SCD-Net: A Lightweight Deep Learning Model for Accurate Soybean Leaf Disease Identification

Abstract

Soybean is a pivotal crop globally, yet its productivity is severely constrained by foliar diseases. Conventional manual scouting for disease diagnosis is often subjective and inconsistent, limiting its utility for large-scale ecological surveillance and timely intervention. Addressing this, we propose SCD-Net, a novel deep learning architecture for intelligent agricultural information processing. Our model integrates an optimized Swin Transformer backbone with a cross-covariance attention mechanism to capture multi-scale symptom relationships and a dynamic tanh module for adaptive feature normalization. This design enhances the extraction of discriminative pathological features from leaf imagery. Evaluated on a dataset of 7,516 images encompassing healthy leaves and three major diseases (i.e., grey spot, macular, and mosaic), SCD-Net achieves a recognition accuracy of 98.47%, outperforming benchmarks like EfficientFormer. With a compact design of 36.59M parameters and an inference speed of 0.204 seconds per image, it is well-suited for real-time, field-deployable applications. Our work provides a scalable and pathology-aware framework, demonstrating strong potential for advancing plant phenotyping, ecological monitoring, and sustainable agroecosystem management through advanced information systems.

Keywords: Agricultural information processing, Soybean leaf, deep learning, Optimized Swin Transformer, Cross-covariance attention mechanism