SSMM-DS: A semantic segmentation model for mangroves based on Deeplabv3+ with swin transformer

Mangrove wetlands play a crucial role in maintaining species diversity. However, they face threats from habitat degradation, deforestation, pollution, and climate change. Detecting changes in mangrove wetlands is essential for understanding their ecological implications, but it remains a challenging task. In this study, we propose a semantic segmentation model for mangroves based on Deeplabv3+ with Swin Transformer, abbreviated as SSMM-DS. Using Deeplabv3+ as the basic framework, we first constructed a data concatenation module to improve the contrast between mangroves and other vegetation or water. We then employed Swin Transformer as the backbone network, enhancing the capability of global information learning and detail feature extraction. Finally, we optimized the loss function by combining cross-entropy loss and dice loss, addressing the issue of sampling imbalance caused by the small areas of mangroves. Using GF-1 and GF-6 images, taking mean precision (mPrecision), mean intersection over union (mIoU), floating-point operations (FLOPs), and the number of parameters (Params) as evaluation metrics, we evaluate SSMM-DS against state-of-the-art models, including FCN, PSPNet, OCRNet, uPerNet, and SegFormer. The results demonstrate SSMM-DS's superiority in terms of mIoU, mPrecision, and parameter efficiency. SSMM-DS achieves a higher mIoU (95.11%) and mPrecision (97.79%) while using fewer parameters (17.48M) compared to others. Although its FLOPs are slightly higher than SegFormer's (15.11G vs. 9.9G), SSMM-DS offers a balance between performance and efficiency. Experimental results highlight SSMM-DS's effectiveness in extracting mangrove features, making it a valuable tool for monitoring and managing these critical ecosystems.