As a vital staple crop, rice is susceptible to leaf diseases such as bacterial leaf blight, rice blast, brown spot, and sheath blight during its growth cycle, which severely impact yield and quality. Precise disease segmentation is crucial for early prevention and control, yet it faces challenges including complex backgrounds, diverse lesion morphologies, and similar features. This paper proposes SwinDA-DeepLabv3+, a segmentation network based on Swin Transformer multi-scale feature fusion: it employs Swin Transformer as the backbone network to capture global disease features; replaces traditional ASPP with a dual-attention-enhanced dilated convolution module to enhance focus on sparse lesions; and mitigates sample imbalance through a Dice-Focal hybrid loss function. Experiments demonstrate that this network achieves an mIoU of 83.34% on a four-class rice disease dataset, representing a 4.68% improvement over the original DeepLabv3+.