Abstract:

[Objective] Sugarcane is an important cash crop, and its health status affects crop yields. However, under natural environmental conditions, the identification of sugarcane leaf diseases is a challenging problem. There are various issues such as disease spots on sugarcane leaves being blocked and interference from lighting, which make it extremely difficult to comprehensively obtain disease information, thus significantly increasing the difficulty of disease identification. Early image recognition algorithms cannot accurately extract disease features and are prone to misjudgment and missed judgment in practical applications. To solve the problem of identifying sugarcane leaf diseases under natural conditions and break through the limitations of traditional methods, a novel identification model, XEffDa was proposed in this research. [Methods] The XEffDa model proposed implemented a series of improvement measures based on the ensemble learning framework, aiming to significantly improve the accuracy of classifying and identifying sugarcane leaf diseases. Firstly, the images in the sugarcane leaf disease dataset under natural conditions were pre-processed. Real-time data augmentation techniques were used to expand the scale of the dataset. Meanwhile, HSV image segmentation and edge-processing techniques were adopted to effectively remove redundant backgrounds and interference factors in the images. Considering that sugarcane leaf disease images were fine-grained images, in order to fully extract the semantic information of the images, the transfer learning strategy was employed. The pre-trained models of EfficientNetB0, Xception, and DenseNet201 were loaded respectively, and with the help of the pre-trained weight parameters based on the ImageNet dataset, the top layers of the models were frozen. The performance of the validation set was monitored through the Bayesian optimization method, and the parameters of the top-layer structure were replaced, thus achieving a good balance between optimizing the number of model parameters and the overall performance. In the top-layer structure, the improved ElasticNet regularization and Dropout layer were integrated. These two mechanisms cooperated with each other to double-suppress overfitting and significantly enhance the generalization ability of the model. During the training process, the MSprop optimizer was selected and combined with the sparse categorical cross - entropy loss function to better adapt to the multi-classification problem of sugarcane disease identification. After each model completed training independently, an exponential weight-allocation strategy was used to organically integrate the prediction features of each model and accurately map them to the final disease categories. To comprehensively evaluate the model performance, the accuracy indicator was continuously monitored, and an early-stopping mechanism was introduced to avoid overfitting and further strengthen the generalization ability of the model. Through the implementation of this series of refined optimization and integration strategies, the XEffDa model for sugarcane leaf diseases was finally successfully constructed. [Results and Discussions] The results of the confusion matrix showed that the XEffDa model performed very evenly across various disease categories, and all indicators achieved excellent results. Especially in the identification of red rot disease, its F₁-Score was as high as 99.09%. This result was not only higher than that of other single models (such as EfficientNetB0 and Xception) but also superior to the combination of EfficientNetB0 and other deep networks (such as DenseNet121 and DenseNet201). This indicated that the XEffDa model significantly improved the ability to extract and classify features of complex pathological images by integrating the advantages of different network architectures. The comparison experiments of different models showed that the recognition accuracy of the XEffDa model reached 97.62%. Compared with the single models of EfficientNetB0 and Xception, as well as the combined models of EfficientNetB0 and other deep networks, the recognition accuracy increased by 9.96, 6.04, 8.09, 4.19, and 1.78 percentage points, respectively. The fusion experiments further showed that the accuracy, precision, recall, and F₁-Score of the network improved by ElasticNet regularization increased by 3.76, 3.76, 3.67, and 3.72 percentage points respectively compared with the backbone network. The results of the maximum-probability scatter plot showed that the proportion of the maximum prediction probability value not lower than 0.5 was as high as 99.4%. [Conclusions] The XEffDa model demonstrated stronger robustness and stability. In the identification task of small sugarcane leaf disease datasets, it showed good generalization ability. This model can provide a powerful reference for the accurate prevention and control of sugarcane crop leaf diseases in practical scenarios, and it has positive significance for promoting the intelligent and precise management of sugarcane production.

Key words: sugarcane leaf diseases, image recognition, EfficientNet, Xception, DenseNet201, model ensemble