Corresponding Author: Palvi Sharma
Multi-crop plant disease automatic classification in precision agriculture requires an effective solution but the state-of-the-art deep learning architectures are prone to suffering background shortcut problem, higher inference latency, and inadequate visualization interpretability. This paper presents a better performing framework combining ConvNeXt-Tiny with an innovative Residual Spatial Attention Module (RSAM) to solve the fine-grained diagnosis problem for 38 plant pathology targets. On a dataset of 10,876 unseen images, the proposed framework demonstrates a Top-1 classification accuracy of 96.85%, a precision of 96.95%, a recall of 96.48%, and a macro F1-score of 96.71% while being superior to ResNet-50 (+2.20%) and being 0.97 ms faster per image (7.15 ms/img with 28.12 M parameters). Explainable AI (XAI) analysis with the help of Grad-CAM shows the ability of the RSAM block to suppress background and soil artifacts, directing 88.42% of activation energy into lesion areas and producing insignificant background suppression effect of 1.12%. The t-SNE projections of latent feature manifolds demonstrate good inter-class clustering.