SDM Net: a lightweight network for intelligent diagnosis of apple leaf diseases based on self-distillation learning

Apple leaf diseases seriously affect apple yield and fruit quality. Traditional disease recognition methods rely on manual feature extraction, which is inefficient and difficult to adapt to complex field environments. With the development of agricultural informatisation and intelligence, visual recognition technology based on deep learning provides a new approach for accurate diagnosis of crop diseases. However, existing models often suffer from high computational complexity, insufficient feature extraction for small lesions, and poor robustness in complex backgrounds, which limits their practical application in resource-constrained agricultural scenarios. To address these limitations, we propose SDM-Net (Self-Distillation-MobileNetV2). This study proposes an apple leaf disease recognition model, SDM-Net, which integrates self-distillation with a CAA attention mechanism. First, channel pruning is applied to cut computational parameters and lighten the model. Next, we propose a CAA attention module to enhance the detection of minute lesions. Finally, we leverage depthwise separable convolution and a self-distillation strategy to build an adaptive learning module, which enables the model to maintain high recognition robustness and accuracy even in complex scenarios. The results showed that the recognition accuracy of the SDM-Net model reached 93.49 %, with an accuracy of 89.08 %. Compared with ShuffleNetV2, MobilenetV3, RegNet, EfficientNetV2, Swin-TransformerV2, and StarNet, it performed better, and the model’s floating-point operation was only 5.88 M, with a parameter size of 2.67 M, especially suitable for resource-constrained agricultural application scenarios. This study provides an effective solution for identifying pests and diseases in agricultural information processing, promoting the development of precision agriculture.