ShiftingNet : Lightweight Crop Leaf Disease Classification Model With Channel‐Wise Feature Shifting

What the paper says

This paper proposes ShiftingNet, a lightweight classification model based on the improved EfficientNetV2 architecture, which embeds the Channel‐wise Feature Shifting (CFS) operation. Designed as an efficient diagnostic component for agricultural expert systems, ShiftingNet aims to automate classification performance for diverse crop leaf diseases under challenging agricultural conditions. To address the limitations in balancing global and local feature representations and cross‐environment generalisation, we design two novel modules: Channel‐wise Feature Shifting Convolution (CFSConv) and Fused Channel‐wise Feature Shifting Convolution (Fused‐CFSConv). These modules integrate the CFS residual connection and DropPath regularisation into the original MBConv and Fused‐MBConv, while introducing the Squeeze‐and‐Excitation (SE) and Coordinate Attention (CA) mechanisms, respectively. We construct a multi‐source corn dataset, MixCorn, which fuses PlantVillage laboratory images, PlantDoc network images, and CD&S field samples, covering different illumination conditions, backgrounds, and disease scales. Experiments show that ShiftingNet achieves classification accuracies of 99.84% and 99.08% on the PlantVillage and MixCorn datasets, respectively, with only 9.92 M parameters. This demonstrates advantages in knowledge acquisition efficiency and computational cost, providing a theoretical foundation for constructing resource‐constrained mobile expert systems. Robustness evaluation under common perturbations and Grad‐CAM‐based interpretability analysis further validate the model's reliability for automated decision‐making. Ablation studies further confirm that the CFS operation improves feature representation and classification performance.