Multi-Scale Attentive Deep Learning Framework for Soybean Leaf Disease Classification with Diversity-by-Design Ensembling

Abstract

Soybean foliar diseases are separated less by global leaf appearance than by the morphology of lesions occupying a small fraction of the frame, and the categories that matter most clinically rust, target leaf spot and bacterial pustule differ only in texture at the scale of a few pixels. Conventional transfer-learning pipelines discard exactly this evidence, because the final feature map of a standard backbone represents each activation with a 32-pixel stride at typical input resolutions. This article proposes three complementary architectures that address the problem from different directions. MSAF-DCNet extracts three resolution taps from an EfficientNetV2-S trunk, recalibrates each independently with convolutional block attention, and fuses concatenated first- and second-order pooled statistics. SoyLeaf-LiteNet couples a compact trunk with squeeze-and-excitation gating and generalised-mean pooling for on-device deployment. DSCA-Net processes a shared view through two trunks selected for divergent inductive biases so that their errors decorrelate. A class-balanced sampler, MixUp regularisation, a two-phase schedule with checkpoint guarding, eight-view dihedral test-time augmentation and validation-weighted soft-voting complete the framework. On a 13-class, 1,158-image benchmark, MSAF-DCNet attains 91.95% test accuracy and 93.51% macro-F1, exceeding six established backbones trained under an identical protocol, while SoyLeaf-LiteNet recovers 89.08% using 7.27 M parameters and one third of the training time. The proposed ensemble attains the best macro-precision (94.12%) and macro ROC-AUC (0.9976). A class-wise analysis localises almost the entire residual error to one phytopathologically coherent confusion cluster and shows it to be a decision-threshold rather than a discrimination failure, since rust one-vs-rest AUC remains above 0.979 while its recall falls as low as 0.529.