Soybean, as one of the most economically important crops worldwide, plays an irreplaceable role in multiple domains, including food production, animal feed, and industrial raw materials. However, soybean leaf diseases pose a severe threat to both yield and quality, making accurate disease identification critically important for ensuring stable agricultural production. This study systematically analyzes common soybean leaf diseases and constructs a high-quality dataset through manual screening and dataset integration. Building upon this foundation, this paper proposes a high-precision soybean leaf disease recognition model termed Multi-scale Feature Fusion Attention Network (MSFFANet). First, a novel Multi-scale Spatial and Channel Fusion Module (MSCFM) is introduced, which constructs differentiated receptive fields to simultaneously capture local detail and global morphological information from both spatial and channel dimensions. Furthermore, an innovative attention mechanism—Median and Standard Deviation Attention Mechanism (MSAM) is proposed. By refining the pooling strategy, MSAM quantifies feature saliency through the computation of the median and standard deviation of feature maps, which effectively suppresses background interference while directing model attention toward lesion regions, thereby improving recognition accuracy. In addition, a label smoothing technique is incorporated to mitigate overconfident predictions toward any particular category. Experimental results demonstrate that MSFFANet achieves competitive and reliable performance in identifying various soybean leaf diseases, with accuracy, precision, recall, and F1-score reaching 96.47%, 96.85%, 96.86%, and 96.85%, respectively. Moreover, the model exhibits certain generalization ability on leaf disease datasets from other crops, such as apple and coffee. The findings of this study offer supportive technical references for the early diagnosis and precise prevention of soybean leaf diseases.
An et al. (Mon,) studied this question.