Deep learning models are widely used in the field of crop detection and recognition. Their advantage lies in optimizing the model by constructing different functional perception layers，which can automatically extract features from input data and achieve end-to-end learning. However，the unknown data processing process in this model leads to a lack of interpretability，which becomes the main obstacle to the application of deep learning. To overcome the shortcomings of insufficient interpretability in deep learning models，researchers have proposed an interpretability method based on class activation mapping. This article summarizes the research progress of the class activation mapping algorithm Grad-CAM in crop disease classification and detection，crop pest detection and recognition，crop variety classification，target crop detection， and other applications. It explains the advantages of the class activation mapping algorithm in visualizing convolutional neural networks with arbitrary structures，and analyzes the shortcomings of class activation mapping algorithms such as low interpretation precision，unstable gradients，lack of evaluation standards，and single application background. It proposes the development trend of building models with high accuracy and interpretability，construction of new interpretive algorithms， establishing unified evaluation standards for interpretable algorithms，and ensuring the correctness of interpretable algorithms.

In response to the problems of numerous model parameters and low detection accuracy in SSD-based cotton leaf disease detection methods, a cotton leaf disease detection method based on improved SSD was proposed by combining the characteristics of cotton leaf diseases. First, the lightweight network MobileNetV2 was introduced to improve the backbone feature extraction network, which provides more abundant semantic information and details while significantly reducing the amount of model parameters and computing complexity, and accelerates the detection speed to achieve real-time detection. Then, the SE attention mechanism, ECA attention mechanism, and CBAM attention mechanism were fused to filter out disease target features and effectively suppress the feature information of jamming targets, generating feature maps with strong semantics and precise location information. The test results on the self-built cotton leaf disease dataset show that the parameter quantity of the SSD_MobileNetV2 model with backbone network of MobileNetV2 was 50.9% of the SSD_VGG model taking VGG as the backbone. Compared with SSD_VGG model, the P, R, F1 values, and mAP of the MobileNetV2 model increased by 4.37%, 3.3%, 3.8%, and 8.79% respectively, while FPS increased by 22.5 frames/s. The SE, ECA, and CBAM attention mechanisms were introduced into the SSD_VGG model and SSD_MobileNetV2 model. Using gradient weighted class activation mapping algorithm to explain the model detection process and visually compare the detection results of each model. The results indicate that the P, R, F1 values, mAP and FPS of the SSD_MobileNetV2+ECA model were higher than other models that introduced the attention mechanisms. Moreover, this model has less parameter with faster running speed, and is more suitable for detecting cotton diseases in complex environments, showing the best detection effect. Therefore, the improved SSD_MobileNetV2+ECA model significantly enhanced the semantic information of the shallow feature map of the model, and has a good detection effect on cotton leaf diseases in complex environments. The research can provide a lightweight, real-time, and accurate solution for detecting of cotton diseases in complex environments.

In view of the problems in operation process of fixed rake-type residual recycling component, such as poor individual profiling effect in film picking, easy clogging of the compound of films, soil and maize stubbles, high power consumption in film picking, and strong disturbance to seedbed soil, in this study, an operation model of intermittent film-picking on full-film mulched double ditches was proposed and an intermittent film picking component was designed. The DEM-MBD coupled algorithm was adopted for numerical simulation on the operation process of the intermittent film-picking component on full-film mulched double ditches, and a comparative analysis was carried out on the seedbed disturbance effect and resistance variation characteristics in film-picking by fixed and intermittent film-picking components. By taking the forward speed in film-picking, cam arrangement angle of the film-picking component and rotating speed of the cam shaft as independent variables, film-picking rate as the response value, a mathematical model between test factors and the film-picking rate was established, to explore the influence order of the factors on film-picking rate, and the optimal working parameters of the intermittent film-picking component were obtained as follows: the forward speed in film-picking was 2 km/h, cam arrangement angle was 180°, rotating speed of the cam shaft was 120 r/min. Under the optimal parameter combination, the average film-picking rate of the simulation test was 96.1%. Field test showed that, the average film-picking rate of the intermittent film-picking component was 95.6%, and 0.5% higher than that of the simulation test. The working condition of the sample machine was basically consistent with the simulation process, and can accurately represent the operation mechanism of intermittent film-picking on full-film mulched double ditches, showing that the established discrete element simulation model and its parameters were accurate and reasonable.