Interpretable YOLO and mask-guided pooling for maize leaf disease classification in complex field backgrounds

Accurate identification of maize leaf diseases is essential for intelligent agricultural field management and targeted pesticide application. However, most current deep learning recognition research is limited to uniform laboratory backgrounds, presenting a significant domain gap with real-world farmland environments. Existing end-to-end classification models are prone to “Shortcut Learning” when facing complex field backgrounds containing soil, weeds, and varying illumination. They tend to fit environmental noise rather than true pathological characteristics, leading to poor generalization and the risk of illusory high performance. To fundamentally address this overfitting paradox, this paper proposes an innovative mask-guided pooling classification model requiring zero additional parameters (Mgp-YOLO11n). Specifically, we constructed a dataset of 3,092 real-world field images with complex backgrounds covering five maize health categories: Healthy, Gray Leaf Spot (GLS), Northern Leaf Blight (NLB), Northern Leaf Spot (NLS), and Curvularia Leaf Spot (CLS). We utilized a YOLO11-seg model embedded with a Convolutional Block Attention Module (CBAM) to achieve high-quality instance segmentation of disease spots. Subsequently, the traditional Global Average Pooling (GAP) layer is replaced in the downstream classification network. The prior mask serves as spatial hard attention weights, and a dynamic effective area normalization mechanism (MGP operator) is introduced during feature aggregation to achieve physical isolation at the feature space level. Classic large models were introduced as benchmarks for a “background retention” versus “black background” cross-validation, alongside deep interpretability analysis via Grad-CAM. Experimental results indicate that while the baseline YOLO11n-cls model achieves an initial accuracy of 92.30% in complex backgrounds, it decreases sharply to 77.30% in the black background test (in which pixels outside the annotated leaf mask are set to black while the leaf region and image dimensions remain unchanged—this masks rather than crops the background, eliminating background cues without forcing the model to attend specifically to lesions), confirming the defect of relying on background induced spurious correlations. In contrast, Mgp-YOLO11n improves accuracy in complex backgrounds to 93.24%, and its accuracy in black background tests increases to 95.94% with zero additional learnable parameters and negligible impact on real-time inference speed. Grad-CAM heatmaps conclusively demonstrate that the MGP mechanism successfully eliminates background shortcuts, overcomes the receptive field’s edge penetration issue, and forces the network to focus entirely on intrinsic pathological characteristics. This study reconstructs the feature extraction paradigm of the classification head, mitigates the accuracy-centric pitfall in agricultural vision, and provides a robust framework for deploying agricultural edge-intelligent terminals, such as plant protection drones, in highly complex farmland environments.