-carbon cold chain logistics development level was relatively fast in the past five years. Although China's low-carbon cold chain logistics development has shown an overall upward trend, it is still in the development stage.
Accurate and efficient monitoring of animal information, timely analysis of animal physiological and physical health conditions, and automatic feeding and farming management combined with intelligent technologies are of great significance for large-scale livestock farming. Deep learning techniques, with automatic feature extraction and powerful image representation capabilities, solve many visual challenges, and are more suitable for application in monitoring animal information in complex livestock farming environments. In order to further analyze the research and application of artificial intelligence technology in intelligent animal farming, this paper presents the current state of research on deep learning techniques for tag detection recognition, body condition evaluation and weight estimation, and behavior recognition and quantitative analysis for cattle, sheep and pigs. Among them, target detection and recognition is conducive to the construction of electronic archives of individual animals, on which basis the body condition and weight information, behavior information and health status of animals can be related, which is also the trend of intelligent animal farming. At present, intelligent animal farming still faces many problems and challenges, such as the existence of multiple perspectives, multi-scale, multiple scenarios and even small sample size of a certain behavior in data samples, which greatly increases the detection difficulty and the generalization of intelligent technology application. In addition, animal breeding and animal habits are a long-term process. How to accurately monitor the animal health information in real time and effectively feed it back to the producer is also a technical difficulty. According to the actual feeding and management needs of animal farming, the development of intelligent animal farming is prospected and put forward. First, enrich the samples and build a multi perspective dataset, and combine semi supervised or small sample learning methods to improve the generalization ability of in-depth learning models, so as to realize the perception and analysis of the animal's physical environment. Secondly, the unified cooperation and harmonious development of human, intelligent equipment and breeding animals will improve the breeding efficiency and management level as a whole. Third, the deep integration of big data, deep learning technology and animal farming will greatly promote the development of intelligent animal farming. Last, research on the interpretability and security of artificial intelligence technology represented by deep learning model in the breeding field. And other development suggestions to further promote intelligent animal farming. Aiming at the progress of research application of deep learning in livestock smart farming, it provides reference for the modernization and intelligent development of livestock farming.
Enhancing resource use efficiency in agricultural field management and breeding high-performance crop varieties are crucial approaches for securing crop yield and mitigating negative environmental impact of crop production. Crop stress sensing and plant phenotyping systems are integral to variable-rate (VR) field management and high-throughput plant phenotyping (HTPP), with both sharing similarities in hardware and data processing techniques. Crop stress sensing systems for VR field management have been studied for decades, aiming to establish more sustainable management practices. Concurrently, significant advancements in HTPP system development have provided a technological foundation for reducing conventional phenotyping costs. In this paper, we present a systematic review of crop stress sensing systems employed in VR field management, followed by an introduction to the sensors and data pipelines commonly used in field HTPP systems. State-of-the-art sensing and decision-making methodologies for irrigation scheduling, nitrogen application, and pesticide spraying are categorized based on the degree of modern sensor and model integration. We highlight the data processing pipelines of three ground-based field HTPP systems developed at the University of Nebraska-Lincoln. Furthermore, we discuss current challenges and propose potential solutions for field HTPP research. Recent progress in artificial intelligence, robotic platforms, and innovative instruments is expected to significantly enhance system performance, encouraging broader adoption by breeders. Direct quantification of major plant physiological processes may represent one of next research frontiers in field HTPP, offering valuable phenotypic data for crop breeding under increasingly unpredictable weather conditions. This review can offer a distinct perspective, benefiting both research communities in a novel manner.
Agriculture is a basic industry deeply related to the national economy and people's livelihood, while it is also a weak industry. There are some problems with traditional agricultural risk management research methods, such as insufficient mining of nonlinear information, low accuracy and poor robustness. Artificial intelligence(AI) has powerful functions such as strong nonlinear fitting, end-to-end modeling, feature self-learning based on big data, which can solve the above problems well. The research progress of artificial intelligence technology in agricultural vulnerability assessment, agricultural risk prediction and agricultural damage assessment were first analyzed in this paper, and the following conclusions were obtained: 1. The feature importance assessment of AI in agricultural vulnerability assessment lacks scientific and effective verification indicators, and the application method makes it impossible to compare the advantages and disadvantages of multiple AI models. Therefore, it is suggested to use subjective and objective methods for evaluation; 2. In risk prediction, it is found that with the increase of prediction time, the prediction ability of machine learning model tends to decline. Overfitting is a common problem in risk prediction, and there are few researches on the mining of spatial information of graph data; 3. Complex agricultural production environment and varied application scenarios are important factors affecting the accuracy of damage assessment. Improving the feature extraction ability and robustness of deep learning models is a key and difficult issue to be overcome in future technological development. Then, in view of the performance improvement problem and small sample problem existing in the application process of AI technology, corresponding solutions were put forward. For the performance improvement problem, according to the user's familiarity with artificial intelligence, a variety of model comparison method, model group method and neural network structure optimization method can be used respectively to improve the performance of the model; For the problem of small samples, data augmentation, GAN (Generative Adversarial Network) and transfer learning can often be combined to increase the amount of input data of the model, enhance the robustness of the model, accelerate the training speed of the model and improve the accuracy of model recognition. Finally, the applications of AI in agricultural risk management were prospected: In the future, AI algorithm could be considered in the construction of agricultural vulnerability curve; In view of the relationship between upstream and downstream of agricultural industry chain and agriculture-related industries, the graph neural network can be used more in the future to further study the agricultural price risk prediction; In the modeling process of future damage assessment, more professional knowledge related to the assessment target can be introduced to enhance the feature learning of the target, and expanding the small sample data is also the key subject of future research.
Waterfowl farming in China is developing rapidly in the direction of large-scale, standardization and intelligence. The research and application of intelligent farming equipment and information technology is the key to promote the healthy and sustainable development of waterfowl farming, which is important to improve the output efficiency of waterfowl farming, reduce the reliance on labor in the production process, fit the development concept of green and environmental protection and achieve high-quality transformational development. In this paper, the latest research and inventions of intelligent waterfowl equipment, waterfowl shed environment intelligent control technology and intelligent waterfowl feeding, drinking water, dosing and disinfection and automatic manure treatment equipment were introduced. At present, compared to pigs, chickens and cattle, the intelligent equipment of waterfowl are still relatively backward. Most waterfowl houses are equipped with chicken equipment directly, lacking improvements for waterfowl. Moreover, the linkage between the equipment is poor and not integrated with the breeding mode and shed structure of waterfowl, resulting in low utilization. Therefore, there is a need to develop and improve equipment for the physiological growth characteristics of waterfowl from the perspective of their breeding welfare. In addition, the latest research advances in the application of real-time production information collection and intelligent management technologies were present. The information collection technologies included visual imaging technology, sound capture systems, and wearable sensors were present. Since the researches of ducks and geese is few, the research of poultry field, which can provide a reference for the waterfowl were also summarized. The research of information perception and processing of waterfowl is currently in its initial stage. Information collection techniques need to be further tailored to the physiological growth characteristics of waterfowl, and better deep learning models need to be established. The waterfowl management platform, taking the intelligent management platform developed by South China Agricultural University as an example were also described. Finally, the intelligent application of the waterfowl industry was pointed out, and the future trends of intelligent farming with the development of mechanized and intelligent equipment for waterfowl in China to improve the recommendations were analyzed. The current waterfowl farming is in urgent need of intelligent equipment reform and upgrading of the industry for support. In the future, intelligent equipment for waterfowl, information perception methods and control platforms are in urgent to be developed. When upgrading the industry, it is necessary to develop a development strategy that fits the current waterfowl farming model in China.
To deal with the issues of low accuracy of apple leaf disease images recognition under weak supervision with only image category labeling, an improved CBAM-ResNet-based algorithm was proposed in this research. Using ResNet18 as the base model, the multilayer perceptron (MLP) in the lightweight convolutional block attention module (CBAM) attention mechanism channel was improved by up-dimensioning to amplify the details of apple leaf disease features. The improved CBAM attention module was incorporated into the residual module to enhance the key details of AlphaDropout with SeLU (Scaled Exponential Linearunits) to prevent overfitting of its network and accelerate the convergence effect of the model. Finally, the learning rate was adjusted using a single-cycle cosine annealing algorithm to obtain the disease recognition model. The training test was performed under weak supervision with only image-level annotation of all sample images, which greatly reduced the annotation cost. Through ablation experiments, the best dimensional improvement of MLP in CBAM was explored as 2. Compared with the original CBAM, the accuracy rate was increased by 0.32%, and the training time of each round was reduced by 8 s when the number of parameters increased by 17.59%. Tests were conducted on a dataset of 6185 images containing five diseases, including apple spotted leaf drop, brown spot, mosaic, gray spot, and rust, and the results showed that the model achieved an average recognition accuracy of 98.44% for the five apple diseases under weakly supervised learning. The improved CBAM-ResNet18 had increased by 1.47% compared with the pre-improved ResNet18, and was higher than VGG16, DesNet121, ResNet50, ResNeXt50, EfficientNet-B0 and Xception control model. In terms of learning efficiency, the improved CBAM-ResNet18 compared to ResNet18 reduced the training time of each round by 6 s under the condition that the number of parameters increased by 24.9%, and completed model training at the fastest speed of 137 s per round in VGG16, DesNet121, ResNet50, ResNeXt50, Efficient Net-B0 and Xception control models. Through the results of the confusion matrix, the average precision, average recall rate, and average F1 score of the model were calculated to reach 98.43%, 98.46%, and 0.9845, respectively. The results showed that the proposed improved CBAM-ResNet18 model could perform apple leaf disease identification and had good identification results, and could provide technical support for intelligent apple leaf disease identification providing.
Pre-training refers to the process of training deep neural network parameters on a large corpus before a specific task model performs a particular task. This approach enables downstream tasks to fine-tune the pre-trained model parameters based on a small amount of labeled data, eliminating the need to train a new model from scratch. Currently, research on named entity recognition (NER) using pre-trained language model (PLM) only uses the last layer of the PLM to express output when facing challenges such as complex entity naming methods and fuzzy entity boundaries in the agricultural field. This approach ignores the rich information contained in the internal layers of the model themselves. To address these issues, a named entity recognition method based on progressive convolutional networks has been proposed. This method stores natural sentences and outputs representations of each layer obtained through PLM. The intermediate outputs of the pre-trained model are sequentially convolved to extract shallow feature information that may have been overlooked previously. Using the progressive convolutional network module proposed in this research, the adjacent two-layer representations are convolved from the first layer, and the fusion result continues to be convolved with the next layer, resulting in enhanced sentence embedding that includes the entire information dimension of the model layer. The method does not require the introduction of external information, which makes the sentence representation contain richer information. Research has shown that the sentence embedding output of the model layer near the input contains more fine-grained information, such as phrases and phrases, which can assist with NER problems in the agricultural field. Fully utilizing the computational power already used, the results obtained can enhance the representation embedding of sentences. Finally, the conditional random field (CRF) model was used to generate the global optimal sequence. On a constructed agricultural dataset containing four types of agricultural entities, the proposed method's comprehensive indicator F1 value increased by 3.61% points compared to the basic BERT (Bidirectional Encoder Representation from Transformers) model. On the open dataset MSRA, the F1 value also increased to 94.96%, indicating that the progressive convolutional network can enhance the model's ability to represent natural language and has advantages in NER tasks.
Crops have diverse structures and complex growth environments. RGB image data can reflect the texture and color features of plants accurately, while 3D data contains information about crop volume. The combination of RGB image and 3D point cloud data can achieve the extraction of two-dimensional and three-dimensional phenotypic parameters of crops, which is of great significance for the research of phenomics methods. In this study, potatoe plants were chosen as the research subject, and RGB cameras and laser scanners were used to collect 50 potato RGB images and 3D laser point cloud data. The segmentation accuracy of four deep learning semantic segmentation methods, OCRNet, UpNet, PaNet, and DeepLab v3+, were compared and analyzed for the RGB images. OCRNet, which demonstrated higher accuracy, was used to perform semantic segmentation on top-view RGB images of potatoes. Mean shift clustering algorithm was optimized for laser point cloud data processing, and single-plant segmentation of laser point cloud data was completed. Stem and leaf segmentation of single-plant potato point cloud data were accurately performed using Euclidean clustering and K-Means clustering algorithms. In addition, a strategy was proposed to establish a one-to-one correspondence between RGB images and point clouds of single-plant potatoes using pot numbering. 8 2D phenotypic parameters and 10 3D phenotypic parameters, including maximum width, perimeter, area, plant height, volume, leaf length, and leaf width, etc., were extracted from RGB images and laser point clouds, respectively. Finally, the accuracy of three representative and easily measurable phenotypic parameters, leaf number, plant height, and maximum width were evaluated. The mean absolute percentage errors (MAPE) were 8.6%, 8.3% and 6.0%, respectively, while the root mean square errors (RMSE) were 1.371 pieces, 3.2 cm and 1.86 cm, respectively, and the determination coefficients (R2) were 0.93, 0.95 and 0.91, respectively. The research results indicated that the extracted phenotype parameters can accurately and efficiently reflect the growth status of potatoes. Combining the RGB image data of potatoes with three-dimensional laser point cloud data can fully exploit the advantages of the rich texture and color characteristics of RGB images and the volumetric information provided by three-dimensional point clouds, achieving non-destructive, efficient, and high-precision extraction of two-dimensional and three-dimensional phenotype parameters of potato plants. The achievements of this study could not only provide important technical support for the cultivation and breeding of potatoes but also provide strong support for phenotype-based research.
To achieve rapid on-site detection of aflatoxin B1 (AFB1) in agricultural and sideline products, a portable detection system based on differential pulse voltammetry (DPV) and STM32F103ZET6 as the core processor was designed. The system consists of two main parts: hardware detection devices and a mobile App, which are connected through Wi-Fi communication. The hardware detection equipment includes a DPV waveform generation circuit, constant potential circuit, and micro current detection module. The upper computer App was developed in an Android environment and completed tasks such as signal acquisition and data storage. After completing the design, experiments were conducted to verify the accuracy of the constant potential circuit and micro current detection module. The constant potential circuit accurately applied the voltage set by the program to the electrode, with a maximum error of 4 mV. The micro current detection module converts the current into a voltage signal according to the theoretical formula and amplifies it according to the theoretical amplification factor. The laboratory-made AFB1 sensor was used to effectively detect AFB1 in the range of 0.1 fg/ml to 100 pg/ml. The maximum relative error between the test results in the standard solution and the electrochemical workstation CHI760e was 7.37%. Furthermore, peanut oil samples with different concentrations of AFB1 were tested, and the results were compared to the CHI760e detection results as the standard, with a recovery rate of 96.8%~106.0%. Peanut samples with different degrees of mold were also tested and compared with CHI760e, with a maximum relative error of 7.10%.The system's portability allows it to be easily transported to different locations for on-site testing, making it an ideal solution for testing in remote or rural areas where laboratory facilities may be limited. Furthermore, the use of a mobile App for data acquisition and storage makes it easy to track and manage testing results. In summary, this portable detection system has great potential for widespread application in the rapid on-site detection of AFB1 in agricultural and sideline products.
To meet the needs of environmental monitoring and regulation in rabbit houses, a real-time environmental monitoring system for rabbit houses was proposed based on narrow band Internet of Things (NB-IoT). The system overcomes the limitations of traditional wired networks, reduces network costs, circuit components, and expenses is low. An Arduino development board and the Quectel BC260Y-NB-IoT network module were used, along with the message queuing telemetry transport (MQTT) protocol for remote telemetry transmission, which enables network connectivity and communication with an IoT cloud platform. Multiple sensors, including SGP30, MQ137, and 5516 photoresistors, were integrated into the system to achieve real-time monitoring of various environmental parameters within the rabbit house, such as sound decibels, light intensity, humidity, temperature, and gas concentrations. The collected data was stored for further analysis and could be used to inform environmental regulation and monitoring in rabbit houses, both locally and in the cloud. Signal alerts based on circuit principles were triggered when thresholds were exceeded, creating an optimal living environment for the rabbits. The advantages of NB-IoT networks and other networks, such as Wi-Fi and LoRa were compared. The technology and process of building a system based on the three-layer architecture of the Internet of Things was introduced. The prices of circuit components were analyzed, and the total cost of the entire system was less than 400 RMB. The system underwent network and energy consumption tests, and the transmission stability, reliability, and energy consumption were reasonable and consistent across different time periods, locations, and network connection methods. An average of 0.57 transactions per second (TPS) was processed by the NB-IoT network using the MQTT communication protocol, and 34.2 messages per minute were sent and received with a fluctuation of 1 message. The monitored device was found to have an average voltage of approximately 12.5 V, a current of approximately 0.42 A, and an average power of 5.3 W after continuous monitoring using an electricity meter. No additional power consumption was observed during communication. The performance of various sensors was tested through a 24-hour indoor test, during which temperature and lighting conditions showed different variations corresponding to day and night cycles. The readings were stably and accurately captured by the environmental sensors, demonstrating their suitability for long-term monitoring purposes. This system is can provide equipment cost and network selection reference values for remote or large-scale livestock monitoring devices.
[Objective] The fruit load estimation of fruit tree is essential for horticulture management. Traditional estimation method by manual sampling is not only labor-intensive and time-consuming but also prone to errors. Most existing models can not apply to edge computing equipment with limited computing resources because of their high model complexity. This study aims to develop a lightweight model for edge computing equipment to estimate fruit load automatically in the orchard. [Methods] The experimental data were captured using the smartphone in the citrus orchard in Jiangnan district, Nanning city, Guangxi province. In the dataset, 30 videos were randomly selected for model training and other 10 for testing. The general idea of the proposed algorithm was divided into two parts: Detecting fruits and extracting ReID features of fruits in each image from the video, then tracking fruit and estimating the fruit load. Specifically, the CSPDarknet53 network was used as the backbone of the model to achieve feature extraction as it consumes less hardware computing resources, which was suitable for edge computing equipment. The path aggregation feature pyramid network PAFPN was introduced as the neck part for the feature fusion via the jump connection between the low-level and high-level features. The fused features from the PAFPN were fed into two parallel branches. One was the fruit detection branch and another was the identity embedding branch. The fruit detection branch consisted of three prediction heads, each of which performed 3×3 convolution and 1×1 convolution on the feature map output by the PAFPN to predict the fruit's keypoint heat map, local offset and bounding box size, respectively. The identity embedding branch distinguished between different fruit identity features. In the fruit tracking stage, the byte mechanism from the ByteTrack algorithm was introduced to improve the data association of the FairMOT method, enhancing the performance of fruit load estimation in the video. The Byte algorithm considered both high-score and low-score detection boxes to associate the fruit motion trajectory, then matches the identity features' similarity of fruits between frames. The number of fruit IDs whose tracking duration longer than five frames was counted as the amount of citrus fruit in the video. [Results and Discussions] All experiments were conducted on edge computing equipment. The fruit detection experiment was conducted under the same test dataset containing 211 citrus tree images. The experimental results showed that applying CSPDarkNet53+PAFPN structure in the proposed model achieved a precision of 83.6%, recall of 89.2% and F1 score of 86.3%, respectively, which were superior to the same indexes of FairMOT (ResNet34) model, FairMOT (HRNet18) model and Faster RCNN model. The CSPDarkNet53+PAFPN structure adopted in the proposed model could better detect the fruits in the images, laying a foundation for estimating the amount of citrus fruit on trees. The model complexity experimental results showed that the number of parameters, FLOPs (Floating Point Operations) and size of the proposed model were 5.01 M, 36.44 G and 70.2 MB, respectively. The number of parameters for the proposed model was 20.19% of FairMOT (ResNet34) model's and 41.51% of FairMOT (HRNet18) model's. The FLOPs for the proposed model was 78.31% less than FairMOT (ResNet34) model's and 87.63% less than FairMOT (HRNet18) model's. The model size for the proposed model was 23.96% of FairMOT (ResNet34) model's and 45.00% of FairMOT (HRNet18) model's. Compared with the Faster RCNN, the model built in this study showed advantages in the number of parameters, FLOPs and model size. The low complexity proved that the proposed model was more friendly to edge computing equipment. Compared with the lightweight backbone network EfficientNet-Lite, the CSPDarkNet53 applied in the proposed model's backbone performed better fruit detection and model complexity. For fruit load estimation, the improved tracking strategy that integrated the Byte algorithm into the FairMOT positively boosted the estimation accuracy of fruit load. The experimental results on the test videos showed that the AEP (Average Estimating Precision) and FPS (Frames Per Second) of the proposed model reached 91.61% and 14.76 f/s, which indicated that the proposed model could maintain high estimation accuracy while the FPS was 2.4 times and 4.7 times of the comparison models, respectively. The RMSE (Root Mean Square Error) of the proposed model was 4.1713, which was 47.61% less than FairMOT (ResNet34) model's and 22.94% less than FairMOT (HRNet18) model's. The R2 of the determination coefficient between the algorithm-measured value and the manual counted value was 0.9858, which was superior to other comparison models. The proposed model revealed better performance in estimating fruit load and lower model complexity than other comparatives. [Conclusions] The experimental results proved the validity of the proposed model for fruit load estimation on edge computing equipment. This research could provide technical references for the automatic monitoring and analysis of orchard productivity. Future research will continue to enrich the data resources, further improve the model's performance, and explore more efficient methods to serve more fruit tree varieties.
[Objective] Aiming at the problems of low accuracy and incomplete coverage of image recognition of phenological period of apple in natural environment by traditional methods, an improved ResNet50 model was proposed for phenological period recognition of apple. [Methods] With 8 kinds of phenological period images of Red Fuji apple in Sichuan plateau area as the research objects and 3 sets of spherical cameras built in apple orchard as acquisition equipment, the original data set of 9800 images of apple phenological period were obtained, labeled by fruit tree experts. Due to the different duration of each phenological period of apple, there were certain differences in the quantity of collection. In order to avoid the problem of decreasing model accuracy due to the quantity imbalance, data set was enhanced by random cropping, random rotation, horizontal flip and brightness adjustment, and the original data set was expanded to 32,000 images. It was divided into training set (25,600 images), verification set (3200 images) and test set (3200 images) in a ratio of 8:1:1. Based on the ResNet50 model, the SE (Squeeze and Excitation Network) channel attention mechanism and Adam optimizer were integrated. SE channel attention was introduced at the end of each residual module in the benchmark model to improve the model's feature extraction ability for plateau apple tree images. In order to achieve fast convergence of the model, the Adam optimizer was combined with the cosine annealing attenuation learning rate, and ImageNet was selected as the pre-training model to realize intelligent recognition of plateau Red Fuji apple phenological period under natural environment. A "Intelligent Monitoring and Production Management Platform for Fruit Tree Growth Period" has been developed using the identification model of apple tree phenology. In order to reduce the probability of model misjudgment, improve the accuracy of model recognition, and ensure precise control of the platform over the apple orchard, three sets of cameras deployed in the apple orchard were set to capture motion trajectories, and images were collected at three time a day: early, middle, and late, a total of 27 images per day were collected. The model calculated the recognition results of 27 images and takes the category with the highest number of recognition as the output result to correct the recognition rate and improve the reliability of the platform. [Results and Discussions] Experiments were carried out on 32,000 apple tree images. The results showed that when the initial learning rate of Adam optimizer was set as 0.0001, the accuracy of the test model tended to the optimal, and the loss value curve converged the fastest. When the initial learning rate was set to 0.0001 and the iteration rounds are set to 30, 50 and 70, the accuracies of the optimal verification set obtained by the model was 0.9354, 0.9635 and 0.9528, respectively. Therefore, the improved ResNet50 model selects the learning rate of 0.0001 and iteration rounds of 50 as the training parameters of the Adam optimizer. Ablation experiments showed that the accuracy of validation set and test set were increased by 0.8% and 2.99% in the ResNet50 model with increased SE attention mechanism, respectively. The validation set accuracy and test set accuracy of the ResNet50 model increased by 2.19% and 1.42%, respectively, when Adam optimizer was added. The accuracy of validation set and test set was 2.33% and 3.65%, respectively. The accuracy of validation set was 96.35%, the accuracy of test set was 91.94%, and the average detection time was 2.19 ms.Compared with the AlexNet, VGG16, ResNet18, ResNet34, and ResNet101 models, the improved ResNet50 model improved the accuracy of the optimal validation set by 9.63%, 5.07%, 5.81%, 4.55%, and 0.96%, respectively. The accuracy of the test set increased by 12.31%, 6.88%, 8.53%, 8.67%, and 5.58%, respectively. The confusion matrix experiment result showed that the overall recognition rate of the improved ResNet50 model for the phenological period of apple tree images was more than 90%, of which the accuracy rate of bud stage and dormancy stage was the lowest, and the probability of mutual misjudgment was high, and the test accuracy rates were 89.50% and 87.44% respectively. There were also a few misjudgments during the young fruit stage, fruit enlargement stage, and fruit coloring stage due to the similarity in characteristics between adjacent stages. The external characteristics of the Red Fuji apple tree were more obvious during the flowering and fruit ripening stages, and the model had the highest recognition rate for the flowering and fruit ripening stages, with test accuracy reaching 97.50% and 97.49%, respectively. [Conclusions] The improved ResNet50 can effectively identify apple phenology, and the research results can provide reference for the identification of orchard phenological period. After integration into the intelligent monitoring production management platform of fruit tree growth period, intelligent management and control of apple orchard can be realized.
[Objective] Rapid recognition and automatic positioning of table grapes in the natural environment is the prerequisite for the automatic picking of table grapes by the picking robot. [Methods] An rapid recognition and automatic picking points positioning method based on improved K-means clustering algorithm and contour analysis was proposed. First, euclidean distance was replaced by a weighted gray threshold as the judgment basis of K-means similarity. Then the images of table grapes were rasterized according to the K value, and the initial clustering center was obtained. Next, the average gray value of each cluster and the percentage of pixel points of each cluster in the total pixel points were calculated. And the weighted gray threshold was obtained by the average gray value and percentage of adjacent clusters. Then, the clustering was considered as have ended until the weighted gray threshold remained unchanged. Therefore, the cluster image of table grape was obtained. The improved clustering algorithm not only saved the clustering time, but also ensured that the K value could change adaptively. Moreover, the adaptive Otsu algorithm was used to extract grape cluster information, so that the initial binary image of the table grape was obtained. In order to reduce the interference of redundant noise on recognition accuracy, the morphological algorithms (open operation, close operation, images filling and the maximum connected domain) were used to remove noise, so the accurate binary image of table grapes was obtained. And then, the contours of table grapes were obtained by the Sobel operator. Furthermore, table grape clusters grew perpendicular to the ground due to gravity in the natural environment. Therefore, the extreme point and center of gravity point of the grape cluster were obtained based on contour analysis. In addition, the linear bundle where the extreme point and the center of gravity point located was taken as the carrier, and the similarity of pixel points on both sides of the linear bundle were taken as the judgment basis. The line corresponding to the lowest similarity value was taken as the grape stem, so the stem axis of the grape was located. Moreover, according to the agronomic picking requirements of table grapes, and combined with contour analysis, the region of interest (ROI) in picking points could be obtained. Among them, the intersection of the grapes stem and the contour was regarded as the middle point of the bottom edge of the ROI. And the 0.8 times distance between the left and right extreme points was regarded as the length of the ROI, the 0.25 times distance between the gravity point and the intersection of the grape stem and the contour was regarded as the height of the ROI. After that, the central point of the ROI was captured. Then, the nearest point between the center point of the ROI and the grape stem was determined, and this point on the grape stem was taken as the picking point of the table grapes. Finally, 917 grape images (including Summer Black, Moldova, and Youyong) taken by the rear camera of MI8 mobile phone at Jinniu Mountain Base of Shandong Fruit and Vegetable Research Institute were verified experimentally. Results and Discussions] The results showed that the success rate was 90.51% when the error between the table grape picking points and the optimal points were less than 12 pixels, and the average positioning time was 0.87 s. The method realized the fast and accurate localization of table grape picking points. On top of that, according to the two cultivation modes (hedgerow planting and trellis planting) of table grapes, a simulation test platform based on the Dense mechanical arm and the single-chip computer was set up in the study. 50 simulation tests were carried out for the four conditions respectively, among which the success rate of localization for purple grape picking point of hedgerow planting was 86.00%, and the average localization time was 0.89 s; the success rate of localization for purple grape identification and localization of trellis planting was 92.00%, and the average localization time was 0.67 s; the success rate of localization for green grape picking point of hedgerow planting was 78.00%, and the average localization time was 0.72 s; and the success rate of localization for green grape identification and localization of trellis planting was 80.00%, and the average localization time was 0.71 s. [Conclusions] The experimental results showed that the method proposed in the study can meet the requirements of table grape picking, and can provide technical supports for the development of grape picking robot.
[Objective] Pineapple is a common tropical fruit, and its ripeness has an important impact on the storage and marketing. It is particularly important to analyze the maturity of pineapple fruit before picking. Deep learning technology can be an effective method to achieve automatic recognition of pineapple maturity. To improve the accuracy and rate of automatic recognition of pineapple maturity, a new network model named MobileNet V3-YOLOv4 was proposed in this study. [Methods] Firstly, pineapple maturity analysis data set was constructed. A total of 1580 images were obtained, with 1264 images selected as the training set, 158 images as the validation set, and 158 images as the test set. Pineapple photos were taken in natural environment. In order to ensure the diversity of the data set and improve the robustness and generalization of the network, pineapple photos were taken under the influence of different factors such as branches and leaves occlusion, uneven lighting, overlapping shadows, etc. and the location, weather and growing environment of the collection were different. Then, according to the maturity index of pineapple, the photos of pineapple with different maturity were marked, and the labels were divided into yellow ripeness and green ripeness. The annotated images were taken as data sets and input into the network for training. Aiming at the problems of the traditional YOLOv4 network, such as large number of parameters, complex network structure and slow reasoning speed, a more optimized lightweight MobileNet V3-YOLOv4 network model was proposed. The model utilizes the benck structure to replace the Resblock in the CSPDarknet backbone network of YOLOv4. Meanwhile, in order to verify the effectiveness of the MobileNet V3-YOLOv4 network, MobileNet V1-YOLOv4 model and MobileNet V2-YOLOv4 model were also trained. Five different single-stage and two-stage network models, including R-CNN, YOLOv3, SSD300, Retinanet and Centernet were compared with each evaluation index to analyze the performance superiority of MobileNet V3-YOLOv4 model. Results and Discussions] MobileNet V3-YOLOv4 was validated for its effectiveness in pineapple maturity detection through experiments comparing model performance, model classification prediction, and accuracy tests in complex pineapple detection environments.The experimental results show that, in terms of model performance comparison, the training time of MobileNet V3-YOLOv4 was 11,924 s, with an average training time of 39.75 s per round, the number of parameters was 53.7 MB, resulting in a 25.59% reduction in the saturation time compared to YOLOv4, and the parameter count accounted for only 22%. The mean average precision (mAP) of the trained MobileNet V3-YOLOv4 in the verification set was 53.7 MB. In order to validate the classification prediction performance of the MobileNet V3-YOLOv4 model, four metrics, including Recall score, F1 Score, Precision, and average precision (AP), were utilized to classify and recognize pineapples of different maturities. The experimental results demonstrate that MobileNet V3-YOLOv4 exhibited significantly higher Precision, AP, and F1 Score the other. For the semi-ripe stage, there was a 4.49% increase in AP, 0.07 improvement in F1 Score, 1% increase in Recall, and 3.34% increase in Precision than YOLOv4. As for the ripe stage, there was a 6.06% increase in AP, 0.13 improvement in F1 Score, 16.55% increase in Recall, and 6.25% increase in Precision. Due to the distinct color features of ripe pineapples and their easy differentiation from the background, the improved network achieved a precision rate of 100.00%. Additionally, the mAP and reasoning speed (Frames Per Second, FPS) of nine algorithms were examined. The results showed that MobileNet V3-YOLOv4 achieved an mAP of 90.92%, which was 5.28% higher than YOLOv4 and 3.67% higher than YOLOv3. The FPS was measured at 80.85 img/s, which was 40.28 img/s higher than YOLOv4 and 8.91 img/s higher than SSD300. The detection results of MobileNet V3-YOLOv4 for pineapples of different maturities in complex environments indicated a 100% success rate for both the semi-ripe and ripe stages, while YOLOv4, MobileNet V1-YOLOv4, and MobileNet V2-YOLOv4 exhibited varying degrees of missed detections. [Conclusions] Based on the above experimental results, it can be concluded that MobileNet V3-YOLOv4 proposed in this study could not only reduce the training speed and parameter number number, but also improve the accuracy and reasoning speed of pineapple maturity recognition, so it has important application prospects in the field of smart orchard. At the same time, the pineapple photo data set collected in this research can also provide valuable data resources for the research and application of related fields.
[Objective] To address the problems of low efficiency and high cost of traditional manual identification of pests and diseases, improve the automatic recognition of pests and diseases by introducing advanced technical means, and provide feasible technical solutions for agricultural pest and disease monitoring and prevention and control, a rice image recognition model GE-DenseNet (G-ECA DenseNet) based on improved ECA (Efficient Channel Attention) mechanism with DenseNet201 was proposed. [Methods] The leaf images of three pests and diseases, namely, brownspot, hispa, leafblast and healthy rice were selected as experimental materials. The images were captured at the Zhuanghe Rice Professional Cooperative in Yizheng, Jiangsu Province, and the camera was used to manually take pictures from multiple angles such as the top and side of rice every 2 h, thus acquiring 1250 images of rice leaves under different lighting conditions, different perspectives, and different shading environments. In addition, samples about pests and diseases were collected in the Kaggle database. There were 1488 healthy leaves, 523 images of brownspot, 565 images of hispa, and 779 images of leafblast in the dataset. Since the original features of the pest and disease data were relatively close, firstly, the dataset was divided into a training set and a test set according to the ratio of 9:1, and then data enhancement was performed on the training set. A region of interest (ROI) was randomly selected to achieve a local scale of 1.1 to 1.25 for the sample images of the dataset, thus simulating the situation that only part of the leaves were captured in the actual shooting process due to the different distance of the plants from the camera. In addition, a random rotation of a certain angle was used to crop the image to simulate the different angles of the leaves. Finally, the experimental training set contains 18,018 images and the test set contains 352 images. The GE-DenseNet model firstly introduces the idea of Ghost module on the ECA attention mechanism to constitute the G-ECA Layer structure, which replaces the convolution operation with linear transformation to perform efficient fusion of channel features while avoiding dimensionality reduction when learning channel attention information and effectively enhancing its ability to extract features. Secondly, since the original Dense Block only considered the correlation between different layers and ignores the extraction of important channel information in the image recognition process, introducing G-ECA Layer before the original Dense Block of DenseNet201 gives the model a better channel feature extraction capability and thus improved the recognition accuracy. Due to the small dataset used in the experiment, the weight parameters of DenseNet201 pre-trained on the ImageNet dataset were migrated to GE-DenseNet. During the training process, the BatchSize size was set to 32, the number of iterations (Epoch) was set to 50, and the Focal Loss function was used to solve the problem of unbalanced samples for each classification. Meanwhile, the adaptive moment estimation (Adam) optimizer was used to avoid the problem of drastic gradient changes in back propagation due to random initialization of some weights at the early stage of model training, which weakened the uncertainty of network training to a certain extent. [Results and Discussions] Experimental tests were conducted on a homemade dataset of rice pests and diseases, and the recognition accuracy reached 83.52%. Comparing the accuracy change graphs and loss rate change graphs of GE-DenseNet and DenseNet201, it could be found that the proposed method in this study was effective in training stability, which could accelerate the speed of model convergence and improve the stability of the model, making the network training process more stable. And observing the visualization results of GE-DenseNet and DenseNet201 corresponding feature layers, it could be found that the features were more densely reflected around the pests and diseases after adding the G-ECA Layer structure. From the ablation comparison experiments of the GE-DenseNet model, it could be obtained that the model accuracy increased by 2.27% after the introduction of the Focal Loss function with the G-ECA Layer layer. Comparing the proposed model with the classical NasNet (4@1056), VGG-16 and ResNet50 models, the classification accuracy increased by 6.53%, 4.83% and 3.69%, respectively. Compared with the original DenseNet201, the recognition accuracy of hispa improved 20.32%. [Conclusions] The experimental results showed that the addition of G-ECA Layer structure enables the model to more accurately capture feature information suitable for rice pest recognition, thus enabling the GE-DenseNet model to achieve more accurate recognition of different rice pest images. This provides reliable technical support for timely pest and disease control, reducing crop yield loss and pesticide use. Future research can lighten the model and reduce its size without significantly reducing the recognition accuracy, so that it can be deployed in UAVs, tractors and various distributed image detection edge devices to facilitate farmers to conduct real-time inspection of farmland and further enhance the intelligence of agricultural production.
[Objective] To quickly and accurately assess the situation of crop lodging disasters, it is necessary to promptly obtain information such as the location and area of the lodging occurrences. Currently, there are no corresponding technical standards for identifying crop lodging based on UAV remote sensing, which is not conducive to standardizing the process of obtaining UAV data and proposing solutions to problems. This study aims to explore the impact of different spatial resolution remote sensing images and feature optimization methods on the accuracy of identifying wheat lodging areas. [Methods] Digital orthophoto images (DOM) and digital surface models (DSM) were collected by UAVs with high-resolution sensors at different flight altitudes after wheat lodging. The spatial resolutions of these image data were 1.05, 2.09, and 3.26 cm. A full feature set was constructed by extracting 5 spectral features, 2 height features, 5 vegetation indices, and 40 texture features from the pre-processed data. Then three feature selection methods, ReliefF algorithm, RF-RFE algorithm, and Boruta-Shap algorithm, were used to construct an optimized subset of features at different flight altitudes to select the best feature selection method. The ReliefF algorithm retains features with weights greater than 0.2 by setting a threshold of 0.2; the RF-RFE algorithm quantitatively evaluated the importance of each feature and introduces variables in descending order of importance to determine classification accuracy; the Boruta-Shap algorithm performed feature subset screening on the full feature set and labels a feature as green when its importance score was higher than that of the shaded feature, defining it as an important variable for model construction. Based on the above-mentioned feature subset, an object-oriented classification model on remote sensing images was conducted using eCognition9.0 software. Firstly, after several experiments, the feature parameters for multi-scale segmentation in the object-oriented classification were determined, namely a segmentation scale of 1, a shape factor of 0.1, and a tightness of 0.5. Three object-oriented supervised classification algorithms, support vector machine (SVM), random forest (RF), and K nearest neighbor (KNN), were selected to construct wheat lodging classification models. The Overall classification accuracy and Kappa coefficient were used to evaluate the accuracy of wheat lodging identification. By constructing a wheat lodging classification model, the appropriate classification strategy was clarified and a technical path for lodging classification was established. This technical path can be used for wheat lodging monitoring, providing a scientific basis for agricultural production and improving agricultural production efficiency. [Results and Discussions] The results showed that increasing the altitude of the UAV to 90 m significantly improved flight efficiency of wheat lodging areas. In comparison to flying at 30 m for the same monitoring range, data acquisition time was reduced to approximately 1/6th, and the number of photos needed decreased from 62 to 6. In terms of classification accuracy, the overall classification effect of SVM is better than that of RF and KNN. Additionally, when the image spatial resolution varied from 1.05 to 3.26 cm, the full feature set and all three optimized feature subsets had the highest classification accuracy at a resolution of 1.05 cm, which was better than at resolutions of 2.09 and 3.26 cm. As the image spatial resolution decreased, the overall classification effect gradually deteriorated and the positioning accuracy decreased, resulting in poor spatial consistency of the classification results. Further research has found that the Boruta-Shap feature selection method can reduce data dimensionality and improve computational speed while maintaining high classification accuracy. Among the three tested spatial resolution conditions (1.05, 2.09, and 3.26 cm), the combination of SVM and Boruta-Shap algorithms demonstrated the highest overall classification accuracy. Specifically, the accuracy rates were 95.6%, 94.6%, and 93.9% for the respective spatial resolutions. These results highlighted the superior performance of this combination in accurately classifying the data and adapt to changes in spatial resolution. When the image resolution was 3.26 cm, the overall classification accuracy decreased by 1.81% and 0.75% compared to 1.05 and 2.09 cm; when the image resolution was 2.09 cm, the overall classification accuracy decreased by 1.06% compared to 1.05 cm, showing a relatively small difference in classification accuracy under different flight altitudes. The overall classification accuracy at an altitude of 90 m reached 95.6%, with Kappa coefficient of 0.914, meeting the requirements for classification accuracy. [Conclusions] The study shows that the object-oriented SVM classifier and the Boruta-Shap feature optimization algorithm have strong application extension advantages in identifying lodging areas in remote sensing images at multiple flight altitudes. These methods can achieve high-precision crop lodging area identification and reduce the influence of image spatial resolution on model stability. This helps to increase flight altitude, expand the monitoring range, improve UAV operation efficiency, and reduce flight costs. In practical applications, it is possible to strike a balance between classification accuracy and efficiency based on specific requirements and the actual scenario, thus providing guidance and support for the development of strategies for acquiring crop lodging information and evaluating wheat disasters.
[Objective] Leaf chlorophyll content (LCC) of flue-cured Tobacco is an important indicator for characterizing the photosynthesis, nutritional status, and growth of the crop. Tobacco is an important economic crop with leaves as the main harvest object, it is crucial to monitor its LCC. Hyperspectral data can be used for the rapid estimation of LCC in flue-cured tobacco leaves, making it of great significance and application value. The purpose of this study was to efficiently and accurately estimate the LCC of flue-cured tobacco during different growth stages. [Methods] Zhongyan 100 was chose as the research object, five nitrogen fertilization levels were set. In each plot, three plants were randomly and destructively sampled, resulting in a total of 45 ground samples for each data collection. After transplanting, the reflectance data of the flue-cured tobacco canopy at six growth stages (32, 48, 61, 75, 89, and 109 d ) were collected using a UAV equipped with a Resonon Pika L hyperspectral. Spectral indices for the LCC estimation model of flue-cured tobacco were screened in two ways: (1) based on 18 published vegetation indices sensitive to LCC of crop leaves; (2) based on random combinations of any two bands in the wavelength range of 400‒1000 nm. The Difference Spectral Index (DSI), Ratio Spectral Index (RSI), and Normalized Spectral Index (NDSI) were calculated and plotted against LCC. The correlations between the three spectral indices and leaf LCC were calculated and plotted using contour maps. Five regression models, unary linear regression (ULR), multivariable linear regression (MLR), partial least squares regression (PLSR), support vector regression (SVR), and random forest regression (RFR), were used to estimate the chlorophyll content. A regression estimate model of LCC based on various combinations of spectral indices was eventually constructed by comparing the prediction accuracies of single spectral index models multiple spectral index models at different growth stages. Results and Discussions] The results showed that the LCC range for six growth stages was 0.52‒2.95 mg/g. The standard deviation and coefficient of variation values demonstrated a high degree of dispersion in LCC, indicating differences in fertility between different treatments at the test site and ensuring the applicability of the estimation model within a certain range. Except for 109 d after transplanting, most vegetation indices were significantly correlated with LCC (p<0.01). Compared with traditional vegetation indices, the newly combined spectral indices significantly improved the correlation with LCC. The sensitive bands at each growth stage were relatively concentrated, and the spectral index combinations got high correlation with LCC were mainly distributed between 780‒940 nm and 520‒710 nm. The sensitive bands for the whole growth stages were relatively dispersed, and there was little difference in the position of sensitive band between different spectral indices. For the univariate LCC estimation model, the highest modeling accuracy was achieved using the newly combined Normalized Spectral Index and Red Light Ratio Spectral Index at 75 d after transplanting. The coefficients of determination (R2 ) and root mean square errors (RMSE) for the modeling and validation sets were 0.822, 0.814, and 0.226, 0.230, respectively. The prediction results of the five resgression models showed that the RFR algorithm based on multivariate data performed best in LCC estimation. The R2 and RMSE of the modeling set using data at 75 d after transplanting were 0.891 and 0.205, while those of the validation set reached 0.919 and 0.146. In addition, the estimation performance of the univariate model based on the whole growth stages dataset was not ideal, with R2 of 0.636 and 0.686, and RMSE of 0.333 and 0.304 for the modeling and validation sets, respectively. However, the estimation accuracy of the model based on multiple spectral parameters was significantly improved in the whole growth stages dataset, with R2 of 0.854 and 0.802, and RMSE of 0.206 and 0.264 for the modeling and validation sets of the LCC-RFR model, respectively. In addition, in the whole growth stages dataset, the estimation accuracy of the LCC-RFR model was better than that of the LCC-MLR, LCC-PLSR, and LCC-SVR models. Compared with the modeling set, R2 increased by 19.06%, 18.62%, and 29.51%, while RMSE decreased by 31.93%, 29.51%, and 28.24%. Compared with the validation set, R2 increased by 8.21%, 12.62%, and 8.17%, while RMSE decreased by 3.76%, 9.33%, and 4.55%. [Conclusions] The sensitivity of vegetation indices (VIs) to LCC is closely connected to the tobacco growth stage, according to the results this study, which examined the reaction patterns of several spectral indices to LCC in flue-cured tobacco. The sensitivity of VIs to LCC at various growth stages is critical for crop parameter assessment using UAV hyperspectral photography. Five estimation models for LCC in flue-cured tobacco leaves were developed, with the LCC-RFR model demonstrating the greatest accuracy and stability. The RFR model is less prone to overfitting and can efficiently decrease outlier and noise interference. This work could provide theoretical and technological references for LCC estimate and flue-cured tobacco growth monitoring.
[Objective] Accurate prediction of changes in sugarcane yield in Guangxi can provide important reference for the formulation of relevant policies by the government and provide decision-making basis for farmers to guide sugarcane planting, thereby improving sugarcane yield and quality and promoting the development of the sugarcane industry. This research was conducted to provide scientific data support for sugar factories and related management departments, explore the relationship between sugarcane yield and meteorological factors in the main sugarcane producing areas of Guangxi Zhuang Autonomous Region. [Methods] The study area included five sugarcane planting regions which laid in five different counties in Guangxi, China. The average yields per hectare of each planting regions were provided by Guangxi Sugar Industry Group which controls the sugar refineries of each planting region. The daily meteorological data including 14 meteorological factors from 2002 to 2019 were acquired from National Data Center for Meteorological Sciences to analyze their influences placed on sugarcane yield. Since meteorological factors could pose different influences on sugarcane growth during different time spans, a new kind of factor which includes meteorological factors and time spans was defined, such as the average precipitation in August, the average temperature from February to April, etc. And then the inter-correlation of all the meteorological factors of different time spans and their correlations with yields were analyzed to screen out the key meteorological factors of sensitive time spans. After that, four algorithms of BP neural network (BPNN), support vector machine (SVM), random forest (RF), and long short-term memory (LSTM) were employed to establish sugarcane apparent yield prediction models for each planting region. Their corresponding reference models based on the annual meteorological factors were also built. Additionally, the meteorological yields of every planting region were extracted by HP filtering, and a general meteorological yield prediction model was built based on the data of all the five planting regions by using RF, SVM BPNN, and LSTM, respectively. [Results and Discussions] The correlation analysis showed that different planting regions have different sensitive meteorological factors and key time spans. The highly representative meteorological factors mainly included sunshine hours, precipitation, and atmospheric pressure. According to the results of correlation analysis, in Region 1, the highest negative correlation coefficient with yield was observed at the sunshine hours during October and November, while the highest positive correlation coefficient was found at the minimum relative humidity in November. In Region 2, the maximum positive correlation coefficient with yield was observed at the average vapor pressure during February and March, whereas the maximum negative correlation coefficient was associated with the precipitation in August and September. In Region 3, the maximum positive correlation coefficient with yield was found at the 20‒20 precipitation during August and September, while the maximum negative correlation coefficient was related to sunshine hours in the same period. In Region 4, the maximum positive correlation coefficient with yield was observed at the 20‒20 precipitation from March to December, whereas the maximum negative correlation coefficient was associated with the highest atmospheric pressure from August to December. In Region 5, the maximum positive correlation coefficient with yield was found at the average vapor pressure from June and to August, whereas the maximum negative correlation coefficient as related to the lowest atmospheric pressure in February and March. For each specific planting region, the accuracy of apparent yield prediction model based on sensitive meteorological factors during key time spans was obviously better than that based on the annual average meteorological values. The LSTM model performed significantly better than the widely used classic BPNN, SVM, and RF models for both kinds of meteorological factors (under sensitive time spans or annually). The overall root mean square error (RMSE) and mean absolute percentage error (MAPE) of the LSTM model under key time spans were 10.34 t/ha and 6.85%, respectively, with a coefficient of determination Rv2 of 0.8489 between the predicted values and true values. For the general prediction models of the meteorological yield to multiple the sugarcane planting regions, the RF, SVM, and BPNN models achieved good results, and the best prediction performance went to BPNN model, with an RMSE of 0.98 t/ha, MAPE of 9.59%, and Rv2 of 0.965. The RMSE and MAPE of the LSTM model were 0.25 t/ha and 39.99%, respectively, and the Rv2 was 0.77. [Conclusions] Sensitive meteorological factors under key time spans were found to be more significantly correlated with the yields than the annual average meteorological factors. LSTM model shows better performances on apparent yield prediction for specific planting region than the classic BPNN, SVM, and RF models, but BPNN model showed better results than other models in predicting meteorological yield over multiple sugarcane planting regions.
[Objective] Desert vegetation is an indispensable part of desert ecosystems, and its conservation and restoration are crucial. Accurate identification of desert plants is an indispensable task, and is the basis of desert ecological research and conservation. The complex growth environment caused by light, soil, shadow and other vegetation increases the recognition difficulty, and the generalization ability is poor and the recognition accuracy is not guaranteed. The rapid development of modern technology provides new opportunities for plant identification and classification. By using intelligent identification algorithms, field investigators can be effectively assisted in desert plant identification and classification, thus improve efficiency and accuracy, while reduce the associated human and material costs. [Methods] In this research, the following works were carried out for the recognition of desert plant: Firstly, a training dataset of deep learning model of desert plant images in the arid and semi-arid region of Xinjiang was constructed to provide data resources and basic support for the classification and recognition of desert plant images.The desert plant image data was collected in Changji and Tacheng region from the end of September 2021 and July to August 2022, and named DPlants50. The dataset contains 50 plant species in 13 families and 43 genera with a total of 12,507 images, and the number of images for each plant ranges from 183 to 339. Secondly, a migration integration learning-based algorithm for desert plant image recognition was proposed, which could effectively improve the recognition accuracy. Taking the EfficientNet B0-B4 network as the base network, the ImageNet dataset was pre-trained by migration learning, and then an integrated learning strategy was adopted combining Bagging and Stacking, which was divided into two layers. The first layer introduced K-fold cross-validation to divide the dataset and trained K sub-models by borrowing the Stacking method. Considering that the output features of each model were the same in this study, the second layer used Bagging to integrate the output features of the first layer model by voting method, and the difference was that the same sub-models and K sub-models were compared to select the better model, so as to build the integrated model, reduce the model bias and variance, and improve the recognition performance of the model. For 50 types of desert plants, 20% of the data was divided as the test set, and the remaining 5 fold cross validation was used to divide the dataset, then can use DPi(i=1,2,…,5) represents each training or validation set. Based on the pre trained EfficientNet B0-B4 network, training and validation were conducted on 5 data subsets. Finally, the model was integrated using soft voting, hard voting, and weighted voting methods, and tested on the test set. [Results and Discussions] The results showed that the Top-1 accuracy of the single sub-model based on EfficientNet B0 network was 92.26%~93.35%, the accuracy of the Ensemble-Soft model with soft voting, the Ensemble-Hard model with hard voting and the Ensemble-Weight model integrated by weighted voting method were 93.63%, 93.55% and 93.67%, F1 Score and accuracy were comparable, the accuracy and F1 Score of Ensemble-Weight model integrated by weighted voting method were not significantly improved compared with Ensemble-Soft model and Ensemble-hard model, but it showed that the effect of weighted voting method proposed in this study was better than both of them. The three integrated models demonstrate no noteworthy enhancements in accuracy and F1 Score when juxtaposed with the five sub-models. This observation results suggests that the homogeneity among the models constrains the effectiveness of the voting method strategy. Moreover, the recognition effects heavily hinges on the performance of the EfficientNet B0-DP5 model. Therefore, the inclusion of networks with more pronounced differences was considered as sub-models. A single sub-model based on EfficientNet B0-B4 network had the highest Top-1 accuracy of 96.65% and F1 Score of 96.71%, while Ensemble-Soft model, Ensemble-Hard model and Ensemble-Weight model got the accuracy of 99.07%, 98.91% and 99.23%, which further improved the accuracy compared to the single sub-model, and the F1 Score was basically the same as the accuracy rate, and the model performance was significant. The model integrated by the weighted voting method also improved accuracy and F1 Score for both soft and hard voting, with significant model performance and better recognition, again indicating that the weighted voting method was more effective than the other two. Validated on the publicly available dataset Oxford Flowers102, the three integrated models improved the accuracy and F1 Score of the three sub-models compared to the five sub-models by a maximum of 4.56% and 5.05%, and a minimum of 1.94% and 2.29%, which proved that the migration and integration learning strategy proposed in this paper could effectively improve the model performances. [Conclusions] In this study, a method to recognize desert plant images in natural context by integrating migration learning and integration learning was proposed, which could improve the recognition accuracy of desert plants up to 99.23% and provide a solution to the problems of low accuracy, model robustness and weak generalization of plant images in real field environment. After transferring to the server through the cloud, it can realize the accurate recognition of desert plants and serve the scenes of field investigation, teaching science and scientific experiment.
[Objective] Saposhnikovia divaricata (Turcz.) Schischk is a kind of traditional Chinese medicine. Currently, the methods of identifying the origin and quality of Saposhnikovia divaricata (Turcz.) Schischk are mainly based on their physical or chemical characteristics, which is impossible to make an accurate measurement of Groundness identification. With the continuous development of deep learning, its advantages of no manual extraction and high classification accuracy are widely used in different fields, and an attention-embedded ShuffleNet V2-based model was proposed in this study to address the problems of large computation and low accuracy of most convolutional neural network models in the identification of Chinese herbal medicine Saposhnikovia divaricata (Turcz.) Schischk. [Methods] The model architecture was adjusted to reduce the number of model parameters and computation without degrading the network performance, and the traditional residual network was replaced by the Hourglass residual network, while the SE attention mechanism was introduced to embed the hourglass residual network with additional channel attention into ShuffleNet V2. The important features were enhanced and the unimportant features were weakened by controlling the size of the channel ratio to make the extracted features more directional by SE attention. The SiLU activation function was used to replace the ReLU activation function to enhance the generalization ability of the model Enriching local feature learning. Therefore, a lightweight Shuffle-Hourglass SE model was proposed. The samples of Saposhnikovia divaricata (Turcz.) Schischk used in this research were samples from the main production areas, including more than 1000 samples from five production areas in Heilongjiang, Jilin, Hebei, Gansu and Inner Mongolia. A total of 5234 images of Saposhnikovia divaricata (Turcz.) Schischk were obtained by using cell phone photography indoors under white daylight, fully taking into account the geographical distribution differences of different Saposhnikovia divaricata (Turcz.) Schischk. The data set of Saposhnikovia divaricata (Turcz.) Schischk images was expanded to 10,120 by using random flip, random crop, brightness and contrast enhancement processes. In order to verify the effectiveness of the model proposed, four classical network models, VGG16, MobileNet V2, ShuffleNet V2 and SqueezeNet V2, were selected for comparison experiments, ECA ( Efficient Channel Attention ) attention mechanism, CBAM ( Convolutional Block Attention Module ) attention mechanism and CA attention mechanism were chosen to compare with SE. All attention mechanisms were introduced into the same position in the ShuffleNet V2 model, and ReLU, H-swish and ELU activation functions were selected for contrast experiments under the condition in which other parameters unchanged. In order to explore the performance improvement of ShuffleNet V2 model by using the attention mechanism of SE module, Hourglass residual block and activation function, Shuffle-Hourglass SE model ablation experiment was carried out. Finally, loss, accuracy, precision, recall and F1 score in test set and training set were used as evaluation indexes of model performances. [Results and Discussions] The results showed that the Shuffle-Hourglass SE model proposed achieved the best performances. An accuracy of 95.32%, recall of 95.28%, and F1 score of 95.27% were obtained in the test set, which was 2.09%, 2.1 %, and 2.19 % higher than the ShuffleNet V2 model, respectively. The test duration and model size were 246.34 ms and 3.23 M, respectively, which were not only optimal among Traditional CNN such as VGG and Desnet,but had great advantages among lightweight networks such as MobileNet V2、SqueezeNet V2 and ShufffleNet V2. Compared with the classical convolutional network VGG, 7.41% of the accuracy was improved, 71.89% of the test duration was reduced, and 96.76% of the model size was reduced by the Shuffle-Hourglass SE model proposed in this study. Although the test duration of ShuffleNet V2 and MobileNet V2 were similar, the accuracy and speed of the Shuffle-Hourglass SE model improved, which proved its better performance. Compared with MobileNet V2, the test duration was reduced by 69.31 ms, the model size was reduced by 1.98 M, and the accuracy was increased by 10.5 %. In terms of classification accuracy, the improved network maintains higher recognition accuracy and better classification performance. [Conclusions] The model proposed in this research is able to identify the Saposhnikovia divaricata (Turcz.) Schischk originality well while maintaining high identification accuracy and consuming less storage space, which is helpful for realizing real-time identification of Saposhnikovia divaricata (Turcz.) Schischk originality in the future low performance terminals.
[Objective] Conducting research on the recognition of yak meat parts can help avoid confusion and substandard parts during the production and sales of yak meat, improve the transparency and traceability of the yak meat industry, and ensure food safety. To achieve fast and accurate recognition of different parts of yak meat, this study proposed an improved residual network model and developed a smartphone based yak meat part recognition software. [Methods] Firstly, the original data set of 1960 yak tenderloin, high rib, shank and brisket were expanded by 8 different data enhancement methods, including horizontal flip, vertical flip, random direction rotation 30°, random direction rotation 120°, random direction rotation 300°, contrast adjustment, saturation adjustment and hue adjustment. After expansion, 17,640 yak meat images of different parts were obtained. The expanded yak meat images of different parts were divided according to the 4:1 ratio, resulting in 14,112 yak meat sample images in the training set and 3528 yak meat sample images in the test set. Secondly, the convolutional block attention module (CBAM) was integrated into each residual block of the original network model to enhance the extraction of key detail features of yak images in different parts. At the same time, introducing this mechanism into the network model could achieve greater accuracy improvement with less computational overhead and fewer parameters. In addition, in the original network model, the full connection layer was directly added after all residual blocks instead of global average pooling and global maximum pooling, which could improve the accuracy of the network model, prevent overfitting, reduce the number of connections in subsequent network layers, accelerate the execution speed of the network model, and reduce the computing time when the mobile phone recognized images. Thirdly, different learning rates, weight attenuation coefficients and optimizers were used to verify the influence of the improved ResNet18_CBAM network model on convergence speed and accuracy. According to the experiments, the stochastic gradient descent (SGD) algorithm was adopted as the optimizer, and when the learning rate was 0.001 and the weight attenuation coefficient was 0, the improved ReaNet18_CBAM network model had the fastest convergence speed and the highest recognition accuracy on different parts of yak data sets. Finally, the PyTorch Mobile module in PyTorch deep learning framework was used to convert the trained ResNet18_CBAM network model into TorchScript model and saved it in *.ptl. Then, the yak part recognition App was developed using the Android Studio development environment, which included two parts: Front-end interface and back-end processing. The front-end of the App uses *.xml for a variety of price control layout, and the back-end used Java language development. Then TorchScript model in *.ptl was used to identify different parts of yak meat. Results and Discussions] In this study, CBAM, SENet, NAM and SKNet, four popular attentional mechanism modules, were integrated into the original ResNet18 network model and compared by ablation experiments. Their recognition accuracy on different parts of yak meat dataset were 96.31%, 94.12%, 92.51% and 93.85%, respectively. The results showed that among CBAM, SENet, NAM and SKNet, the recognition accuracy of ResNet18 CBAM network model was significantly higher than that of the other three attention mechanism modules. Therefore, the CBAM attention mechanism module was chosen as the improvement module of the original network model. The accuracy of the improved ResNet18_CBAM network model in the test set of 4 different parts of yak tenderloin, high rib, shank and brisket was 96.31%, which was 2.88% higher than the original network model. The recognition accuracy of the improved ResNet18_CBAM network model was compared with AlexNet, VGG11, ResNet34 and ResNet18 network models on different parts of yak test set. The improved ResNet18_CBAM network model had the highest accuracy. In order to verify the actual results of the improved ResNet18_CBAM network model on mobile phones, the test conducted in Xining beef and mutton wholesale market. In the actual scenario testing on the mobile end, a total of 54, 59, 51, and 57 yak tenderloin, high rib, shank and brisket samples were collected, respectively. The number of correctly identified samples and the number of incorrectly identified samples were counted respectively. Finally, the recognition accuracy of tenderloin, high rib, shank and brisket of yak reached 96.30%, 94.92%, 98.04% and 96.49%, respectively. The results showed that the improved ResNet18_CBAM network model could be used in practical applications for identifying different parts of yak meat and has achieved good results. [Conclusions] The research results can help ensure the food quality and safety of the yak industry, improve the quality and safety level of the yak industry, improve the yak trade efficiency, reduce the cost, and provide technical support for the intelligent development of the yak industry in the Qinghai-Tibet Plateau region.
Significance Agricultural environment is dynamic and variable, with numerous factors affecting the growth of animals and plants and complex interactions. There are numerous factors that affect the growth of all kinds of animals and plants. There is a close but complex correlation between these factors such as air temperature, air humidity, illumination, soil temperature, soil humidity, diseases, pests, weeds and etc. Thus, farmers need agricultural knowledge to solve production problems. With the rapid development of internet technology, a vast amount of agricultural information and knowledge is available on the internet. However, due to the lack of effective organization, the utilization rate of these agricultural information knowledge is relatively low.How to analyze and generate production knowledge or decision cases from scattered and disordered information is a big challenge all over the world. Agricultural knowledge intelligent service technology is a good way to resolve the agricultural data problems such as low rank, low correlation, and poor interpretability of reasoning. It is also the key technology to improving the comprehensive prediction and decision-making analysis capabilities of the entire agricultural production process. It can eliminate the information barriers between agricultural knowledge, farmers, and consumers, and is more conducive to improve the production and quality of agricultural products, provide effective information services. Progress The definition, scope, and technical application of agricultural knowledge intelligence services are introduced in this paper. The demand for agricultural knowledge services are analyzed combining with artificial intelligence technology. Agricultural knowledge intelligent service technologies such as perceptual recognition, knowledge coupling, and inference decision-making are conducted. The characteristics of agricultural knowledge services are analyzed and summarized from multiple perspectives such as industrial demand, industrial upgrading, and technological development. The development history of agricultural knowledge services is introduced. Current problems and future trends are also discussed in the agricultural knowledge services field. Key issues in agricultural knowledge intelligence services such as animal and plant state recognition in complex and uncertain environments, multimodal data association knowledge extraction, and collaborative reasoning in multiple agricultural application scenarios have been discussed. Combining practical experience and theoretical research, a set of intelligent agricultural situation analysis service framework that covers the entire life cycle of agricultural animals and plants and combines knowledge cases is proposed. An agricultural situation perception framework has been built based on satellite air ground multi-channel perception platform and Internet real-time data. Multimodal knowledge coupling, multimodal knowledge graph construction and natural language processing technology have been used to converge and manage agricultural big data. Through knowledge reasoning decision-making, agricultural information mining and early warning have been carried out to provide users with multi-scenario agricultural knowledge services. Intelligent agricultural knowledge services have been designed such as multimodal fusion feature extraction, cross domain knowledge unified representation and graph construction, and complex and uncertain agricultural reasoning and decision-making. An agricultural knowledge intelligent service platform composed of cloud computing support environment, big data processing framework, knowledge organization management tools, and knowledge service application scenarios has been built. Rapid assembly and configuration management of agricultural knowledge services could be provide by the platform. The application threshold of artificial intelligence technology in agricultural knowledge services could be reduced. In this case, problems of agricultural users can be solved. A novel method for agricultural situation analysis and production decision-making is proposed. A full chain of intelligent knowledge application scenario is constructed. The scenarios include planning, management, harvest and operations during the agricultural before, during and after the whole process. Conclusions and Prospects The technology trend of agricultural knowledge intelligent service is summarized in five aspects. (1) Multi-scale sparse feature discovery and spatiotemporal situation recognition of agricultural conditions. The application effects of small sample migration discovery and target tracking in uncertain agricultural information acquisition and situation recognition are discussed. (2) The construction and self-evolution of agricultural cross media knowledge graph, which uses robust knowledge base and knowledge graph to analyze and gather high-level semantic information of cross media content. (3) In response to the difficulties in tracing the origin of complex agricultural conditions and the low accuracy of comprehensive prediction, multi granularity correlation and multi-mode collaborative inversion prediction of complex agricultural conditions is discussed. (4) The large language model (LLM) in the agricultural field based on generative artificial intelligence. ChatGPT and other LLMs can accurately mine agricultural data and automatically generate questions through large-scale computing power, solving the problems of user intention understanding and precise service under conditions of dispersed agricultural data, multi-source heterogeneity, high noise, low information density, and strong uncertainty. In addition, the agricultural LLM can also significantly improve the accuracy of intelligent algorithms such as identification, prediction and decision-making by combining strong algorithms with Big data and super computing power. These could bring important opportunities for large-scale intelligent agricultural production. (5) The construction of knowledge intelligence service platforms and new paradigm of knowledge service, integrating and innovating a self-evolving agricultural knowledge intelligence service cloud platform. Agricultural knowledge intelligent service technology will enhance the control ability of the whole agricultural production chain. It plays a technical support role in achieving the transformation of agricultural production from "observing the sky and working" to "knowing the sky and working". The intelligent agricultural application model of "knowledge empowerment" provides strong support for improving the quality and efficiency of the agricultural industry, as well as for the modernization transformation and upgrading.
Significance Agricultural machinery serves as the fundamental support for implementing advanced agricultural production concepts. The key challenge for the future development of smart agriculture lies in how to enhance the design, manufacturing, operation, and maintenance of these machines to fully leverage their capabilities. To address this, the concept of the digital twin has emerged as an innovative approach that integrates various information technologies and facilitates the integration of virtual and real-world interactions. By providing a deeper understanding of agricultural machinery and its operational processes, the digital twin offers solutions to the complexity encountered throughout the entire lifecycle, from design to recycling. Consequently, it contributes to an all-encompassing enhancement of the quality of agricultural machinery operations, enabling them to better meet the demands of agricultural production. Nevertheless, despite its significant potential, the adoption of the digital twin for agricultural machinery is still at an early stage, lacking the necessary theoretical guidance and methodological frameworks to inform its practical implementation. Progress Drawing upon the successful experiences of the author's team in the digital twin for agricultural machinery, this paper presents an overview of the research progress made in digital twin. It covers three main areas: The digital twin in a general sense, the digital twin in agriculture, and the digital twin for agricultural machinery. The digital twin is conceptualized as an abstract notion that combines model-based system engineering and cyber-physical systems, facilitating the integration of virtual and real-world environments. This paper elucidates the relevant concepts and implications of digital twin in the context of agricultural machinery. It points out that the digital twin for agricultural machinery aims to leverage advanced information technology to create virtual models that accurately describe agricultural machinery and its operational processes. These virtual models act as a carrier, driven by data, to facilitate interaction and integration between physical agricultural machinery and their digital counterparts, consequently yielding enhanced value. Additionally, it proposes a comprehensive framework comprising five key components: Physical entities, virtual models, data and connectivity, system services, and business applications. Each component's functions operational mechanism, and organizational structure are elucidated. The development of the digital twin for agricultural machinery is still in its conceptual phase, and it will require substantial time and effort to gradually enhance its capabilities. In order to advance further research and application of the digital twin in this domain, this paper integrates relevant theories and practical experiences to propose an implementation plan for the digital twin for agricultural machinery. The macroscopic development process encompasses three stages: Theoretical exploration, practical application, and summarization. The specific implementation process entails four key steps: Intelligent upgrading of agricultural machinery, establishment of information exchange channels, construction of virtual models, and development of digital twin business applications. The implementation of digital twin for agricultural machinery comprises four stages: Pre-research, planning, implementation, and evaluation. The digital twin serves as a crucial link and bridge between agricultural machinery and the smart agriculture. It not only facilitates the design and manufacturing of agricultural machinery, aligning them with the realities of agricultural production and supporting the advancement of advanced manufacturing capabilities, but also enhances the operation, maintenance, and management of agricultural production to better meet practical requirements. This, in turn, expedites the practical implementation of smart agriculture. To fully showcase the value of the digital twin for agricultural machinery, this paper addresses the existing challenges in the design, manufacturing, operation, and management of agricultural machinery. It expounds the methods by which the digital twin can address these challenges and provides a technical roadmap for empowering the design, manufacturing, operation, and management of agricultural machinery through the use of the digital twin. In tackling the critical issue of leveraging the digital twin to enhance the operational quality of agricultural machinery, this paper presents two research cases focusing on high-powered tractors and large combine harvesters. These cases validate the feasibility of the digital twin in improving the quality of plowing operations for high-powered tractors and the quality of grain harvesting for large combine harvesters. Conclusions and Prospects This paper serves as a reference for the development of research on digital twin for agricultural machinery, laying a theoretical foundation for empowering smart agriculture and intelligent equipment with the digital twin. The digital twin provides a new approach for the transformation and upgrade of agricultural machinery, offering a new path for enhancing the level of agricultural mechanization and presenting new ideas for realizing smart agriculture. However, existing digital twin for agricultural machinery is still in its early stages, and there are a series of issues that need to be explored. It is necessary to involve more professionals from relevant fields to advance the research in this area.
Objective Deep learning is one of the most important technologies in the field of artificial intelligence, which has sparked a research boom in academic and engineering applications. It also shows strong application potential in remote sensing retrieval of geophysical parameters. The cross-disciplinary research is just beginning, and most deep learning applications in geosciences are still "black boxes", with most applications lacking physical significance, interpretability, and universality. In order to promote the application of artificial intelligence in geosciences and agriculture and cultivate interdisciplinary talents, a paradigm theory for geophysical parameter retrieval based on artificial intelligence coupled physics and statistical methods was proposed in this research. Methods The construction of the retrieval paradigm theory for geophysical parameters mainly included three parts: Firstly, physical logic deduction was performed based on the physical energy balance equation, and the inversion equation system was constructed theoretically which eliminated the ill conditioned problem of insufficient equations. Then, a fuzzy statistical method was constructed based on physical deduction. Representative solutions of physical methods were obtained through physical model simulation, and other representative solutions as the training and testing database for deep learning were obtained using multi-source data. Finally, deep learning achieved the goal of coupling physical and statistical methods through the use of representative solutions from physical and statistical methods as training and testing databases. Deep learning training and testing were aimed at obtaining curves of solutions from physical and statistical methods, thereby making deep learning physically meaningful and interpretable. Results and Discussions The conditions for determining the formation of a universal and physically interpretable paradigm were: (1) There must be a causal relationship between input and output variables (parameters); (2) In theory, a closed system of equations (with unknowns less than or equal to the number of equations) can be constructed between input and output variables (parameters), which means that the output parameters can be uniquely determined by the input parameters. If there is a strong causal relationship between input parameters (variables) and output parameters (variables), deep learning can be directly used for inversion. If there is a weak correlation between the input and output parameters, prior knowledge needs to be added to improve the inversion accuracy of the output parameters. The MODIS thermal infrared remote sensing data were used to retrieve land surface temperature, emissivity, near surface air temperature and atmospheric water vapor content as a case to prove the theory. When there was strong correlation between output parameters (LST and LSE) and input variables (BTi), using deep learning coupled with physical and statistical methods could obtain very high accuracy. When there was a weak correlation between the output parameter (NSAT) and the input variable (BTi), adding prior knowledge (LST and LSE) could improve the inversion accuracy and stability of the output parameter (NSAT). When there was partial strong correlation (WVC and BTi), adding prior knowledge (LST and LSE) could slightly improve accuracy and stability, but the error of prior knowledge (LST and LSE) may bring uncertainty, so prior knowledge could also be omitted. According to the inversion analysis of geophysical parameters of MODIS sensor thermal infrared band, bands 27, 28, 29 and 31 were more suitable for inversion of atmospheric water vapor content, and bands 28, 29, 31 and 32 were more suitable for inversion of surface temperature, Emissivity and near surface air temperature. If someone want to achieve the highest accuracy of four parameters, it was recommended to design the instrument with five bands (27, 28, 29, 31, 32) which were most suitable. If only four thermal infrared bands were designed, bands 27, 28, 31, and 32 should be given priority consideration. From the results of land surface temperature, emissivity, near surface air temperature and atmospheric water vapor content retrieved from MODIS data using this theory, it was not only more accurate than traditional methods, but also could reduce some bands, reduce satellite load and improve satellite life. Especially, this theoretical method overcomes the influence of the MODIS official algorithm (day/night algorithm) on sudden changes in surface types and long-term lack of continuous data, which leads to unstable accuracy of the inversion product. The analysis results showed that the proposed theory and conditions are feasible, and the accuracy and applicability were better than traditional methods. The theory and judgment conditions of geophysical parameter retrieval paradigms were also applicable for target recognition such as remote sensing classification, but it needed to be interpreted from a different perspective. For example, the feature information extracted by different convolutional kernels must be able to uniquely determine the target. Under satisfying with the conditions of paradigm theory, the inversion of geophysical parameters based on artificial intelligence is the best choice. Conclusions The geophysical parameter retrieval paradigm theory based on artificial intelligence proposed in this study can overcome the shortcomings of traditional retrieval methods, especially remote sensing parameter retrieval, which simplify the inversion process and improve the inversion accuracy. At the same time, it can optimize the design of satellite sensors. The proposal of this theory is of milestone significance in the history of geophysical parameter retrieval.
[Significance] Oil crops play a significant role in the food supply, as well as the important source of edible vegetable oils and plant proteins. Real-time, dynamic and large-scale monitoring of oil crop growth is essential in guiding agricultural production, stabilizing markets, and maintaining health. Previous studies have made a considerable progress in the yield simulation of staple crops in regional scale based on remote sensing methods, but the yield simulation of oil crops in regional scale is still poor as its complexity of the plant traits and structural characteristics. Therefore, it is urgently needed to study regional oil crop yield estimation based on remote sensing technology. [Progress] This paper summarized the content of remote sensing technology in oil crop monitoring from three aspects: backgrounds, progressions, opportunities and challenges. Firstly, significances and advantages of using remote sensing technology to estimate the of oil crops have been expounded. It is pointed out that both parameter inversion and crop area monitoring were the vital components of yield estimation. Secondly, the current situation of oil crop monitoring was summarized based on remote sensing technology from three aspects of remote sensing parameter inversion, crop area monitoring and yield estimation. For parameter inversion, it is specified that optical remote sensors were used more than other sensors in oil crops inversion in previous studies. Then, advantages and disadvantages of the empirical model and physical model inversion methods were analyzed. In addition, advantages and disadvantages of optical and microwave data were further illustrated from the aspect of oil crops structure and traits characteristics. At last, optimal choice on the data and methods were given in oil crop parameter inversion. For crop area monitoring, this paper mainly elaborated from two parts of optical and microwave remote sensing data. Combined with the structure of oil crops and the characteristics of planting areas, the researches on area monitoring of oil crops based on different types of remote sensing data sources were reviewed, including the advantages and limitations of different data sources in area monitoring. Then, two yield estimation methods were introduced: remote sensing yield estimation and data assimilation yield estimation. The phenological period of oil crop yield estimation, remote sensing data source and modeling method were summarized. Next, data assimilation technology was introduced, and it was proposed that data assimilation technology has great potential in oil crop yield estimation, and the assimilation research of oil crops was expounded from the aspects of assimilation method and grid selection. All of them indicate that data assimilation technology could improve the accuracy of regional yield estimation of oil crops. Thirdly, this paper pointed out the opportunities of remote sensing technology in oil crop monitoring, put forward some problems and challenges in crop feature selection, spatial scale determination and remote sensing data source selection of oil crop yield, and forecasted the development trend of oil crop yield estimation research in the future. [Conclusions and Prospects] The paper puts forward the following suggestions for the three aspects: (1) Regarding crop feature selection, when estimating yields for oil crops such as rapeseed and soybeans, which have active photosynthesis in siliques or pods, relying solely on canopy leaf area index (LAI) as the assimilation state variable for crop yield estimation may result in significant underestimation of yields, thereby impacting the accuracy of regional crop yield simulation. Therefore, it is necessary to consider the crop plant characteristics and the agronomic mechanism of yield formation through siliques or pods when estimating yields for oil crops. (2) In determining the spatial scale, some oil crops are distributed in hilly and mountainous areas with mixed land cover. Using regularized yield simulation grids may result in the confusion of numerous background objects, introducing additional errors and affecting the assimilation accuracy of yield estimation. This poses a challenge to yield estimation research. Thus, it is necessary to choose appropriate methods to divide irregular unit grids and determine the optimal scale for yield estimation, thereby improving the accuracy of yield estimation. (3) In terms of remote sensing data selection, the monitoring of oil crops can be influenced by crop structure and meteorological conditions. Depending solely on spectral data monitoring may have a certain impact on yield estimation results. It is important to incorporate radar off-nadir remote sensing measurement techniques to perceive the response relationship between crop leaves and siliques or pods and remote sensing data parameters. This can bridge the gap between crop characteristics and remote sensing information for crop yield simulation. This paper can serve as a valuable reference and stimulus for further research on regional yield estimation and growth monitoring of oil crops. It supplements existing knowledge and provides insightful considerations for enhancing the accuracy and efficiency of oil crop production monitoring and management.
[Significance] The scientific dataset of agricultural pests and diseases is the foundation for monitoring and warning of agricultural pests and diseases. It is of great significance for the development of agricultural pest control, and is an important component of developing smart agriculture. The quality of the dataset affecting the effectiveness of image recognition algorithms, with the discovery of the importance of deep learning technology in intelligent monitoring of agricultural pests and diseases. The construction of high-quality agricultural pest and disease datasets is gradually attracting attention from scholars in this field. In the task of image recognition, on one hand, the recognition effect depends on the improvement strategy of the algorithm, and on the other hand, it depends on the quality of the dataset. The same recognition algorithm learns different features in different quality datasets, so its recognition performance also varies. In order to propose a dataset evaluation index to measure the quality of agricultural pest and disease datasets, this article analyzes the existing datasets and takes the challenges faced in constructing agricultural pest and disease image datasets as the starting point to review the construction of agricultural pest and disease datasets. [Progress] Firstly, disease and pest datasets are divided into two categories: private datasets and public datasets. Private datasets have the characteristics of high annotation quality, high image quality, and a large number of inter class samples that are not publicly available. Public datasets have the characteristics of multiple types, low image quality, and poor annotation quality. Secondly, the problems faced in the construction process of datasets are summarized, including imbalanced categories at the dataset level, difficulty in feature extraction at the dataset sample level, and difficulty in measuring the dataset size at the usage level. These include imbalanced inter class and intra class samples, selection bias, multi-scale targets, dense targets, uneven data distribution, uneven image quality, insufficient dataset size, and dataset availability. The main reasons for the problem are analyzed by two key aspects of image acquisition and annotation methods in dataset construction, and the improvement strategies and suggestions for the algorithm to address the above issues are summarized. The collection devices of the dataset can be divided into handheld devices, drone platforms, and fixed collection devices. The collection method of handheld devices is flexible and convenient, but it is inefficient and requires high photography skills. The drone platform acquisition method is suitable for data collection in contiguous areas, but the detailed features captured are not clear enough. The fixed device acquisition method has higher efficiency, but the shooting scene is often relatively fixed. The annotation of image data is divided into rectangular annotation and polygonal annotation. In image recognition and detection, rectangular annotation is generally used more frequently. It is difficult to label images that are difficult to separate the target and background. Improper annotation can lead to the introduction of more noise or incomplete algorithm feature extraction. In response to the problems in the above three aspects, the evaluation methods are summarized for data distribution consistency, dataset size, and image annotation quality at the end of the article. [Conclusions and Prospects] The future research and development suggestions for constructing high-quality agricultural pest and disease image datasets based are proposed on the actual needs of agricultural pest and disease image recognition:(1) Construct agricultural pest and disease datasets combined with practical usage scenarios. In order to enable the algorithm to extract richer target features, image data can be collected from multiple perspectives and environments to construct a dataset. According to actual needs, data categories can be scientifically and reasonably divided from the perspective of algorithm feature extraction, avoiding unreasonable inter class and intra class distances, and thus constructing a dataset that meets task requirements for classification and balanced feature distribution. (2) Balancing the relationship between datasets and algorithms. When improving algorithms, consider the more sufficient distribution of categories and features in the dataset, as well as the size of the dataset that matches the model, to improve algorithm accuracy, robustness, and practicality. It ensures that comparative experiments are conducted on algorithm improvement under the same evaluation standard dataset, and improved the pest and disease image recognition algorithm. Research the correlation between the scale of agricultural pest and disease image data and algorithm performance, study the relationship between data annotation methods and algorithms that are difficult to annotate pest and disease images, integrate recognition algorithms for fuzzy, dense, occluded targets, and propose evaluation indicators for agricultural pest and disease datasets. (3) Enhancing the use value of datasets. Datasets can not only be used for research on image recognition, but also for research on other business needs. The identification, collection, and annotation of target images is a challenging task in the construction process of pest and disease datasets. In the process of collecting image data, in addition to collecting images, attention can be paid to the collection of surrounding environmental information and host information. This method is used to construct a multimodal agricultural pest and disease dataset, fully leveraging the value of the dataset. In order to focus researchers on business innovation research, it is necessary to innovate the organizational form of data collection, develop a big data platform for agricultural diseases and pests, explore the correlation between multimodal data, improve the accessibility and convenience of data, and provide efficient services for application implementation and business innovation.
[Objective] Rice blast is considered as the most destructive disease that threatens global rice production and causes severe economic losses worldwide. The detection of rice blast in an early manner plays an important role in resistance breeding and plant protection. At present, most studies on rice blast detection have been devoted to its symptomatic stage, while none of previous studies have used solar-induced chlorophyll fluorescence (SIF) to monitor rice leaf blast (RLB) at early stages. This research was conducted to investigate the early identification of RLB infected leaves based on solar-induced chlorophyll fluorescence at different leaf positions. [Methods] Greenhouse experiments and field trials were conducted separately in Nanjing and Nantong in July and August, 2021, in order to record SIF data of the top 1th to 4th leaves of rice plants at jointing and heading stages with an Analytical Spectral Devices (ASD) spectrometer coupled with a FluoWat leaf clip and a halogen lamp. At the same time, the disease severity levels of the measured samples were manually collected according to the GB/T 15790-2009 standard. After the continuous wavelet transform (CWT) of SIF spectra, separability assessment and feature selection were applied to SIF spectra. Wavelet features sensitive to RLB were extracted, and the sensitive features and their identification accuracy of infected leaves for different leaf positions were compared. Finally, RLB identification models were constructed based on linear discriminant analysis (LDA). [Results and Discussion] The results showed that the upward and downward SIF in the far-red region of infected leaves at each leaf position were significantly higher than those of healthy leaves. This may be due to the infection of the fungal pathogen Magnaporthe oryzae, which may have destroyed the chloroplast structure, and ultimately inhibited the primary reaction of photosynthesis. In addition, both the upward and downward SIF in the red region and the far-red region increased with the decrease of leaf position. The sensitive wavelet features varied by leaf position, while most of them were distributed in the steep slope of the SIF spectrum and wavelet scales 3, 4 and 5. The sensitive features of the top 1th leaf were mainly located at 665-680 nm, 755-790 nm and 815-830 nm. For the top 2th leaf, the sensitive features were mainly found at 665-680 nm and 815-830 nm. For the top 3th one, most of the sensitive features lay at 690 nm, 755-790 nm and 815-830 nm, and the sensitive bands around 690 nm were observed. The sensitive features of the top 4th leaf were primarily located at 665-680 nm, 725 nm and 815-830 nm, and the sensitive bands around 725 nm were observed. The wavelet features of the common sensitive region (665-680 nm), not only had physiological significance, but also coincided with the chlorophyll absorption peak that allowed for reasonable spectral interpretation. There were differences in the accuracy of RLB identification models at different leaf positions. Based on the upward and downward SIF, the overall accuracies of the top 1th leaf were separately 70% and 71%, which was higher than other leaf positions. As a result, the top 1th leaf was an ideal indicator leaf to diagnose RLB in the field. The classification accuracy of SIF wavelet features were higher than the original SIF bands. Based on CWT and feature selection, the overall accuracy of the upward and downward optimal features of the top 1th to 4th leaves reached 70.13%、63.70%、64.63%、64.53% and 70.90%、63.12%、62.00%、64.02%, respectively. All of them were higher than the canopy monitoring feature F760, whose overall accuracy was 69.79%, 61.31%, 54.41%, 61.33% and 69.99%, 58.79%, 54.62%, 60.92%, respectively. This may be caused by the differences in physiological states of the top four leaves. In addition to RLB infection, the SIF data of some top 3th and top 4th leaves may also be affected by leaf senescence, while the SIF data of top 1th leaf, the latest unfolding leaf of rice plants was less affected by other physical and chemical parameters. This may explain why the top 1th leaf responded to RLB earlier than other leaves. The results also showed that the common sensitive features of the four leaf positions were also concentrated on the steep slope of the SIF spectrum, with better classification performance around 675 and 815 nm. The classification accuracy of the optimal common features, ↑WF832,3 and ↓WF809,3, reached 69.45%, 62.19%, 60.35%, 63.00% and 69.98%, 62.78%, 60.51%, 61.30% for the top 1th to top 4th leaf positions, respectively. The optimal common features, ↑WF832,3 and ↓WF809,3, were both located in wavelet scale 3 and 800-840nm, which may be related to the destruction of the cell structure in response to Magnaporthe oryzae infection. [Conclusions] In this study, the SIF spectral response to RLB was revealed, and the identification models of the top 1th leaf were found to be most precise among the top four leaves. In addition, the common wavelet features sensitive to RLB, ↑WF832,3 and ↓WF809,3, were extracted with the identification accuracy of 70%. The results proved the potential of CWT and SIF for RLB detection, which can provide important reference and technical support for the early, rapid and non-destructive diagnosis of RLB in the field.
[Objective] Wine grapes are severely affected by leafroll disease, which affects their growth, and reduces the quality of the color, taste, and flavor of wine. Timely and accurate diagnosis of leafroll disease severity is crucial for preventing and controlling the disease, improving the wine grape fruit quality and wine-making potential. Unmanned aerial vehicle (UAV) remote sensing technology provides high-resolution images of wine grape vineyards, which can capture the features of grapevine canopies with different levels of leafroll disease severity. Deep learning networks extract complex and high-level features from UAV remote sensing images and perform fine-grained classification of leafroll disease infection severity. However, the diagnosis of leafroll disease severity is challenging due to the imbalanced data distribution of different infection levels and categories in UAV remote sensing images. [Method] A novel method for diagnosing leafroll disease severity was developed at a canopy scale using UAV remote sensing technology and deep learning. The main challenge of this task was the imbalanced data distribution of different infection levels and categories in UAV remote sensing images. To address this challenge, a method that combined deep learning fine-grained classification and generative adversarial networks (GANs) was proposed. In the first stage, the GANformer, a Transformer-based GAN model was used, to generate diverse and realistic virtual canopy images of grapevines with different levels of leafroll disease severity. To further analyze the image generation effect of GANformer. The t-distributed stochastic neighbor embedding (t-SNE) to visualize the learned features of real and simulated images. In the second stage, the CA-Swin Transformer, an improved image classification model based on the Swin Transformer and channel attention mechanism was used, to classify the patch images into different classes of leafroll disease infection severity. CA-Swin Transformer could also use a self-attention mechanism to capture the long-range dependencies of image patches and enhance the feature representation of the Swin Transformer model by adding a channel attention mechanism after each Transformer layer. The channel attention (CA) mechanism consisted of two fully connected layers and an activation function, which could extract correlations between different channels and amplify the informative features. The ArcFace loss function and instance normalization layer was also used to enhance the fine-grained feature extraction and downsampling ability for grapevine canopy images. The UAV images of wine grape vineyards were collected and processed into orthomosaic images. They labeled into three categories: healthy, moderate infection, and severe infection using the in-field survey data. A sliding window method was used to extract patch images and labels from orthomosaic images for training and testing. The performance of the improved method was compared with the baseline model using different loss functions and normalization methods. The distribution of leafroll disease severity was mapped in vineyards using the trained CA-Swin Transformer model. [Results and Discussions] The experimental results showed that the GANformer could generate high-quality virtual canopy images of grapevines with an FID score of 93.20. The images generated by GANformer were visually very similar to real images and could produce images with different levels of leafroll disease severity. The T-SNE visualization showed that the features of real and simulated images were well clustered and separated in two-dimensional space, indicating that GANformer learned meaningful and diverse features, which enriched the image dataset. Compared to CNN-based deep learning models, Transformer-based deep learning models had more advantages in diagnosing leafroll disease infection. Swin Transformer achieved an optimal accuracy of 83.97% on the enhanced dataset, which was higher than other models such as GoogLeNet, MobileNetV2, NasNet Mobile, ResNet18, ResNet50, CVT, and T2TViT. It was found that replacing the cross entropy loss function with the ArcFace loss function improved the classification accuracy by 1.50%, and applying instance normalization instead of layer normalization further improved the accuracy by 0.30%. Moreover, the proposed channel attention mechanism, named CA-Swin Transformer, enhanced the feature representation of the Swin Transformer model, achieved the highest classification accuracy on the test set, reaching 86.65%, which was 6.54% higher than using the Swin Transformer on the original test dataset. By creating a distribution map of leafroll disease severity in vineyards, it was found that there was a certain correlation between leafroll disease severity and grape rows. Areas with a larger number of severe leafroll diseases caused by Cabernet Sauvignon were more prone to have missing or weak plants. [Conclusions] A novel method for diagnosing grapevine leafroll disease severity at a canopy scale using UAV remote sensing technology and deep learning was proposed. This method can generate diverse and realistic virtual canopy images of grapevines with different levels of leafroll disease severity using GANformer, and classify them into different classes using CA-Swin Transformer. This method can also map the distribution of leafroll disease severity in vineyards using a sliding window method, and provides a new approach for crop disease monitoring based on UAV remote sensing technology.
[Objective] Wheat, as one of the major global food crops, plays a key role in food production and food supply. Different influencing factors can lead to different types of wheat lodging, e.g., root lodging may be due to improper use of fertilizers. While stem lodging is mostly due to harsh environments, different types of wheat lodging can have different impacts on yield and quality. The aim of this study was to categorize the types of wheat lodging by unmanned aerial vehicle (UAV) image detection and to investigate the effect of UAV flight altitude on the classification performance. [Methods] Three UAV flight altitudes (15, 45, and 91 m) were set to acquire images of wheat test fields. The main research methods contained three parts: an automatic segmentation algorithm, wheat classification model selection, and an improved classification model based on EfficientNetV2-C. In the first part, the automatic segmentation algorithm was used to segment the UAV to acquire the wheat test field at three different heights and made it into the training dataset needed for the classification model. The main steps were first to preprocess the original wheat test field images acquired by the UAV through scaling, skew correction, and other methods to save computation time and improve segmentation accuracy. Subsequently, the pre-processed image information was analyzed, and the green part of the image was extracted using the super green algorithm, which was binarized and combined with the edge contour extraction algorithm to remove the redundant part of the image to extract the region of interest, so that the image was segmented for the first time. Finally, the idea of accumulating pixels to find sudden value added was used to find the segmentation coordinates of two different sizes of wheat test field in the image, and the region of interest of the wheat test field was segmented into a long rectangle and a short rectangle test field twice, so as to obtain the structural parameters of different sizes of wheat test field and then to generate the dataset of different heights. In the second part, four machine learning classification models of support vector machine (SVM), K nearest neighbor (KNN), decision tree (DT), and naive bayes (NB), and two deep learning classification models (ResNet101 and EfficientNetV2) were selected. Under the unimproved condition, six classification models were utilized to classify the images collected from three UAVs at different flight altitudes, respectively, and the optimal classification model was selected for improvement. In the third part, an improved model, EfficientNetV2-C, with EfficientNetV2 as the base model, was proposed to classify and recognized the lodging type of wheat in test field images. The main improvement points were attention mechanism improvement and loss function improvement. The attention mechanism was to replace the original model squeeze and excitation (SE) with coordinate attention (CA), which was able to embed the position information into the channel attention, aggregate the features along the width and height directions, respectively, during feature extraction, and capture the long-distance correlation in the width direction while retaining the long-distance correlation in the length direction, accurate location information, enhancing the feature extraction capability of the network in space. The loss function was replaced by class-balanced focal loss (CB-Focal Loss), which could assign different loss weights according to the number of valid samples in each class when targeting unbalanced datasets, effectively solving the impact of data imbalance on the classification accuracy of the model. [Results and Discussions] Four machine learning classification results: SVM average classification accuracy was 81.95%, DT average classification accuracy was 79.56%, KNN average classification accuracy was 59.32%, and NB average classification accuracy was 59.48%. The average classification accuracy of the two deep learning models, ResNet101 and EfficientNetV2, was 78.04%, and the average classification accuracy of ResNet101 was 81.61%. Comparing the above six classification models, the EfficientNetV2 classification model performed optimally at all heights. And the improved EfficientNetV2-C had an average accuracy of 90.59%, which was 8.98% higher compared to the average accuracy of EfficientNetV2. The SVM classification accuracies of UAVs at three flight altitudes of 15, 45, and 91 m were 81.33%, 83.57%, and 81.00%, respectively, in which the accuracy was the highest when the altitude was 45 m, and the classification results of the SVM model values were similar to each other, which indicated that the imbalance of the input data categories would not affect the model's classification effect, and the SVM classification model was able to solve the problem of high dimensionality of the data efficiently and had a good performance for small and medium-sized data sets. The SVM classification model could effectively solve the problem of the high dimensionality of data and had a better classification effect on small and medium-sized datasets. For the deep learning classification model, however, as the flight altitude increases from 15 to 91 m, the classification performance of the deep learning model decreased due to the loss of image feature information. Among them, the classification accuracy of ResNet101 decreased from 81.57% to 78.04%, the classification accuracy of EfficientNetV2 decreased from 84.40% to 81.61%, and the classification accuracy of EfficientNetV2-C decreased from 97.65% to 90.59%. The classification accuracy of EfficientNetV2-C at each of the three altitudes. The difference between the values of precision, recall, and F1-Score results of classification was small, which indicated that the improved model in this study could effectively solve the problems of unbalanced model classification results and poor classification effect caused by data imbalance. [Conclusions] The improved EfficientNetV2-C achieved high accuracy in wheat lodging type detection, which provides a new solution for wheat lodging early warning and crop management and is of great significance for improving wheat production efficiency and sustainable agricultural development.
[Objective] Lodging constitutes a severe crop-related catastrophe, resulting in a reduction in photosynthesis intensity, diminished nutrient absorption efficiency, diminished crop yield, and compromised crop quality. The utilization of unmanned aerial vehicles (UAV) to acquire agricultural remote sensing imagery, despite providing high-resolution details and clear indications of crop lodging, encounters limitations related to the size of the study area and the duration of the specific growth stages of the plants. This limitation hinders the acquisition of an adequate quantity of low-altitude remote sensing images of wheat fields, thereby detrimentally affecting the performance of the monitoring model. The aim of this study is to explore a method for precise segmentation of lodging areas in limited crop growth periods and research areas. [Methods] Compared to the images captured at lower flight altitudes, the images taken by UAVs at higher altitudes cover a larger area. Consequently, for the same area, the number of images taken by UAVs at higher altitudes is fewer than those taken at lower altitudes. However, the training of deep learning models requires huge amount supply of images. To make up the issue of insufficient quantity of high-altitude UAV-acquired images for the training of the lodging area monitoring model, a transfer learning strategy was proposed. In order to verify the effectiveness of the transfer learning strategy, based on the Swin-Transformer framework, the control model, hybrid training model and transfer learning training model were obtained by training UAV images in 4 years (2019, 2020, 2021, 2023)and 3 study areas(Shucheng, Guohe, Baihe) under 2 flight altitudes (40 and 80 m). To test the model's performance, a comparative experimental approach was adopted to assess the accuracy of the three models for segmenting 80 m altitude images. The assessment relied on five metrics: intersection of union (IoU), accuracy, precision, recall, and F1-score. [Results and Discussions] The transfer learning model shows the highest accuracy in lodging area detection. Specifically, the mean IoU, accuracy, precision, recall, and F1-score achieved 85.37%, 94.98%, 91.30%, 92.52% and 91.84%, respectively. Notably, the accuracy of lodging area detection for images acquired at a 40 m altitude surpassed that of images captured at an 80 m altitude when employing a training dataset composed solely of images obtained at the 40 m altitude. However, when adopting mixed training and transfer learning strategies and augmenting the training dataset with images acquired at an 80 m altitude, the accuracy of lodging area detection for 80 m altitude images improved, inspite of the expense of reduced accuracy for 40 m altitude images. The performance of the mixed training model and the transfer learning model in lodging area detection for both 40 and 80 m altitude images exhibited close correspondence. In a cross-study area comparison of the mean values of model evaluation indices, lodging area detection accuracy was slightly higher for images obtained in Baihu area compared to Shucheng area, while accuracy for images acquired in Shucheng surpassed that of Guohe. These variations could be attributed to the diverse wheat varieties cultivated in Guohe area through drill seeding. The high planting density of wheat in Guohe resulted in substantial lodging areas, accounting for 64.99% during the late mature period. The prevalence of semi-lodging wheat further exacerbated the issue, potentially leading to misidentification of non-lodging areas. Consequently, this led to a reduction in the recall rate (mean recall for Guohe images was 89.77%, which was 4.88% and 3.57% lower than that for Baihu and Shucheng, respectively) and IoU (mean IoU for Guohe images was 80.38%, which was 8.80% and 3.94% lower than that for Baihu and Shucheng, respectively). Additionally, the accuracy, precision, and F1-score for Guohe were also lower compared to Baihu and Shucheng. [Conclusions] This study inspected the efficacy of a strategy aimed at reducing the challenges associated with the insufficient number of high-altitude images for semantic segmentation model training. By pre-training the semantic segmentation model with low-altitude images and subsequently employing high-altitude images for transfer learning, improvements of 1.08% to 3.19% were achieved in mean IoU, accuracy, precision, recall, and F1-score, alongside a notable mean weighted frame rate enhancement of 555.23 fps/m2. The approach proposed in this study holds promise for improving lodging monitoring accuracy and the speed of image segmentation. In practical applications, it is feasible to leverage a substantial quantity of 40 m altitude UAV images collected from diverse study areas including various wheat varieties for pre-training purposes. Subsequently, a limited set of 80 m altitude images acquired in specific study areas can be employed for transfer learning, facilitating the development of a targeted lodging detection model. Future research will explore the utilization of UAV images captured at even higher flight altitudes for further enhancing lodging area detection efficiency.
[Objective] The width of maize stalks is an important indicator affecting the lodging resistance of maize. The measurement of maize stalk width has many problems, such as cumbersome manual collection process and large errors in the accuracy of automatic equipment collection and recognition, and it is of great application value to study a method for in-situ detection and high-precision identification of maize stalk width. [Methods] The ZED2i binocular camera was used and fixed in the field to obtain real-time pictures from the left and right sides of maize stalks together. The picture acquisition system was based on the NVIDIA Jetson TX2 NX development board, which could achieve timed shooting of both sides view of the maize by setting up the program. A total of maize original images were collected and a dataset was established. In order to observe more features in the target area from the image and provide assistance to improve model training generalization ability, the original images were processed by five processing methods: image saturation, brightness, contrast, sharpness and horizontal flipping, and the dataset was expanded to 3500 images. YOLOv8 was used as the original model for identifying maize stalks from a complex background. The coordinate attention (CA) attention mechanism can bring huge gains to downstream tasks on the basis of lightweight networks, so that the attention block can capture long-distance relationships in one direction while retaining spatial information in the other direction, so that the position information can be saved in the generated attention map to focus on the area of interest and help the network locate the target better and more accurately. By adding the CA module multiple times, the CA module was fused with the C2f module in the original Backbone, and the Bottleneck in the original C2f module was replaced by the CA module, and the C2fCA network module was redesigned. Replacing the loss function Efficient IoU Loss(EIoU) splits the loss term of the aspect ratio into the difference between the predicted width and height and the width and height of the minimum outer frame, which accelerated the convergence of the prediction box, improved the regression accuracy of the prediction box, and further improved the recognition accuracy of maize stalks. The binocular camera was then calibrated so that the left and right cameras were on the same three-dimensional plane. Then the three-dimensional reconstruction of maize stalks, and the matching of left and right cameras recognition frames was realized through the algorithm, first determine whether the detection number of recognition frames in the two images was equal, if not, re-enter the binocular image. If they were equal, continue to judge the coordinate information of the left and right images, the width and height of the bounding box, and determine whether the difference was less than the given Ta. If greater than the given Ta, the image was re-imported; If it was less than the given Ta, the confidence level of the recognition frame of the image was determined whether it was less than the given Tb. If greater than the given Tb, the image is re-imported; If it is less than the given Tb, it indicates that the recognition frame is the same maize identified in the left and right images. If the above conditions were met, the corresponding point matching in the binocular image was completed. After the three-dimensional reconstruction of the binocular image, the three-dimensional coordinates (Ax, Ay, Az) and (Bx, By, Bz) in the upper left and upper right corners of the recognition box under the world coordinate system were obtained, and the distance between the two points was the width of the maize stalk. Finally, a comparative analysis was conducted among the improved YOLOv8 model, the original YOLOv8 model, faster region convolutional neural networks (Faster R-CNN), and single shot multiBox detector (SSD)to verify the recognition accuracy and recognition accuracy of the model. [Results and Discussions] The precision rate (P)、recall rate (R)、average accuracy mAP0.5、average accuracy mAP0.5:0.95 of the improved YOLOv8 model reached 96.8%、94.1%、96.6% and 77.0%. Compared with YOLOv7, increased by 1.3%、1.3%、1.0% and 11.6%, compared with YOLOv5, increased by 1.8%、2.1%、1.2% and 15.8%, compared with Faster R-CNN, increased by 31.1%、40.3%、46.2%、and 37.6%, and compared with SSD, increased by 20.6%、23.8%、20.9% and 20.1%, respectively. Respectively, and the linear regression coefficient of determination R2, root mean square error RMSE and mean absolute error MAE were 0.373, 0.265 cm and 0.244 cm, respectively. The method proposed in the research can meet the requirements of actual production for the measurement accuracy of maize stalk width. [Conclusions] In this study, the in-situ recognition method of maize stalk width based on the improved YOLOv8 model can realize the accurate in-situ identification of maize stalks, which solves the problems of time-consuming and laborious manual measurement and poor machine vision recognition accuracy, and provides a theoretical basis for practical production applications.
[Objective] The root system is an important component of plant composition, and its growth and development are crucial for plants. Root image segmentation is an important method for obtaining root phenotype information and analyzing root growth patterns. Research on root image segmentation still faces difficulties, because of the noise and image quality limitations, the intricate and diverse soil environment, and the ineffectiveness of conventional techniques. This paper proposed a multi-scale feature extraction root segmentation algorithm that combined data augmentation and transfer learning to enhance the generalization and universality of the root image segmentation models in order to increase the speed, accuracy, and resilience of root image segmentation. [Methods] Firstly, the experimental datasets were divided into a single dataset and a mixed dataset. The single dataset acquisition was obtained from the experimental station of Hebei Agricultural University in Baoding city. Additionally, a self-made RhizoPot device was used to collect images with a resolution pixels of 10,200×14,039, resulting in a total of 600 images. In this experiment, 100 sheets were randomly selected to be manually labeled using Adobe Photoshop CC2020 and segmented into resolution pixels of 768×768, and divided into training, validation, and test sets according to 7:2:1. To increase the number of experimental samples, an open source multi-crop mixed dataset was obtained in the network as a supplement, and it was reclassified into training, validation, and testing sets. The model was trained using the data augmentation strategy, which involved performing data augmentation operations at a set probability of 0.3 during the image reading phase, and each method did not affect the other. When the probability was less than 0.3, changes would be made to the image. Specific data augmentation methods included changing image attributes, randomly cropping, rotating, and flipping those images. The UNet structure was improved by designing eight different multi-scale image feature extraction modules. The module structure mainly included two aspects: Image convolution and feature fusion. The convolution improvement included convolutional block attention module (CBAM), depthwise separable convolution (DP Conv), and convolution (Conv). In terms of feature fusion methods, improvements could be divided into concatenation and addition. Subsequently, ablation tests were conducted based on a single dataset, data augmentation, and random loading of model weights, and the optimal multi-scale feature extraction module was selected and compared with the original UNet. Similarly, a single dataset, data augmentation, and random loading of model weights were used to compare and validate the advantages of the improved model with the PSPNet, SegNet, and DeeplabV3Plus algorithms. The improved model used pre-trained weights from a single dataset to load and train the model based on mixed datasets and data augmentation, further improving the model's generalization ability and root segmentation ability. [Results and Discussions] The results of the ablation tests indicated that Conv_ 2+Add was the best improved algorithm. Compared to the original UNet, the mIoU, mRecall, and root F1 values of the model increased by 0.37%, 0.99%, and 0.56%, respectively. And, comparative experiments indicate Unet+Conv_2+Add model was superior to the PSPNet, SegNet, and DeeplabV3Plus models, with the best evaluation results. And the values of mIoU, mRecall, and the harmonic average of root F1 were 81.62%, 86.90%, and 77.97%, respectively. The actual segmented images obtained by the improved model were more finely processed at the root boundary compared to other models. However, for roots with deep color and low contrast with soil particles, the improved model could only achieve root recognition and the recognition was sparse, sacrificing a certain amount of information extraction ability. This study used the root phenotype evaluation software Rhizovision to analyze the root images of the Unet+Conv_2+Add improved model, PSPNet, SegNet, and DeeplabV3Plu, respectively, to obtain the values of the four root phenotypes (total root length, average diameter, surface area, and capacity), and the results showed that the average diameter and surface area indicator values of the improved model, Unet+Conv_2+Add had the smallest differences from the manually labeled indicator values and the SegNet indicator values for the two indicators. Total root length and volume were the closest to those of the manual labeling. The results of transfer learning experiments proved that compared with ordinary training, the transfer training of the improved model UNet+Conv_2+Add increased the IoU value of the root system by 1.25%. The Recall value of the root system was increased by 1.79%, and the harmonic average value of F1 was increased by 0.92%. Moreover, the overall convergence speed of the model was fast. Compared with regular training, the transfer training of the original UNet improved the root IoU by 0.29%, the root Recall by 0.83%, and the root F1 value by 0.21%, which indirectly confirmed the effectiveness of transfer learning. [Conclusions] The multi-scale feature extraction strategy proposed in this study can accurately and efficiently segment roots, and further improve the model's generalization ability using transfer learning methods, providing an important research foundation for crop root phenotype research.
[Objective] Wheat serves as the primary source of dietary carbohydrates for the human population, supplying 20% of the required caloric intake. Currently, the primary objective of wheat breeding is to develop wheat varieties that exhibit both high quality and high yield, ensuring an overall increase in wheat production. Additionally, the consideration of phenotype parameters, such as grain length and width, holds significant importance in the introduction, screening, and evaluation of germplasm resources. Notably, a noteworthy positive association has been observed between grain size, grain shape, and grain weight. Simultaneously, within the scope of wheat breeding, the occurrence of inadequate harvest and storage practices can readily result in damage to wheat grains, consequently leading to a direct reduction in both emergence rate and yield. In essence, the integrity of wheat grains directly influences the wheat breeding process. Nevertheless, distinguishing between intact and damaged grains remains challenging due to the minimal disparities in certain characteristics, thereby impeding the accurate identification of damaged wheat grains through manual means. Consequently, this study aims to address this issue by focusing on the detection of wheat kernel integrity and completing the attainment of grain phenotype parameters. [Methods] This study presented an enhanced approach for addressing the challenges of low detection accuracy, unclear segmentation of wheat grain contour, and missing detection. The proposed strategy involves utilizing the Cascade Mask R-CNN model and replacing the backbone network with ResNeXt to mitigate gradient dispersion and minimize the model's parameter count. Furthermore, the inclusion of Mish as an activation function enhanced the efficiency and versatility of the detection model. Additionally, a multilayer convolutional structure was introduced in the detector to thoroughly investigate the latent features of wheat grains. The Soft-NMS algorithm was employed to identify the candidate frame and achieve accurate segmentation of the wheat kernel adhesion region. Additionally, the ImCascade R-CNN model was developed. Simultaneously, to address the issue of low accuracy in obtaining grain contour parameters due to disordered grain arrangement, a grain contour-based algorithm for parameter acquisition was devised. Wheat grain could be approximated as an oval shape, and the grain edge contour could be obtained according to the mask, the distance between the farthest points could be iteratively obtained as the grain length, and the grain width could be obtained according to the area. Ultimately, a method for wheat kernel phenotype identification was put forth. The ImCascade R-CNN model was utilized to analyze wheat kernel images, extracting essential features and determining the integrity of the kernels through classification and boundary box regression branches. The mask generation branch was employed to generate a mask map for individual wheat grains, enabling segmentation of the grain contours. Subsequently, the number of grains in the image was determined, and the length and width parameters of the entire wheat grain were computed. [Results and Discussions] In the experiment on wheat kernel phenotype recognition, a comparison and improvement were conducted on the identification results of the Cascade Mask R-CNN model and the ImCascade R-CNN model across various modules. Additionally, the efficacy of the model modification scheme was verified. The comparison of results between the Cascade Mask R-CNN model and the ImCascade R-CNN model served to validate the proposed model's ability to significantly decrease the missed detection rate. The effectiveness and advantages of the ImCascade R-CNN model were verified by comparing its loss value, P-R value, and mAP_50 value with those of the Cascade Mask R-CNN model. In the context of wheat grain identification and segmentation, the detection results of the ImCascade R-CNN model were compared to those of the Cascade Mask R-CNN and Deeplabv3+ models. The comparison confirmed that the ImCascade R-CNN model exhibited superior performance in identifying and locating wheat grains, accurately segmenting wheat grain contours, and achieving an average accuracy of 90.2% in detecting wheat grain integrity. These findings serve as a foundation for obtaining kernel contour parameters. The grain length and grain width exhibited average error rates of 2.15% and 3.74%, respectively, while the standard error of the aspect ratio was 0.15. The statistical analysis and fitting of the grain length and width, as obtained through the proposed wheat grain shape identification method, yielded determination coefficients of 0.9351 and 0.8217, respectively. These coefficients demonstrated a strong agreement with the manually measured values, indicating that the method is capable of meeting the demands of wheat seed testing and providing precise data support for wheat breeding. [Conclusions] The findings of this study can be utilized for the rapid and precise detection of wheat grain integrity and the acquisition of comprehensive grain contour data. In contrast to current wheat kernel recognition technology, this research capitalizes on enhanced grain contour segmentation to furnish data support for the acquisition of wheat kernel contour parameters. Additionally, the refined contour parameter acquisition algorithm effectively mitigates the impact of disordered wheat kernel arrangement, resulting in more accurate parameter data compared to existing kernel appearance detectors available in the market, providing data support for wheat breeding and accelerating the cultivation of high-quality and high-yield wheat varieties.
[Objective] To construct the 3D point cloud model of green plants a large number of clear images are needed. Due to the limitation of the depth of field of the lens, part of the image would be out of focus when the green plant image with a large depth of field is collected, resulting in problems such as edge blurring and texture detail loss, which greatly affects the accuracy of the 3D point cloud model. However, the existing processing algorithms are difficult to take into account both processing quality and processing speed, and the actual effect is not ideal. The purpose of this research is to improve the quality of the fused image while taking into account the processing speed. [Methods] A plant image fusion method based on non-subsampled shearlet transform (NSST) based parameter-adaptive dual channel pulse-coupled neural network (PADC-PCNN) and stationary wavelet transform (SWT) was proposed. Firstly, the RGB image of the plant was separated into three color channels, and the G channel with many features such as texture details was decomposed by NSST in four decomposition layers and 16 directions, which was divided into one group of low frequency subbands and 64 groups of high frequency subbands. The low frequency subband used the gradient energy fusion rule, and the high frequency subband used the PADC-PCNN fusion rule. In addition, the weighting of the eight-neighborhood modified Laplacian operator was used as the link strength of the high-frequency fusion part, which enhanced the fusion effect of the detailed features. At the same time, for the R and B channels with more contour information and background information, a SWT with fast speed and translation invariance was used to suppress the pseudo-Gibbs effect. Through the high-precision and high-stability multi-focal length plant image acquisition system, 480 images of 8 experimental groups were collected. The 8 groups of data were divided into an indoor light group, natural light group, strong light group, distant view group, close view group, overlooking group, red group, and yellow group. Meanwhile, to study the application range of the algorithm, the focus length of the collected clear plant image was used as the reference (18 mm), and the image acquisition was adjusted four times before and after the step of 1.5 mm, forming the multi-focus experimental group. Subjective evaluation and objective evaluation were carried out for each experimental group to verify the performance of the algorithm. Subjective evaluation was analyzed through human eye observation, detail comparison, and other forms, mainly based on the human visual effect. The image fusion effect of the algorithm was evaluated using four commonly used objective indicators, including average gradient (AG), spatial frequency (SF), entropy (EN), and standard deviation (SD). [Results and Discussions] The proposed PADC-PCNN-SWT algorithm and other five algorithms of common fast guided filtering algorithm (FGF), random walk algorithm (RW), non-subsampled shearlet transform based PCNN (NSST-PCNN) algorithm, SWT algorithm and non-subsampled shearlet transform based parameter-adaptive dual-channel pulse-coupled neural network (NSST-PADC) and were compared. In the objective evaluation data except for the red group and the yellow group, each index of the PADC-PCNN-SWT algorithm was second only to the NSST-PADC algorithm, but the processing speed was 200.0% higher than that of the NSST-PADC algorithm on average. At the same time, compared with the FDF, RW, NSST-PCNN, and SWT algorithms, the PADC-PCN -SWT algorithm improved the clarity index by 5.6%, 8.1%, 6.1%, and 17.6%, respectively, and improved the spatial frequency index by 2.9%, 4.8%, 7.1%, and 15.9%, respectively. However, the difference between the two indicators of information entropy and standard deviation was less than 1%, and the influence was ignored. In the yellow group and the red group, the fusion quality of the non-green part of the algorithm based on PADC-PCNN-SWT was seriously degraded. Compared with other algorithms, the sharpness index of the algorithm based on PADC-PCNN-SWT decreased by an average of 1.1%, and the spatial frequency decreased by an average of 5.1%. However, the indicators of the green part of the fused image were basically consistent with the previous several groups of experiments, and the fusion effect was good. Therefore, the algorithm based on PADC-PCNN-SWT only had a good fusion effect on green plants. Finally, by comparing the quality of four groups of fused images with different focal length ranges, the results showed that the algorithm based on PADC-PCNN-SWT had a better contour and color restoration effect for out-of-focus images in the range of 15-21 mm, and the focusing range based on PADC-PCNN-SWT was about 6 mm. [Conclusions] The multi-focal length image fusion algorithm based on PADC-PCNN-SWT achieved better detail fusion performance and higher image fusion efficiency while ensuring fusion quality, providing high-quality data, and saving a lot of time for building 3D point cloud model of green plants.
[Objective] JUNCAO, a perennial herbaceous plant that can be used as medium for cultivating edible and medicinal fungi. It has important value for promotion, but the problem of overwintering needs to be overcome when planting in the temperate zone. Low-temperature stress can adversely impact the growth of JUNCAO plants. Malondialdehyde (MDA) is a degradation product of polyunsaturated fatty acid peroxides, which can serve as a useful diagnostic indicator for studying plant growth dynamics. Because the more severe the damage caused by low temperature stress on plants, the higher their MDA content. Therefore, the detection of MDA content can provide instruct for low-temperature stress diagnosis and JUNCAO plants breeding. With the development of optical sensors and machine learning technologies, visible/near-infrared spectroscopy technology combined with algorithmic models has great potential in rapid, non-destructive and high-throughput inversion of MDA content and evaluation of JUNCAO growth dynamics. [Methods] In this research, six varieties of JUNCAO plants were selected as experimental subjects. They were divided into a control group planted at ambient temperature (28°C) and a stress group planted at low temperature (4°C). The hyperspectral reflectances of JUNCAO seedling leaves during the seedling stage were collected using an ASD spectroradiomete and a near-infrared spectrometer, and then the leaf physiological indicators were measured to obtain leaf MDA content. Machine learning methods were used to establish the MDA content inversion models based on the collected spectral reflectance data. To enhance the prediction accuracy of the model, an improved one-dimensional deep convolutional generative adversarial network (DCAGN ) was proposed to increase the sample size of the training set. Firstly, the original samples were divided into a training set (96 samples) and a prediction set (48 samples) using the Kennard stone (KS) algorithm at a ratio of 2:1. Secondly, the 96 training set samples were generated through the DCGAN model, resulting in a total of 384 pseudo samples that were 4 times larger than the training set. The pseudo samples were randomly shuffled and sequentially added to the training set to form an enhanced modeling set. Finally, the MDA quantitative detection models were established based on random forest (RF), partial least squares regression (PLSR), and convolutional neural network (CNN) algorithms. By comparing the prediction accuracies of the three models after increasing the sample size of the training set, the best MDA regression detection model of JUNCAO was obtained. [Results and Discussions] (1) The MDA content of the six varieties of JUNCAO plants ranged from 12.1988 to 36.7918 nmol/g. Notably, the MDA content of JUNCAO under low-temperature stress was remarkably increased compared to the control group with significant differences (P<0.05). Moreover, the visible/near-infrared spectral reflectance in the stressed group also exhibited an increasing trend compared to the control group. (2) Samples generated by the DCAGN model conformed to the distribution patterns of the original samples. The spectral curves of the generated samples retained the shape and trends of the original data. The corresponding MDA contented of generated samples consistently falling within the range of the original samples, with the average and standard deviation only decreased by 0.6650 and 0.9743 nmol/g, respectively. (3) Prior to the inclusion of generated samples, the detection performance of the three models differed significantly, with a correlation coefficient (R2) of 0.6967 for RF model, that of 0.6729 for CNN model, and that of 0.5298 for the PLSR model. After the introduction of generated samples, as the number of samples increased, all three models exhibited an initial increase followed by a decrease in R2 on the prediction set, while the root mean square error of prediction (RMSEP) first decreased and then increased. (4) The prediction results of the three regression models indicated that augmenting the sample size by using DCGAN could effectively enhance the prediction performance of models. Particularly, utilizing DCGAN in combination with the RF model achieved the optimal MDA content detection performance, with the R2 of 0.7922 and the RMSEP of 2.1937. [Conclusions] Under low temperature stress, the MDA content and spectral reflectance of the six varieties of JUNCAO leaves significantly increased compared to the control group, which might due to the damage of leaf pigments and tissue structure, and the decrease in leaf water content. Augmenting the sample size using DCGAN effectively enhanced the reliability and detection accuracy of the models. This improvement was evident across different regression models, illustrating the robust generalization capabilities of this DCGAN deep learning network. Specifically, the combination of DCGAN and RF model achieved optimal MDA content detection performance, as expanding to a sufficient sample dataset contributed to improve the modeling accuracy and stability. This research provides valuable insights for JUNCAO plants breeding and the diagnosis of low-temperature stress based on spectral technology and machine learning methods, offering a scientific basis for achieving high, stable, and efficient utilization of JUNCAO plants.
[Objectives] Chlorophyll content and water content are key physiological indicators of crop growth, and their non-destructive detection is a key technology to realize the monitoring of crop growth status such as drought stress. This study took maize as an object to develop a hyperspectral-based approach for the rapid and non-destructive acquisition of the leaf chlorophyll content and water content for drought stress assessment. [Methods] Drought treatment experiments were carried out in a greenhouse of the College of Agriculture, Guangxi University. Maize plants were subjected to drought stress treatment at the seedling stage (four leaves). Four drought treatments were set up for normal water treatment [CK], mild drought [W1], moderate drought [W2], and severe drought [W3], respectively. Leaf samples were collected at the 3rd, 6th, and 9th days after drought treatments, and 288 leaf samples were collected in total, with the corresponding chlorophyll content and water content measured in a standard laboratory protocol. A pair of push-broom hyperspectral cameras were used to collect images of the 288 seedling maize leaf samples, and image processing techniques were used to extract the mean spectra of the leaf lamina part. The algorithm flow framework of "pre-processing - feature extraction - machine learning inversion" was adopted for processing the extracted spectral data. The effects of different pre-processing methods, feature wavelength extraction methods and machine learning regression models were analyzed systematically on the prediction performance of chlorophyll content and water content, respectively. Accordingly, the optimal chlorophyll content and water content inversion models were constructed. Firstly, 70% of the spectral data was randomly sampled and used as the training dataset for training the inversion model, whereas the remaining 30% was used as the testing dataset to evaluate the performance of the inversion model. Subsequently, the effects of different spectral pre-processing methods on the prediction performance of chlorophyll content and water content were compared. Different feature wavelengths were extracted from the optimal pre-processed spectra using different algorithms, then their capabilities in preserve the information useful for the inversion of leaf chlorophyll content and water content were compared. Finally, the performances of different machine learning regression model were compared, and the optimal inversion model was constructed and used to visualize the chlorophyll content and water content. Additionally, the construction of vegetation coefficients were explored for the inversion of chlorophyll content and water content and evaluated their inversion ability. The performance evaluation indexes used include determination coefficient and root mean squared error (RMSE). [Results and Discussions] With the aggravation of stress, the reflectivity of leaves in the wavelength range of 400~1700 nm gradually increased with the degree of drought stress. For the inversion of leaf chlorophyll content and water content, combining stepwise regression (SR) feature extraction with Stacking regression could obtain an optimal performance for chlorophyll content prediction, with an R2 of 0.878 and an RMSE of 0.317 mg/g. Compared with the full-band stacking model, SR-Stacking not only improved R2 by 2.9%, reduced RMSE by 0.0356mg/g, but also reduced the number of model input variables from 1301 to 9. Combining the successive projection algorithm (SPA) feature extraction with Stacking regression could obtain the optimal performance for water content prediction, with an R2 of 0.859 and RMSE of 3.75%. Compared with the full-band stacking model, SPA-Stacking not only increased R2 by 0.2%, reduced RMSE by 0.03%, but also reduced the number of model input variables from 1301 to 16. As the newly constructed vegetation coefficients, normalized difference vegetation index(NDVI) [(R410-R559)/(R410+R559)] and ratio index (RI) (R400/R1171) had the highest accuracy and were significantly higher than the traditional vegetation coefficients for chlorophyll content and water content inversion, respectively. Their R2 were 0.803 and 0.827, and their RMSE were 0.403 mg/g and 3.28%, respectively. The chlorophyll content and water content of leaves were visualized. The results showed that the physiological parameters of leaves could be visualized and the differences of physiological parameters in different regions of the same leaves can be found more intuitively and in detail. [Conclusions] The inversion models and vegetation indices constructed based on hyperspectral information can achieve accurate and non-destructive measurement of chlorophyll content and water content in maize leaves. This study can provide a theoretical basis and technical support for real-time monitoring of corn growth status. Through the leaf spectral information, according to the optimal model, the water content and chlorophyll content of each pixel of the hyperspectral image can be predicted, and the distribution of water content and chlorophyll content can be intuitively displayed by color. Because the field environment is more complex, transfer learning will be carried out in future work to improve its generalization ability in different environments subsequently and strive to develop an online monitoring system for field drought and nutrient stress.
[Objective] Chlorophyll fluorescence (ChlF) emission from photosystem II (PSII) is closely coupled with photochemical reactions. As an efficient and non-destructive means of obtaining plant photosynthesis efficiency and physiological state information, the collection of fluorescence signals is often used in many fields such as plant physiological research, smart agricultural information sensing, etc. Chlorophyll fluorescence imaging systems, which is the experimental device for collecting the fluorescence signal, have difficulties in application due to their high price and complex structure. In order to solve the issues, this paper investigates and constructs a low-cost chlorophyll fluorescence imaging system based on a micro complementary metal oxide semiconductor (CMOS) camera and a smartphone, and carries out experimental verifications and applications on it. [Method] The chlorophyll fluorescence imaging system is mainly composed of three parts: excitation light, CMOS camera and its control circuit, and a upper computer based on a smartphone. The light source of the excitation light group is based on the principle and characteristics of chlorophyll fluorescence, and uses a blue light source of 460 nm band to achieve the best fluorescence excitation effect. In terms of structure, the principle of integrating sphere was borrowed, the bowl-shaped light source structure was adopted, and the design of the LED surface light source was used to meet the requirements of chlorophyll fluorescence signal measurement for the uniformity of the excitation light field. For the adjustment of light source intensity, the control scheme of pulse width modulation was adopted, which could realize sequential control of different intensities of excitation light. Through the simulation analysis of the light field, the light intensity and distribution characteristics of the light field were stuidied, and the calibration of the excitation light group was completed according to the simulation results. The OV5640 micro CMOS camera was used to collect fluorescence images. Combined with the imaging principle of the CMOS camera, the fluorescence imaging intensity of the CMOS camera was calculated, and its ability to collect chlorophyll fluorescence was analyzed and discussed. The control circuit of the CMOS camera uses an STM32 microcontroller as the microcontroller unit, and completes the data communication between the synchronous light group control circuit and the smartphone through the RS232 to TTL serial communication module and the full-speed universal serial bus, respectively. The smartphone upper computer software is the operating software of the chlorophyll fluorescence imaging system user terminal and the overall control program for fluorescence image acquisition. The overall workflow could be summarized as the user sets the relevant excitation light parameters and camera shooting instructions in the upper computer as needed, sends the instructions to the control circuit through the universal serial bus and serial port, and completes the control of excitation light and CMOS camera image acquisition. After the chlorophyll fluorescence image collection was completed, the data would be sent back to the smart phone or server for analysis, processing, storage, and display. In order to verify the design of the proposed scheme, a prototype of the chlorophyll fluorescence imaging system based on this scheme was made for experimental verification. Firstly, the uniformity of the light field was measured on the excitation light to test the actual performance of the excitation light designed in this article. On this basis, a chlorophyll fluorescence imaging experiment under continuous light excitation and modulated pulse light protocols was completed. Through the analysis and processing of the experimental results and comparison with mainstream chlorophyll fluorometers, the fluorescence imaging capabilities and low-cost advantages of this chlorophyll fluorometer were further verified. [Results and Discussions] The maximum excitation light intensity of the chlorophyll fluorescence imaging system designed in this article was 6250 µmol/(m2·s). Through the simulation analysis of the light field and the calculation and analysis of the fluorescence imaging intensity of the CMOS camera, the feasibility of collecting chlorophyll fluorescence images by the OV5640 micro CMOS camera was demonstrated, which provided a basis for the specific design and implementation of the fluorometer. In terms of hardware circuits, it made full use of the software and hardware advantages of smartphones, and only consisted of the control circuits of the excitation light and CMOS camera and the corresponding communication modules to complete the fluorescence image collection work, simplifying the circuit structure and reducing hardware costs to the greatest extent. The final fluorescence instrument achieved a collection resolution of 5 million pixels, a spectral range of 400~1000 nm, and a stable acquisition frequency of up to 42 f/s. Experimental results showed that the measured data was consistent with theoretical analysis and simulation, which could meet the requirements of fluorescence detection. The instrument was capable of collecting images of chlorophyll fluorescence under continuous light excitation or the protocol of modulated pulsed light. The acquired chlorophyll fluorescence images could reflect the two-dimensional heterogeneity of leaves and could effectively distinguish the photosynthetic characteristics of different leaves. Typical chlorophyll fluorescence parameter images of Fv/Fm, Rfd, etc. were in line with expectations. Compared with the existing chlorophyll fluorescence imaging system, the chlorophyll fluorescence imaging system designed in this article has obvious cost advantages while realizing the rapid detection function of chlorophyll fluorescence. [Conclusions] The instrument is with a simple structure and low cost, and has good application value for the detection of plant physiology and environmental changes. The system is useful for developing other fluorescence instruments.
