Diagnosis of soybean bacterial blight progress stage based on deep learning in the context of data-deficient

Abstract

Soybean bacterial spot disease is a predominant disease that impairs the quality and yield of soybean produce. In order to minimize production losses to the greatest extent possible, rapid detection and identification of disease progression stages are crucial. However, the conventional techniques used for identifying bacterial spot disease exhibit limitations in terms of providing real-time and objective analysis for dynamically identifying significant differences in the disease manifestation. In the present study, a novel sliding segmentation method was proposed for solving the problem of insufficient data volume, which can diagnose soybean disease stages in a timely and accurate manner, take measures to reduce the impact of diseases on soybeans, and solve the problem of difficulty in obtaining disease datasets. Further, a convolutional neural network (CNN) based image recognition method for soybean leaves at different disease stages was proposed. Different depth learning classical network models were used for transfer learning, and the characteristics of soybean leaf spots in different stages and the differences between diseased leaves and normal leaves were considered. The focus of the experiment was on optimizing a network model by adjusting three factors: the optimization algorithm, epoch number, and batch size number. The aim was to enhance the recognition accuracy of the model. The experimental results show that the optimal network model was based on Swin Transformer, with an optimization algorithm of SGD, epoch, and batch size of 200 and 8, respectively. In the experiment, the optimal model achieved an accuracy of 99.64 % in identifying soybean leaf disease stages. The model can accurately identify soybean leaf diseases at different stages, meet practical production needs, prevent pesticide misuse, reduce yield losses caused by soybean leaf diseases, provide a reference for identifying the development process of leaf diseases, and also provide theoretical support for the establishment of soybean disease warning systems.