Loading...
Welcome to Smart Agriculture 中文

Table of Content

    30 December 2023, Volume 5 Issue 4
    Special Issue--Artificial Intelligence and Robot Technology for Smart Agriculture
    Agricultural Robots: Technology Progress, Challenges and Trends | Open Access
    ZHAO Chunjiang, FAN Beibei, LI Jin, FENG Qingchun
    2023, 5(4):  1-15.  doi:10.12133/j.smartag.SA202312030
    Asbtract ( 1676 )   HTML ( 291)   PDF (2498KB) ( 2145 )  
    Figures and Tables | References | Related Articles | Metrics

    [Significance] Autonomous and intelligent agricultural machinery, characterized by green intelligence, energy efficiency, and reduced emissions, as well as high intelligence and man-machine collaboration, will serve as the driving force behind global agricultural technology advancements and the transformation of production methods in the context of smart agriculture development. Agricultural robots, which utilize intelligent control and information technology, have the unique advantage of replacing manual labor. They occupy the strategic commanding heights and competitive focus of global agricultural equipment and are also one of the key development directions for accelerating the construction of China's agricultural power. World agricultural powers and China have incorporated the research, development, manufacturing, and promotion of agricultural robots into their national strategies, respectively strengthening the agricultural robot policy and planning layout based on their own agricultural development characteristics, thus driving the agricultural robot industry into a stable growth period. [Progress] This paper firstly delves into the concept and defining features of agricultural robots, alongside an exploration of the global agricultural robot development policy and strategic planning blueprint. Furthermore, sheds light on the growth and development of the global agricultural robotics industry; Then proceeds to analyze the industrial backdrop, cutting-edge advancements, developmental challenges, and crucial technology aspects of three representative agricultural robots, including farmland robots, orchard picking robots, and indoor vegetable production robots. Finally, summarizes the disparity between Chinese agricultural robots and their foreign counterparts in terms of advanced technologies. (1) An agricultural robot is a multi-degree-of-freedom autonomous operating equipment that possesses accurate perception, autonomous decision-making, intelligent control, and automatic execution capabilities specifically designed for agricultural environments. When combined with artificial intelligence, big data, cloud computing, and the Internet of Things, agricultural robots form an agricultural robot application system. This system has relatively mature applications in key processes such as field planting, fertilization, pest control, yield estimation, inspection, harvesting, grafting, pruning, inspection, harvesting, transportation, and livestock and poultry breeding feeding, inspection, disinfection, and milking. Globally, agricultural robots, represented by plant protection robots, have entered the industrial application phase and are gradually realizing commercialization with vast market potential. (2) Compared to traditional agricultural machinery and equipment, agricultural robots possess advantages in performing hazardous tasks, executing batch repetitive work, managing complex field operations, and livestock breeding. In contrast to industrial robots, agricultural robots face technical challenges in three aspects. Firstly, the complexity and unstructured nature of the operating environment. Secondly, the flexibility, mobility, and commoditization of the operation object. Thirdly, the high level of technology and investment required. (3) Given the increasing demand for unmanned and less manned operations in farmland production, China's agricultural robot research, development, and application have started late and progressed slowly. The existing agricultural operation equipment still has a significant gap from achieving precision operation, digital perception, intelligent management, and intelligent decision-making. The comprehensive performance of domestic products lags behind foreign advanced counterparts, indicating that there is still a long way to go for industrial development and application. Firstly, the current agricultural robots predominantly utilize single actuators and operate as single machines, with the development of multi-arm cooperative robots just emerging. Most of these robots primarily engage in rigid operations, exhibiting limited flexibility, adaptability, and functionality. Secondly, the perception of multi-source environments in agricultural settings, as well as the autonomous operation of agricultural robot equipment, relies heavily on human input. Thirdly, the progress of new teaching methods and technologies for human-computer natural interaction is rather slow. Lastly, the development of operational infrastructure is insufficient, resulting in a relatively low degree of "mechanization". [Conclusions and Prospects] The paper anticipates the opportunities that arise from the rapid growth of the agricultural robotics industry in response to the escalating global shortage of agricultural labor. It outlines the emerging trends in agricultural robot technology, including autonomous navigation, self-learning, real-time monitoring, and operation control. In the future, the path planning and navigation information perception of agricultural robot autonomy are expected to become more refined. Furthermore, improvements in autonomous learning and cross-scenario operation performance will be achieved. The development of real-time operation monitoring of agricultural robots through digital twinning will also progress. Additionally, cloud-based management and control of agricultural robots for comprehensive operations will experience significant growth. Steady advancements will be made in the innovation and integration of agricultural machinery and techniques.

    Three-Dimensional Environment Perception Technology for Agricultural Wheeled Robots: A Review | Open Access
    CHEN Ruiyun, TIAN Wenbin, BAO Haibo, LI Duan, XIE Xinhao, ZHENG Yongjun, TAN Yu
    2023, 5(4):  16-32.  doi:10.12133/j.smartag.SA202308006
    Asbtract ( 390 )   HTML ( 56)   PDF (1885KB) ( 807 )  
    Figures and Tables | References | Related Articles | Metrics

    [Significance] As the research focus of future agricultural machinery, agricultural wheeled robots are developing in the direction of intelligence and multi-functionality. Advanced environmental perception technologies serve as a crucial foundation and key components to promote intelligent operations of agricultural wheeled robots. However, considering the non-structured and complex environments in agricultural on-field operational processes, the environmental information obtained through conventional 2D perception technologies is limited. Therefore, 3D environmental perception technologies are highlighted as they can provide more dimensional information such as depth, among others, thereby directly enhancing the precision and efficiency of unmanned agricultural machinery operation. This paper aims to provide a detailed analysis and summary of 3D environmental perception technologies, investigate the issues in the development of agricultural environmental perception technologies, and clarify the future key development directions of 3D environmental perception technologies regarding agricultural machinery, especially the agricultural wheeled robot. [Progress] Firstly, an overview of the general status of wheeled robots was introduced, considering their dominant influence in environmental perception technologies. It was concluded that multi-wheel robots, especially four-wheel robots, were more suitable for the agricultural environment due to their favorable adaptability and robustness in various agricultural scenarios. In recent years, multi-wheel agricultural robots have gained widespread adoption and application globally. The further improvement of the universality, operation efficiency, and intelligence of agricultural wheeled robots is determined by the employed perception systems and control systems. Therefore, agricultural wheeled robots equipped with novel 3D environmental perception technologies can obtain high-dimensional environmental information, which is significant for improving the accuracy of decision-making and control. Moreover, it enables them to explore effective ways to address the challenges in intelligent environmental perception technology. Secondly, the recent development status of 3D environmental perception technologies in the agriculture field was briefly reviewed. Meanwhile, sensing equipment and the corresponding key technologies were also introduced. For the wheeled robots reported in the agriculture area, it was noted that the applied technologies of environmental perception, in terms of the primary employed sensor solutions, were divided into three categories: LiDAR, vision sensors, and multi-sensor fusion-based solutions. Multi-line LiDAR had better performance on many tasks when employing point cloud processing algorithms. Compared with LiDAR, depth cameras such as binocular cameras, TOF cameras, and structured light cameras have been comprehensively investigated for their application in agricultural robots. Depth camera-based perception systems have shown superiority in cost and providing abundant point cloud information. This study has investigated and summarized the latest research on 3D environmental perception technologies employed by wheeled robots in agricultural machinery. In the reported application scenarios of agricultural environmental perception, the state-of-the-art 3D environmental perception approaches have mainly focused on obstacle recognition, path recognition, and plant phenotyping. 3D environmental perception technologies have the potential to enhance the ability of agricultural robot systems to understand and adapt to the complex, unstructured agricultural environment. Furthermore, they can effectively address several challenges that traditional environmental perception technologies have struggled to overcome, such as partial sensor information loss, adverse weather conditions, and poor lighting conditions. Current research results have indicated that multi-sensor fusion-based 3D environmental perception systems outperform single-sensor-based systems. This superiority arises from the amalgamation of advantages from various sensors, which concurrently serve to mitigate individual shortcomings. [Conclusions and Prospects] The potential of 3D environmental perception technology for agricultural wheeled robots was discussed in light of the evolving demands of smart agriculture. Suggestions were made to improve sensor applicability, develop deep learning-based agricultural environmental perception technology, and explore intelligent high-speed online multi-sensor fusion strategies. Currently, the employed sensors in agricultural wheeled robots may not fully meet practical requirements, and the system's cost remains a barrier to widespread deployment of 3D environmental perception technologies in agriculture. Therefore, there is an urgent need to enhance the agricultural applicability of 3D sensors and reduce production costs. Deep learning methods were highlighted as a powerful tool for processing information obtained from 3D environmental perception sensors, improving response speed and accuracy. However, the limited datasets in the agriculture field remain a key issue that needs to be addressed. Additionally, multi-sensor fusion has been recognized for its potential to enhance perception performance in complex and changeable environments. As a result, it is clear that 3D environmental perception technology based on multi-sensor fusion is the future development direction of smart agriculture. To overcome challenges such as slow data processing speed, delayed processed data, and limited memory space for storing data, it is essential to investigate effective fusion schemes to achieve online multi-source information fusion with greater intelligence and speed.

    An Rapeseed Unmanned Seeding System Based on Cloud-Terminal High Precision Maps | Open Access
    LU Bang, DONG Wanjing, DING Youchun, SUN Yang, LI Haopeng, ZHANG Chaoyu
    2023, 5(4):  33-44.  doi:10.12133/j.smartag.SA202310004
    Asbtract ( 239 )   HTML ( 43)   PDF (2408KB) ( 238 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] Unmanned seeding of rapeseed is an important link to construct unmanned rapeseed farm. Aiming at solving the problems of cumbersome manual collection of small and medium-sized field boundary information in the south, the low efficiency of turnaround operation of autonomous tractor, and leaving a large leakage area at the turnaround point, this study proposes to build an unmanned rapeseed seeding operation system based on cloud-terminal high-precision maps, and to improve the efficiency of the turnaround operation and the coverage of the operation. [Methods] The system was mainly divided into two parts: the unmanned seeding control cloud platform for oilseed rape is mainly composed of a path planning module, an operation monitoring module and a real-time control module; the navigation and control platform for rapeseed live broadcasting units is mainly composed of a Case TM1404 tractor, an intelligent seeding and fertilizing machine, an angle sensor, a high-precision Beidou positioning system, an electric steering wheel, a navigation control terminal and an on-board controller terminal. The process of constructing the high-precision map was as follows: determining the operating field, laying the ground control points; collecting the positional data of the ground control points and the orthophoto data from the unmanned aerial vehicle (UAV); processing the image data and constructing the complete map; slicing the map, correcting the deviation and transmitting it to the webpage. The field boundary information was obtained through the high-precision map. The equal spacing reduction algorithm and scanning line filling algorithm was adopted, and the spiral seeding operation path outside the shuttle row was automatically generated. According to the tractor geometry and kinematics model and the size of the distance between the tractor position and the field boundary, the specific parameters of the one-back and two-cut turning model were calculated, and based on the agronomic requirements of rapeseed sowing operation, the one-back-two-cut turn operation control strategy was designed to realize the rapeseed direct seeding unit's sowing operation for the omitted operation area of the field edges and corners. The test included map accuracy test, operation area simulation test and unmanned seeding operation field test. For the map accuracy test, the test field at the edge of Lake Yezhi of Huazhong Agricultural Universit was selected as the test site, where high-precision maps were constructed, and the image and position (POS) data collected by the UAV were processed, synthesized, and sliced, and then corrected for leveling according to the actual coordinates of the correction point and the coordinates of the image. Three rectangular fields of different sizes were selected for the operation area simulation test to compare the operation area and coverage rate of the three operation modes: set row, shuttle row, and shuttle row outer spiral. The Case TM1404 tractor equipped with an intelligent seeding and fertilizer application integrated machine was used as the test platform for the unmanned seeding operation test, and data such as tracking error and operation speed were recorded in real time by software algorithms. The data such as tracking error and operation speed were recorded in real-time. After the flowering of rapeseed, a series of color images of the operation fields were obtained by aerial photography using a drone during the flowering period of rapeseed, and the color images of the operation fields were spliced together, and then the seedling and non-seedling areas were mapped using map surveying and mapping software. [Results and Discussions] The results of the map accuracy test showed that the maximum error of the high-precision map ground verification point was 3.23 cm, and the results of the operation area simulation test showed that the full-coverage path of the helix outside the shuttle row reduced the leakage rate by 18.58%-26.01% compared with that of the shuttle row and the set of row path. The results of unmanned seeding operation field test showed that the average speed of unmanned seeding operation was 1.46 m/s, the maximum lateral deviation was 7.94 cm, and the maximum average absolute deviation was 1.85 cm. The test results in field showed that, the measured field area was 1 018.61 m2, and the total area of the non-growing oilseed rape area was 69.63 m2, with an operating area of 948.98 m2, and an operating coverage rate of 93.16%. [Conclusions] The effectiveness and feasibility of the constructed unmanned seeding operation system for rapeseed were demonstrated. This study can provide technical reference for unmanned seeding operation of rapeseed in small and medium-sized fields in the south. In the future, the unmanned seeding operation mode of rapeseed will be explored in irregular field conditions to further improve the applicability of the system.

    Traversal Path Planning for Farmland in Hilly Areas Based on Floyd and Improved Genetic Algorithm | Open Access
    ZHOU Longgang, LIU Ting, LU Jinzhu
    2023, 5(4):  45-57.  doi:10.12133/j.smartag.SA202308004
    Asbtract ( 305 )   HTML ( 28)   PDF (2023KB) ( 389 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] To addresses the problem of traversing multiple fields for agricultural robots in hilly terrain, a traversal path planning method is proposed by combining the Floyd algorithm with an improved genetic algorithm. The method provides a solution that can reduce the cost of agricultural robot operation and optimize the order of field traversal in order to improve the efficiency of farmland operation in hilly areas and realizes to predict how an agricultural robot can transition to the next field after completing its coverage path in the current field. [Methods] In the context of hilly terrain characterized by small and densely distributed field blocks, often separated by field ridges, where there was no clear connectivity between the blocks, a method to establish connectivity between the fields was proposed in the research. This method involved projecting from the corner node of the headland path in the current field to each segment of the headland path in adjacent fields vertically. The shortest projected segment was selected as the candidate connectivity path between the two fields, thus establishing potential connectivity between them. Subsequently, the connectivity was verified, and redundant segments or nodes were removed to further simplify the road network. This method allowed for a more accurate assessment of the actual distances between field blocks, thereby providing a more precise and feasible distance cost between field blocks for multi-block traversal sequence planning. Next, the classical graph algorithm, Floyd algorithm, was employed to address the shortest path problem for all pairs of nodes among the fields. The resulting shortest path matrix among headland path nodes within fields, obtained through the Floyd algorithm, allowed to determine the shortest paths and distances between any two endpoint nodes in different fields. This information was used to ascertain the actual distance cost required for agricultural machinery to transfer between fields. Furthermore, for the genetic algorithm in path planning, there were problems such as difficult parameter setting, slow convergence speed and easy to fall into the local optimal solution. This study improved the traditional genetic algorithm by implementing an adaptive strategy. The improved genetic algorithm in this study dynamically adjusted the crossover and mutation probabilities in each generation based on the fitness of the previous generation, adapting to the problem's characteristics. Simultaneously, it dynamically modified the ratio of parent preservation to offspring generation in the current generation, enhancing population diversity and improving global solution search capabilities. Finally, this study employed genetic algorithms and optimization techniques to address the field traversal order problem, akin to the Traveling Salesman Problem (TSP), with the aim of optimizing the traversal path for agricultural robots. The shortest transfer distances between field blocks obtained through the Floyd algorithm were incorporated as variables into the genetic algorithm for optimization. This process leads to the determination of an optimized sequence for traversing the field blocks and the distribution of entry and exit points for each field block. [Results and Discussions] A traversal path planning simulation experiment was conducted to compare the improved genetic algorithm with the traditional genetic algorithm. After 20 simulation experiments, the average traversal path length and the average convergence iteration count of the two algorithms were compared. The simulation results showed that, compared to the traditional genetic algorithm, the proposed improved genetic algorithm in this study shortened the average shortest path by 13.8%, with fewer iterations for convergence, and demonstrated better capability to escape local optimal solutions. To validate the effectiveness of the multi-field path planning method proposed in this study for agricultural machinery coverage, simulations were conducted using real agricultural field data and field operation parameters. The actual operating area located at coordinates (103.61°E, 30.47°N) was selected as the simulation subject. The operating area consisted of 10 sets of field blocks, with agricultural machinery operating parameters set at a minimum turning radius of 1.5 and a working width of 2. The experimental results showed that in terms of path length and path repetition rate, the present method showed more superior performance, and the field traversal order and the arrangement of imports and exports could effectively reduce the path length and path repetition rate. [Conclusions] The experimental results proved the superiority and feasibility of this study on the traversing path planning of agricultural machines in multiple fields, and the output trajectory coordinates of the algorithm can serve as a reference for both human operators and unmanned agricultural machinery during large-scale operations. In future research, particular attention will be given to addressing practical implementation challenges of intelligent algorithms, especially those related to the real-time aspects of navigation systems and challenges such as Kalman linear filtering. These efforts aim to enhance the applicability of the research findings in real-world scenarios.

    Path Tracking Control Algorithm of Tractor-Implement | Open Access
    LIU Zhiyong, WEN Changkai, XIAO Yuejin, FU Weiqiang, WANG Hao, MENG Zhijun
    2023, 5(4):  58-67.  doi:10.12133/j.smartag.SA202308012
    Asbtract ( 205 )   HTML ( 16)   PDF (1386KB) ( 531 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] The usual agricultural machinery navigation focuses on the tracking accuracy of the tractor, while the tracking effect of the trailed implement in the trailed agricultural vehicle is the core of the work quality. The connection mode of the tractor and the implement is non-rigid, and the implement can rotate around the hinge joint. In path tracking, this non-rigid structure, leads to the phenomenon of non-overlapping trajectories of the tractor and the implement, reduce the path tracking accuracy. In addition, problems such as large hysteresis and poor anti-interference ability are also very obvious. In order to solve the above problems, a tractor-implement path tracking control method based on variable structure sliding mode control was proposed, taking the tractor front wheel angle as the control variable and the trailed implement as the control target. [Methods] Firstly, the linear deviation model was established. Based on the structural relationship between the tractor and the trailed agricultural implements, the overall kinematics model of the vehicle was established by considering the four degrees of freedom of the vehicle: transverse, longitudinal, heading and articulation angle, ignoring the lateral force of the vehicle and the slip in the forward process. The geometric relationship between the vehicle and the reference path was integrated to establish the linear deviation model of vehicle-road based on the vehicle kinematic model and an approximate linearization method. Then, the control algorithm was designed. The switching function was designed considering three evaluation indexes: lateral deviation, course deviation and hinged angle deviation. The exponential reaching law was used as the reaching mode, the saturation function was used instead of the sign function to reduce the control variable jitter, and the convergence of the control law was verified by combining the Lyapunov function. The system was three-dimensional, in order to improve the dynamic response and steady-state characteristics of the system, the two conjugate dominant poles of the system were assigned within the required range, and the third point was kept away from the two dominant poles to reduce the interference on the system performance. The coefficient matrix of the switching function was solved based on the Ackermann formula, then the calculation formula of the tractor front wheel angle was obtained, and the whole control algorithm was designed. Finally, the path tracking control simulation experiment was carried out. The sliding mode controller was built in the MATLAB/Simulink environment, the controller was composed of the deviation calculation module and the control output calculation module. The tractor-implement model in Carsim software was selected with the front car as a tractor and the rear car as the single-axle implement, and tracking control simulation tests of different reference paths were conducted in the MATLAB/Carsim co-simulation environment. [Results and Discussions] Based on the co-simulation environment, the tracking simulation experiments of three reference paths were carried out. When tracking the double lane change path, the lateral deviation and heading deviation of the agricultural implement converged to 0 m and 0° after 8 s. When the reference heading changed, the lateral deviation and heading deviation were less than 0.1 m and less than 7°. When tracking the circular reference path, the lateral deviation of agricultural machinery tended to be stable after 7 s and was always less than 0.03 m, and the heading deviation of agricultural machinery tended to be stable after 7 s and remained at 0°. The simulation results of the double lane change path and the circular path showed that the controller could maintain good performance when tracking the constant curvature reference path. When tracking the reference path of the S-shaped curve, the tracking performance of the agricultural machinery on the section with constant curvature was the same as the previous two road conditions, and the maximum lateral deviation of the agricultural machinery at the curvature change was less than 0.05 m, the controller still maintained good tracking performance when tracking the variable curvature path. [Conclusions] The sliding mode variable structure controller designed in this study can effectively track the linear and circular reference paths, and still maintain a good tracking effect when tracking the variable curvature paths. Agricultural machinery can be on-line in a short time, which meets the requirements of speediness. In the tracking simulation test, the angle of the tractor front wheel and the articulated angle between the tractor and agricultural implement are kept in a small range, which meets the needs of actual production and reduces the possibility of safety accidents. In summary, the agricultural implement can effectively track the reference path and meet the requirements of precision, rapidity and safety. The model and method proposed in this study provide a reference for the automatic navigation of tractive agricultural implement. In future research, special attention will be paid to the tracking control effect of the control algorithm in the actual field operation and under the condition of large speed changes.

    Traceability Model of Plantation Agricultural Products Based on Blockchain and InterPlanetary File System | Open Access
    CHEN Dandan, ZHANG Lijie, JIANG Shuangfeng, ZHANG En, ZHANG Jie, ZHAO Qing, ZHENG Guoqing, LI Guoqiang
    2023, 5(4):  68-78.  doi:10.12133/j.smartag.SA202307004
    Asbtract ( 254 )   HTML ( 28)   PDF (2011KB) ( 314 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] The InterPlanetary File System (IPFS) is a peer-to-peer distributed file system, aiming to establish a global, open, and decentralized network for storage and sharing. Combining the IPFS and blockchain technology could alleviate the pressure on blockchain storage. The distinct features of the supply chain for agricultural products in the plantation industry, including extended production cycles, multiple, heterogeneous data sources, and relatively fragmented production, which can readily result in information gaps and opacity throughout the supply chain; in the traceability process of agricultural products, there are issues with sensitive data being prone to leakage and a lack of security, and the supply chain of plantation agricultural products is long, and the traceability data is often stored in multiple blocks, which requires frequent block tracing operations during tracing, resulting in low efficiency. Consequently, the aim of this study is to fully encapsulate the decentralized nature of blockchain, safeguard the privacy of sensitive data, and alleviate the storage strain of blockchain. [Method] A traceability model for plantation-based agricultural products was developed, leveraging the hyperledger fabric consortium chain and the IPFS. Based on data type, traceability data was categorized into structured and unstructured data. Given that blockchain ledgers were not optimized for direct storage of unstructured data, such as images and videos, to alleviate the storage strain on the blockchain, unstructured data was persisted in the IPFS, while structured data remains within the blockchain ledger. Based on data privacy categories, traceability data was categorized into public data and sensitive data. Public data was stored in the public ledger of hyperledger fabric, while sensitive data was stored in the private data collection of hyperledger fabric. This method allowed for efficient data access while maintaining data security, enhancing the efficiency of traceability. Hyperledger Fabric was the foundational platform for the development of the prototype system. The front-end website was based on the TCP/IP protocol stack. The website visualization was implemented through the React framework. Smart contracts were crafted using the Java programming language. The performance of the application layer interface was tested using the testing tool Postman. [Conclusions and Discussions] The blockchain-based plantation agricultural product traceability system was structured into a five-tiered architecture, starting from the top: the application layer, gateway layer, contract layer, consensus layer, and data storage layer. The primary service providers at the application layer were the enterprises and consumers involved in each stage of the traceability process. The gateway layer served as the middleware between users and the blockchain, primarily providing interface support for the front-end interface of the application layer. The contract layer mainly included smart contracts for planting, processing, warehousing, transportation, and sales. The consensus layer used the EtcdRaft consensus algorithm. The data storage layer was divided into the on-chain storage layer of the blockchain ledger and the off-chain storage layer of the IPFS cluster. In terms of data types, each piece of traceability data was categorized into structured data items and unstructured data items. Unstructured data was stored in the Interstellar File System cluster, and the returned content identifiers were integrated with the structured data items into the blockchain nodes within the traceability system. In the realm of data privacy, smart contracts were employed to segregate public and sensitive data, with public data directly integrating onto the blockchain, and sensitive data, adhering to predefined sharing policies, being stored in a private dataset designated by hyperledger fabric. In terms of user queries, consumers could retrieve product traceability information via a traceability system overseen by a reputable authority. The developed model website consisted of three parts: a login section, an agricultural product circulation information management and user data management section for enterprises in various links, and a traceability data query section for consumers. When using synchronous and asynchronous Application Program Interfaces, the average data on-chain latency was 2 138.9 and 37.6 ms, respectively, and the average data query latency was 12.3 ms. Blockchain, as the foundational data storage technology, enhances the credibility and transaction efficiency in agricultural product traceability. [Conclusions] This study designed and implemented a plantation agricultural product traceability model leveraging blockchain technology's private dataset and the IPFS cluster. This model ensured secure sharing and storage of traceability data, particularly sensitive data, across all stages. Compared to traditional centralized traceability models, it enhanced the reliability of the traceability data. Based on the evaluation through experimental systems, the traceability model proposed in this study effectively safeguarded the privacy of sensitive data in enterprises. Additionally, it offered high efficiency in data linking and querying. Applicable to the real-world traceability environment of plantation agricultural products, it showed potential for widespread application and promotion, offering fresh insights for designing blockchain traceability models in this sector. The model is still in its experimental phase and lacks applications across various types of crops in the farming industry. The subsequent step is to apply the model in real-world scenarios, continually enhance its efficiency, refine the model, advance the practical application of blockchain technology, and lay the foundation for agricultural modernization.

    Collaborative Computing of Food Supply Chain Privacy Data Elements Based on Federated Learning | Open Access
    XU Jiping, LI Hui, WANG Haoyu, ZHOU Yan, WANG Zhaoyang, YU Chongchong
    2023, 5(4):  79-91.  doi:10.12133/j.smartag.SA202309012
    Asbtract ( 225 )   HTML ( 22)   PDF (1719KB) ( 424 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] The flow of private data elements plays a crucial role in the food supply chain, and the safe and efficient operation of the food supply chain can be ensured through the effective management and flow of private data elements. Through collaborative computing among the whole chain of the food supply chain, the production, transportation and storage processes of food can be better monitored and managed, so that possible quality and safety problems can be detected and solved in a timely manner, and the health and rights of consumers can be safeguarded. It can also be applied to the security risk assessment and early warning of the food supply chain. By analyzing big data, potential risk factors and abnormalities can be identified, and timely measures can be taken for early warning and intervention to reduce the possibility of quality and safety risks. This study combined the industrial Internet identification and resolution system with the federated learning algorithm, which can realize collaborative learning among multiple enterprises, and each enterprise can carry out collaborative training of the model without sharing the original data, which protects the privacy and security of the data while realizing the flow of the data, and it can also make use of the data resources distributed in different segments, which can realize more comprehensive and accurate collaborative calculations, and improve the safety and credibility of the industrial Internet system's security and credibility. [Methods] To address the problem of not being able to share and participate in collaborative computation among different subjects in the grain supply chain due to the privacy of data elements, this study first analyzed and summarized the characteristics of data elements in the whole link of grain supply chain, and proposed a grain supply chain data flow and collaborative computation architecture based on the combination of the industrial Internet mark resolution technology and the idea of federated learning, which was constructed in a layered and graded model to provide a good infrastructure for the decentralized between the participants. The data identification code for the flow of food supplied chain data elements and the task identification code for collaborative calculation of food supply chain, as well as the corresponding parameter data model, information data model and evaluation data model, were designed to support the interoperability of federated learning data. A single-link horizontal federation learning model with isomorphic data characteristics of different subjects and a cross-link vertical federation learning model with heterogeneous data characteristics were constructed, and the model parameters were quickly adjusted and calculated based on logistic regression algorithm, neural network algorithm and other algorithms, and the food supply chain security risk assessment scenario was taken as the object of the research, and the research was based on the open source FATE (Federated AI Technology) federation learning model. Enabler (Federated AI Technology) federated learning platform for testing and validation, and visualization of the results to provide effective support for the security management of the grain supply chain. [Results and Discussion] Compared with the traditional single-subject assessment calculation method, the accuracy of single-session isomorphic horizontal federation learning model assessment across subjects was improved by 6.7%, and the accuracy of heterogeneous vertical federation learning model assessment across sessions and subjects was improved by 8.3%. This result showed that the single-session isomorphic horizontal federated learning model assessment across subjects could make full use of the data information of each subject by merging and training the data of different subjects in the same session, thus improving the accuracy of security risk assessment. The heterogeneous vertical federated learning model assessment of cross-session and cross-subject further promotes the application scope of collaborative computing by jointly training data from different sessions and subjects, which made the results of safety risk assessment more comprehensive and accurate. The advantage of combining federated learning and logo resolution technology was that it could conduct model training without sharing the original data, which protected data privacy and security. At the same time, it could also realize the effective use of data resources and collaborative computation, improving the efficiency and accuracy of the assessment process. [Conclusions] The feasibility and effectiveness of this study in practical applications in the grain industry were confirmed by the test validation of the open-source FATE federated learning platform. This provides reliable technical support for the digital transformation of the grain industry and the security management of the grain supply chain, and helps to improve the intelligence level and competitiveness of the whole grain industry. Therefore, this study can provide a strong technical guarantee for realizing the safe, efficient and sustainable development of the grain supply chain.

    Individual Tree Skeleton Extraction and Crown Prediction Method of Winter Kiwifruit Trees | Open Access
    LI Zhengkai, YU Jiahui, PAN Shijia, JIA Zefeng, NIU Zijie
    2023, 5(4):  92-104.  doi:10.12133/j.smartag.SA202308015
    Asbtract ( 205 )   HTML ( 18)   PDF (2529KB) ( 275 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] The proliferation of kiwifruit trees severely overlaps, resulting in a complex canopy structure, rendering it impossible to extract their skeletons or predict their canopies using conventional methods. The objective of this research is to propose a crown segmentation method that integrates skeleton information by optimizing image processing algorithms and developing a new scheme for fusing winter and summer information. In cases where fruit trees are densely distributed, achieving accurate segmentation of fruit tree canopies in orchard drone images can efficiently and cost-effectively obtain canopy information, providing a foundation for determining summer kiwifruit growth size, spatial distribution, and other data. Furthermore, it facilitates the automation and intelligent development of orchard management. [Methods] The 4- to 8-year-old kiwifruit trees were chosen and remote sensing images of winter and summer via unmanned aerial vehicles were obtain as the primary analysis visuals. To tackle the challenge of branch extraction in winter remote sensing images, a convolutional attention mechanism was integrated into the PSP-Net network, along with a joint attention loss function. This was designed to boost the network's focus on branches, enhance the recognition and targeting capabilities of the target area, and ultimately improve the accuracy of semantic segmentation for fruit tree branches.For the generation of the skeleton, digital image processing technology was employed for screening. The discrete information of tree branches was transformed into the skeleton data of a single fruit tree using growth seed points. Subsequently, the semantic segmentation results were optimized through mathematical morphology calculations, enabling smooth connection of the branches. In response to the issue of single tree canopy segmentation in summer, the growth characteristics of kiwifruit trees were taken into account, utilizing the outward expansion of branches growing from the trunk.The growth of tree branches was simulated by using morphological expansion to predict the summer canopy. The canopy prediction results were analyzed under different operators and parameters, and the appropriate expansion operators along with their corresponding operation lengths were selected. The skeleton of a single tree was extracted from summer images. By combining deep learning with mathematical morphology methods through the above steps, the optimized single tree skeleton was used as a prior condition to achieve canopy segmentation. [Results and Discussions] In comparison to traditional methods, the accuracy of extracting kiwifruit tree canopy information images at each stage of the process has been significantly enhanced. The enhanced PSP Net was evaluated using three primary regression metrics: pixel accuracy (PA), mean intersection over union ratio (MIoU), and weighted F1 Score (WF1). The PA, MIoU and WF1 of the improved PSP-Net were 95.84%, 95.76% and 95.69% respectively, which were increased by 12.30%, 22.22% and 17.96% compared with U-Net, and 21.39% , 21.51% and 18.12% compared with traditional PSP-Net, respectively. By implementing this approach, the skeleton extraction function for a single fruit tree was realized, with the predicted PA of the canopy surpassing 95%, an MIoU value of 95.76%, and a WF1 of canopy segmentation approximately at 94.07%.The average segmentation precision of the approach surpassed 95%, noticeably surpassing the original skeleton's 81.5%. The average conformity between the predicted skeleton and the actual summer skeleton stand at 87%, showcasing the method's strong prediction performance. Compared with the original skeleton, the PA, MIoU and WF1 of the optimized skeleton increased by 13.2%, 10.9% and 18.4%, respectively. The continuity of the predicted skeleton had been optimized, resulting in a significant improvement of the canopy segmentation index. The solution effectively addresses the issue of semantic segmentation fracture, and a single tree canopy segmentation scheme that incorporates skeleton information could effectively tackle the problem of single fruit tree canopy segmentation in complex field environments. This provided a novel technical solution for efficient and low-cost orchard fine management. [Conclusions] A method for extracting individual kiwifruit plant skeletons and predicting canopies based on skeleton information was proposed. This demonstrates the enormous potential of drone remote sensing images for fine orchard management from the perspectives of method innovation, data collection, and problem solving. Compared with manual statistics, the overall efficiency and accuracy of kiwifruit skeleton extraction and crown prediction have significantly improved, effectively solving the problem of case segmentation in the crown segmentation process.The issue of semantic segmentation fragmentation has been effectively addressed, resulting in the development of a single tree canopy segmentation method that incorporates skeleton information. This approach can effectively tackle the challenges of single fruit tree canopy segmentation in complex field environments, thereby offering a novel technical solution for efficient and cost-effective orchard fine management. While the research is primarily centered on kiwifruit trees, the methodology possesses strong universality. With appropriate modifications, it can be utilized to monitor canopy changes in other fruit trees, thereby showcasing vast application potential.

    Agricultural Technology Knowledge Intelligent Question-Answering System Based on Large Language Model | Open Access
    WANG Ting, WANG Na, CUI Yunpeng, LIU Juan
    2023, 5(4):  105-116.  doi:10.12133/j.smartag.SA202311005
    Asbtract ( 1202 )   HTML ( 210)   PDF (1475KB) ( 1646 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] The rural revitalization strategy presents novel requisites for the extension of agricultural technology. However, the conventional method encounters the issue of a contradiction between supply and demand. Therefore, there is a need for further innovation in the supply form of agricultural knowledge. Recent advancements in artificial intelligence technologies, such as deep learning and large-scale neural networks, particularly the advent of large language models (LLMs), render anthropomorphic and intelligent agricultural technology extension feasible. With the agricultural technology knowledge service of fruit and vegetable as the demand orientation, the intelligent agricultural technology question answering system was built in this research based on LLM, providing agricultural technology extension services, including guidance on new agricultural knowledge and question-and-answer sessions. This facilitates farmers in accessing high-quality agricultural knowledge at their convenience. [Methods] Through an analysis of the demands of strawberry farmers, the agricultural technology knowledge related to strawberry cultivation was categorized into six themes: basic production knowledge, variety screening, interplanting knowledge, pest diagnosis and control, disease diagnosis and control, and drug damage diagnosis and control. Considering the current situation of agricultural technology, two primary tasks were formulated: named entity recognition and question answering related to agricultural knowledge. A training corpus comprising entity type annotations and question-answer pairs was constructed using a combination of automatic machine annotation and manual annotation, ensuring a small yet high-quality sample. After comparing four existing Large Language Models (Baichuan2-13B-Chat, ChatGLM2-6B, Llama 2-13B-Chat, and ChatGPT), the model exhibiting the best performance was chosen as the base LLM to develop the intelligent question-answering system for agricultural technology knowledge. Utilizing a high-quality corpus, pre-training of a Large Language Model and the fine-tuning method, a deep neural network with semantic analysis, context association, and content generation capabilities was trained. This model served as a Large Language Model for named entity recognition and question answering of agricultural knowledge, adaptable to various downstream tasks. For the task of named entity recognition, the fine-tuning method of Lora was employed, fine-tuning only essential parameters to expedite model training and enhance performance. Regarding the question-answering task, the Prompt-tuning method was used to fine-tune the Large Language Model, where adjustments were made based on the generated content of the model, achieving iterative optimization. Model performance optimization was conducted from two perspectives: data and model design. In terms of data, redundant or unclear data was manually removed from the labeled corpus. In terms of the model, a strategy based on retrieval enhancement generation technology was employed to deepen the understanding of agricultural knowledge in the Large Language Model and maintain real-time synchronization of knowledge, alleviating the problem of LLM hallucination. Drawing upon the constructed Large Language Model, an intelligent question-answering system was developed for agricultural technology knowledge. This system demonstrates the capability to generate high-precision and unambiguous answers, while also supporting the functionalities of multi-round question answering and retrieval of information sources. [Results and Discussions] Accuracy rate and recall rate served as indicators to evaluate the named entity recognition task performance of the Large Language Models. The results indicated that the performance of Large Language Models was closely related to factors such as model structure, the scale of the labeled corpus, and the number of entity types. After fine-tuning, the ChatGLM Large Language Model demonstrated the highest accuracy and recall rate. With the same number of entity types, a higher number of annotated corpora resulted in a higher accuracy rate. Fine-tuning had different effects on different models, and overall, it improved the average accuracy of all models under different knowledge topics, with ChatGLM, Llama, and Baichuan values all surpassing 85%. The average recall rate saw limited increase, and in some cases, it was even lower than the values before fine-tuning. Assessing the question-answering task of Large Language Models using hallucination rate and semantic similarity as indicators, data optimization and retrieval enhancement generation techniques effectively reduced the hallucination rate by 10% to 40% and improved semantic similarity by more than 15%. These optimizations significantly enhanced the generated content of the models in terms of correctness, logic, and comprehensiveness. [Conclusion] The pre-trained Large Language Model of ChatGLM exhibited superior performance in named entity recognition and question answering tasks in the agricultural field. Fine-tuning pre-trained Large Language Models for downstream tasks and optimizing based on retrieval enhancement generation technology mitigated the problem of language hallucination, markedly improving model performance. Large Language Model technology has the potential to innovate agricultural technology knowledge service modes and optimize agricultural knowledge extension. This can effectively reduce the time cost for farmers to obtain high-quality and effective knowledge, guiding more farmers towards agricultural technology innovation and transformation. However, due to challenges such as unstable performance, further research is needed to explore optimization methods for Large Language Models and their application in specific scenarios.

    Phenotype Analysis of Pleurotus Geesteranus Based on Improved Mask R-CNN | Open Access
    ZHOU Huamao, WANG Jing, YIN Hua, CHEN Qi
    2023, 5(4):  117-126.  doi:10.12133/j.smartag.SA202309024
    Asbtract ( 202 )   HTML ( 17)   PDF (1384KB) ( 274 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] Pleurotus geesteranus is a rare edible mushroom with a fresh taste and rich nutritional elements, which is popular among consumers. It is not only cherished for its unique palate but also for its abundant nutritional elements. The phenotype of Pleurotus geesteranus is an important determinant of its overall quality, a specific expression of its intrinsic characteristics and its adaptation to various cultivated environments. It is crucial to select varieties with excellent shape, integrity, and resistance to cracking in the breeding process. However, there is still a lack of automated methods to measure these phenotype parameters. The method of manual measurement is not only time-consuming and labor-intensive but also subjective, which lead to inconsistent and inaccurate results. Thus, the traditional approach is unable to meet the demand of the rapid development Pleurotus geesteranus industry. [Methods] To solve the problems which mentioned above, firstly, this study utilized an industrial-grade camera (Daheng MER-500-14GM) and a commonly available smartphone (Redmi K40) to capture high-resolution images in DongSheng mushroom industry (Jiujiang, Jiangxi province). After discarding blurred and repetitive images, a total of 344 images were collected, which included two commonly distinct varieties, specifically Taixiu 57 and Gaoyou 818. A series of data augmentation algorithms, including rotation, flipping, mirroring, and blurring, were employed to construct a comprehensive Pleurotus geesteranus image dataset. At the end, the dataset consisted of 3 440 images and provided a robust foundation for the proposed phenotype recognition model. All images were divided into training and testing sets at a ratio of 8:2, ensuring a balanced distribution for effective model training. In the second part, based upon foundational structure of classical Mask R-CNN, an enhanced version specifically tailored for Pleurotus geesteranus phenotype recognition, aptly named PG-Mask R-CNN (Pleurotus geesteranus-Mask Region-based Convolutional Neural Network) was designed. The PG-Mask R-CNN network was refined through three approaches: 1) To take advantage of the attention mechanism, the SimAM attention mechanism was integrated into the third layer of ResNet101feature extraction network after analyzing and comparing carefully, it was possible to enhance the network's performance without increasing the original network parameters. 2) In order to avoid the problem of Mask R-CNN's feature pyramid path too long to split low-level feature and high-level feature, which may impair the semantic information of the high-level feature and lose the positioning information of the low-level feature, an improved feature pyramid network was used for multiscale fusion, which allowed us to amalgamate information from multiple levels for prediction. 3) To address the limitation of IoU (Intersection over Union) bounding box, which only considered the overlapping area between the prediction box and target box while ignoring the non-overlapping area, a more advanced loss function called GIoU (Generalized Intersection over Union) was introduced. This replacement improved the calculation of image overlap and enhanced the performance of the model. Furthermore, to evaluate crack state of Pleurotus geesteranus more scientifically, reasonably and accurately, the damage rate as a new crack quantification evaluation method was introduced, which was calculated by using the proportion of cracks in the complete pileus of the mushroom and utilized the MRE (Mean Relative Error) to calculate the mean relative error of the Pleurotus geesteranus's damage rate. Thirdly, the PG-Mask R-CNN network was trained and tested based on the Pleurotus geesteranus image dataset. According to the detection and segmentation results, the measurement and accuracy verification were conducted. Finally, considering that it was difficult to determine the ground true of the different shapes of Pleurotus geesteranus, the same method was used to test 4 standard blocks of different specifications, and the rationality of the proposed method was verified. [Results and Discussions] In the comparative analysis, the PG-Mask R-CNN model was superior to Grabcut algorithm and other 4 instance segmentation models, including YOLACT (You Only Look At Coefficien Ts), InstaBoost, QueryInst, and Mask R-CNN. In object detection tasks, the experimental results showed that PG-Mask R-CNN model achieved a mAP of 84.8% and a mAR (mean Average Recall) of 87.7%, respectively, higher than the five methods were mentioned above. Furthermore, the MRE of the instance segmentation results was 0.90%, which was consistently lower than that of other instance segmentation models. In addition, from a model size perspective, the PG-Mask R-CNN model had a parameter count of 51.75 M, which was slightly larger than that of the unimproved Mask R-CNN model but smaller than other instance segmentation models. With the instance segmentation results on the pileus and crack, the MRE were 1.30% and 7.54%, respectively, while the MAE of the measured damage rate was 0.14%. [Conclusions] The proposed PG-Mask R-CNN model demonstrates a high accuracy in identifying and segmenting the stipe, pileus, and cracks of Pleurotus geesteranus. Thus, it can help the automated measurements of phenotype measurements of Pleurotus geesteranus, which lays a technical foundation for subsequent intelligent breeding, smart cultivation and grading of Pleurotus geesteranus.

    Path Planning and Motion Control Method for Sick and Dead Animal Transport Robots Integrating Improved A * Algorithm and Fuzzy PID | Open Access
    XU Jishuang, JIAO Jun, LI Miao, LI Hualong, YANG Xuanjiang, LIU Xianwang, GUO Panpan, MA Zhirun
    2023, 5(4):  127-136.  doi:10.12133/j.smartag.SA202308001
    Asbtract ( 170 )   HTML ( 12)   PDF (1068KB) ( 314 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] A key challenge for the harmless treatment center of sick and dead animal is to prevent secondary environmental pollution, especially during the process of transporting the animals from cold storage to intelligent treatment facilities. In order to solve this problem and achieve the intelligent equipment process of transporting sick and dead animal from storage cold storage to harmless treatment equipment in the harmless treatment center, it is necessary to conduct in-depth research on the key technical problems of path planning and autonomous walking of transport robots. [Methods] A * algorithm is mainly adopted for the robot path planning algorithm for indoor environments, but traditional A * algorithms have some problems, such as having many inflection points, poor smoothness, long calculation time, and many traversal nodes. In order to solve these problems, a path planning method for the harmless treatment of diseased and dead animal using transport robots based on the improved A algorithm was constructed, as well as a motion control method based on fuzzy proportional integral differential (PID). The Manhattan distance method was used to replace the heuristic function of the traditional A * algorithm, improving the efficiency of calculating the distance between the starting and ending points in the path planning process. Referring to the actual location of the harmless treatment site for sick and dead animal, vector cross product calculation was performed based on the vector from the starting point to the target point and the vector from the current position to the endpoint target. Additional values were added and dynamic adjustments were implemented, thereby changing the value of the heuristic function. In order to further improve the efficiency of path planning and reduce the search for nodes in the planning process, a method of adding function weights to the heuristic function was studied based on the actual situation on site, to change the weights according to different paths. When the current location node was relatively open, the search efficiency was improved by increasing the weight. When encountering situations such as corners, the weight was reduced to improve the credibility of the path. By improving the heuristic function, a driving path from the starting point to the endpoint was quickly obtained, but the resulting path was not smooth enough. Meanwhile, during the tracking process, the robot needs to accelerate and decelerate frequently to adapt to the path, resulting in energy loss. Therefore, according to the different inflection points and control points of the path, different orders of Bessel functions were introduced to smooth the planning process for the path, in order to achieve practical application results. By analyzing the kinematics of robot, the differential motion method of the track type was clarified. On this basis, a walking control algorithm for the robot based on fuzzy PID control was studied and proposed. Based on the actual operation status of the robot, the fuzzy rule conditions were recorded into a fuzzy control rule table, achieving online identification of the characteristic parameters of the robot and adjusting the angular velocity deviation of robot. When the robot controller received a fuzzy PID control signal, the angular velocity output from the control signal was converted into a motor rotation signal, which changed the motor speed on both sides of the robot to achieve differential control and adjust the steering of the robot. [Results and Discussions] Simulation experiments were conducted using the constructed environmental map obtained, verifying the effectiveness of the path planning method for the harmless treatment of sick and dead animal using the improved A algorithm. The comparative experiments between traditional A * algorithm and improved algorithm were conducted. The experimental results showed that the average traversal nodes of the improved A * algorithm decreased from 3 067 to 1 968, and the average time of the algorithm decreased from 20.34 s to 7.26 s. Through on-site experiments, the effectiveness and reliability of the algorithm were further verified. Different colors were used to identify the planned paths, and optimization comparison experiments were conducted on large angle inflection points, U-shaped inflection points, and continuous inflection points in the paths, verifying the optimization effect of the Bessel function on path smoothness. The experimental results showed that the path optimized by the Bessel function was smoother and more suitable for the walking of robot in practical scenarios. Fuzzy PID path tracking experiment results showed that the loading truck can stay close to the original route during both straight and turning driving, demonstrating the good effect of fuzzy PID on path tracking. Further experiments were conducted on the harmless treatment center to verify the effectiveness and practical application of the improved algorithm. Based on the path planning algorithm, the driving path of robot was quickly planned, and the fuzzy PID control algorithm was combined to accurately output the angular velocity, driving the robot to move. The transport robots quickly realized the planning of the transportation path, and during the driving process, could always be close to the established path, and the deviation error was maintained within a controllable range. [Conclusions] A path planning method for the harmless treatment of sick and dead animal using an transport robots based on an improved A * algorithm combined with a fuzzy PID motion control was proposed in this study. This method could effectively shorten the path planning time, reduce traversal nodes, and improve the efficiency and smoothness of path planning.

    Image Segmentation Method Combined with VoVNetv2 and Shuffle Attention Mechanism for Fish Feeding in Aquaculture | Open Access
    WANG Herong, CHEN Yingyi, CHAI Yingqian, XU Ling, YU Huihui
    2023, 5(4):  137-149.  doi:10.12133/j.smartag.SA202310003
    Asbtract ( 227 )   HTML ( 21)   PDF (2425KB) ( 292 )  
    Figures and Tables | References | Related Articles | Metrics

    [Objective] Intelligent feeding methods are significant for improving breeding efficiency and reducing water quality pollution in current aquaculture. Feeding image segmentation of fish schools is a critical step in extracting the distribution characteristics of fish schools and quantifying their feeding behavior for intelligent feeding method development. While, an applicable approach is lacking due to images challenges caused by blurred boundaries and similar individuals in practical aquaculture environment. In this study, a high-precision segmentation method was proposed for fish school feeding images and provides technical support for the quantitative analysis of fish school feeding behavior. [Methods] The novel proposed method for fish school feeding images segmentation combined VoVNetv2 with an attention mechanism named Shuffle Attention. Firstly, a fish feeding segmentation dataset was presented. The dataset was collected at the intensive aquaculture base of Laizhou Mingbo Company in Shandong province, with a focus on Oplegnathus punctatus as the research target. Cameras were used to capture videos of the fish school before, during, and after feeding. The images were annotated at the pixel level using Labelme software. According to the distribution characteristics of fish feeding and non-feeding stage, the data was classified into two semantic categories— non-occlusion and non-aggregation fish (fish1) and occlusion or aggregation fish (fish2). In the preprocessing stage, data cleaning and image augmentation were employed to further enhance the quality and diversity of the dataset. Initially, data cleaning rules were established based on the distribution of annotated areas within the dataset. Images with outlier annotations were removed, resulting in an improvement in the overall quality of the dataset. Subsequently, to prevent the risk of overfitting, five data augmentation techniques (random translation, random flip, brightness variation, random noise injection, random point addition) were applied for mixed augmentation on the dataset, contributing to an increased diversity of the dataset. Through data augmentation operations, the dataset was expanded to three times its original size. Eventually, the dataset was divided into a training dataset and testing dataset at a ratio of 8:2. Thus, the final dataset consisted of 1 612 training images and 404 testing images. In detail, there were a total of 116 328 instances of fish1 and 20 924 instances of fish2. Secondly, a fish feeding image segmentation method was proposed. Specifically, VoVNetv2 was used as the backbone network for the Mask R-CNN model to extract image features. VoVNetv2 is a backbone network with strong computational capabilities. Its unique feature aggregation structure enables effective fusion of features at different levels, extracting diverse feature representations. This facilitates better capturing of fish schools of different sizes and shapes in fish feeding images, achieving accurate identification and segmentation of targets within the images. To maximize feature mappings with limited resources, the experiment replaced the channel attention mechanism in the one-shot aggregation (OSA) module of VoVNetv2 with a more lightweight and efficient attention mechanism named shuffle attention. This improvement allowed the network to concentrate more on the location of fish in the image, thus reducing the impact of irrelevant information, such as noise, on the segmentation results. Finally, experiments were conducted on the fish segmentation dataset to test the performance of the proposed method. [Results and Discussions] The results showed that the average segmentation accuracy of the Mask R-CNN network reached 63.218% after data cleaning, representing an improvement of 7.018% compared to the original dataset. With both data cleaning and augmentation, the network achieved an average segmentation accuracy of 67.284%, indicating an enhancement of 11.084% over the original dataset. Furthermore, there was an improvement of 4.066% compared to the accuracy of the dataset after cleaning alone. These results demonstrated that data preprocessing had a positive effect on improving the accuracy of image segmentation. The ablation experiments on the backbone network revealed that replacing the ResNet50 backbone with VoVNetv2-39 in Mask R-CNN led to a 2.511% improvement in model accuracy. After improving VoVNetv2 through the Shuffle Attention mechanism, the accuracy of the model was further improved by 1.219%. Simultaneously, the parameters of the model decreased by 7.9%, achieving a balance between accuracy and lightweight design. Comparing with the classic segmentation networks SOLOv2, BlendMask and CondInst, the proposed model achieved the highest segmentation accuracy across various target scales. For the fish feeding segmentation dataset, the average segmentation accuracy of the proposed model surpassed BlendMask, CondInst, and SOLOv2 by 3.982%, 12.068%, and 18.258%, respectively. Although the proposed method demonstrated effective segmentation of fish feeding images, it still exhibited certain limitations, such as omissive detection, error segmentation, and false classification. [Conclusions] The proposed instance segmentation algorithm (SA_VoVNetv2_RCNN) effectively achieved accurate segmentation of fish feeding images. It can be utilized for counting the number and pixel quantities of two types of fish in fish feeding videos, facilitating quantitative analysis of fish feeding behavior. Therefore, this technique can provide technical support for the analysis of piscine feeding actions. In future research, these issues will be addressed to further enhance the accuracy of fish feeding image segmentation.

    Overview Article
    The Development Logic, Influencing Factors and Realization Path for Low-Carbon Agricultural Mechanization | Open Access
    YANG Yinsheng, WEI Xin
    2023, 5(4):  150-159.  doi:10.12133/j.smartag.SA202304008
    Asbtract ( 261 )   HTML ( 30)   PDF (870KB) ( 393 )  
    Figures and Tables | References | Related Articles | Metrics

    [Significance] With the escalating global climate change and ecological pollution issues, the "dual carbon" target of Carbon Peak and Carbon Neutrality has been incorporated into various sectors of China's social development. To ensure the green and sustainable development of agriculture, it is imperative to minimize energy consumption and reduce pollution emissions at every stage of agricultural mechanization, meet the diversified needs of agricultural machinery and equipment in the era of intelligent information, and develop low-carbon agricultural mechanization. The development of low-carbon agricultural mechanization is not only an important part of the transformation and upgrading of agricultural mechanization in China but also an objective requirement for the sustainable development of agriculture under the "dual carbon" target. Progress] The connotation and objectives of low-carbon agricultural mechanization are clarified and the development logic of low-carbon agricultural mechanization from three dimensions: theoretical, practical, and systematic are expounded. The "triple-win" of life, production, and ecology is proposed, it is an important criterion for judging the functional realization of low-carbon agricultural mechanization system from a theoretical perspective. The necessity and urgency of low-carbon agricultural mechanization development from a practical perspective is revealed. The "human-machine-environment" system of low-carbon agricultural mechanization development is analyzed and the principles and feasibility of coordinated development of low-carbon agricultural mechanization based on a systemic perspective is explained. Furthermore, the deep-rooted reasons affecting the development of low-carbon agricultural mechanization from six aspects are analyzed: factor conditions, demand conditions, related and supporting industries, production entities, government, and opportunities. Conclusion and Prospects] Four approaches are proposed for the realization of low-carbon agricultural mechanization development: (1) Encouraging enterprises to implement agricultural machinery ecological design and green manufacturing throughout the life cycle through key and core technology research, government policies, and financial support; (2) Guiding agricultural entities to implement clean production operations in agricultural mechanization, including but not limited to innovative models of intensive agricultural land, exploration and promotion of new models of clean production in agricultural mechanization, and the construction of a carbon emission measurement system for agricultural low-carbonization; (3) Strengthening the guidance and implementation of the concept of socialized services for low-carbon agricultural machinery by government departments, constructing and improving a "8S" system of agricultural machinery operation services mainly consisting of Sale, Spare part, Service, Survey, Show, School, Service, and Scrap, to achieve the long-term development of dematerialized agricultural machinery socialized services and green shared operation system; (4) Starting from concept guidance, policy promotion, and financial support, comprehensively advancing the process of low-carbon disposal and green remanufacturing of retired and waste agricultural machinery by government departments.