Abstract
Rapid and accurate identification of crop leaf diseases is essential for informed agricultural decision-making. However, achieving reliable classification remains challenging under conditions such as extreme lighting, complex color variations, and intricate structural backgrounds, particularly when early-stage symptoms are subtle and easily masked by surrounding tissues. To address these challenges, this study proposes a novel network architecture, SREM-Net, which incorporates stylistic and multiscale feature extraction strategies. Specifically, the model introduces the style recalibration MBconv (SRMB) to mitigate feature dilution caused by the coexistence of lesions and complex backgrounds. In addition, the EMF dynamically adjusts the receptive field, enabling the model to capture lesion distributions across the entire leaf while simultaneously emphasizing morphological details, edges, and fine-scale features. To improve interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to generate visual explanations of the detected diseases. On our self-constructed, weather-augmented MCCD dataset, the experimental results demonstrate that SREM-Net outperforms state-of-the-art networks such as LWMobileViT, MobileNetV3-CA, and LWDN, achieving F1-score improvements of 2.13%, 1.21%, and 1.18%, respectively.