DRFormer: Dynamic Recursively-Aware Transformer for Urban Watershed Segmentation

Accurate segmentation of water bodies in urban high-resolution remote sensing images is essential for effective water resource management and environmental monitoring. However, due to complex backgrounds and low contrast in such images, existing methods often fail to effectively capture both global context and local details, resulting in limited segmentation accuracy, especially for small water bodies. To address this, Dynamic Recursively-aware Transformer (DRFormer) is proposed, leveraging dynamic recursive perception for efficient and accurate segmentation. Specifically, we introduce an improved BiFormer encoder with Hybrid Attention (HA) Transformer Blocks, replacing the conventional attention mechanism to enhance feature extraction for small water bodies. Additionally, to preserve detail fidelity while reducing computational overhead, we propose the Lightweight Feature-adaptive Upsampling (LFU) module, which effectively preserves details and aggregates multiscale features with minimal computational cost. This ensures high segmentation accuracy while maintaining efficiency. Extensive experiments on the ZKXT, GF-UWS, and WHDLD-Water datasets demonstrate the effectiveness and robustness of our model. DRFormer achieves F1 scores of 96.45%, 95.28%, and 96.67%, and IoU scores of 91.95%, 91.00%, and 92.61%, respectively, across these datasets. Furthermore, DRFormer significantly reduces computational overhead, with 19M parameters and 4.64 GFLOPs, achieving a favorable balance between performance and efficiency. These results highlight DRFormer’s potential as a robust and efficient solution for urban water resource management and real-time environmental monitoring applications.