A Lightweight Segmentation Model for Northern Corn Leaf Blight Based on an Enhanced UNet Architecture

To address the low segmentation accuracy and high computational complexity of classical deep learning algorithms—caused by the complex morphology of Northern Corn Leaf Blight (NCLB) and blurred boundaries between diseased and healthy leaf regions—this study proposes an improved lightweight segmentation model (termed MSA-UNet) based on the UNet architecture, specifically tailored for NCLB segmentation. In MSA-UNet, three core modules are integrated synergistically to balance efficiency and accuracy: (1) MobileNetV3 (a mobile-optimized convolutional network) replaces the original UNet encoder to reduce parameters while enhancing fine-grained feature extraction; (2) an Enhanced Atrous Spatial Pyramid Pooling (E-ASPP) module is embedded in the bottleneck layer to capture multi-scale lesion features; and (3) the parameter-free Simple Attention Module (SimAM) is added to skip connections to strengthen focus on blurred lesion boundaries. Compared with the baseline UNet model, the proposed MSA-UNet achieves statistically significant performance improvements: mPA, mIoU, and F1-score increase by 3.59%, 5.32%, and 5.75%, respectively; moreover, it delivers substantial reductions in both computational complexity and parameter scale, with GFLOPs decreased by 394.50 G (an 87% reduction) and parameter count reduced by 16.71 M (a 67% reduction). These experimental results confirm that the proposed model markedly improves NCLB leaf lesion segmentation accuracy while retaining a lightweight architecture—rendering it better suited for practical agricultural applications that demand both efficiency and accuracy.