PIPNet EA for Robust and Explainable Crop Weed Classification Using Enhanced Prototype Evidence Aggregation

Abstract Deep learning models can achieve high accuracy on clean crop–weed images but often lose performance when field images are affected by noise, blur, poor illumination, motion, or compression. This study presents PIPNet-EA, an enhanced prototype-based interpretable prediction framework that combines top-κ spatial evidence aggregation with confidence-weighted prototype reasoning. Five deep learning architectures—ResNet50, ResNet101, ViT-B/16, PIPNet-R50, and PIPNet-R101—were evaluated on two crop–weed image datasets under seven common image corruptions at three severity levels: mild, moderate, and severe. The evaluation generated 210 model–corruption–severity accuracy measurements, while the accuracy retention rate (ARR) was used to quantify the proportion of clean-image performance preserved under degradation. PIPNet-R50 showed the strongest overall robustness. On the challenging field dataset, its mean accuracy was 0.641, 0.568, and 0.437 under mild, moderate, and severe corruption, respectively, compared with 0.594, 0.525, and 0.391 for ResNet50. It retained 98.9%, 87.7%, and 67.5% of its clean performance across the three severity levels, exceeding the corresponding retention of ResNet50. Prototype activation maps remained focused on meaningful plant structures, including leaf tips, midribs, and petiole regions, even under moderate degradation. By combining evidence from multiple spatial locations rather than relying on a single dominant activation, PIPNet-EA reduces the effect of local image corruption while maintaining interpretable predictions. These findings support prototype-based evidence aggregation as a practical approach for developing robust and explainable crop–weed classification systems for precision agriculture.