Dual branch negotiation model for robust water resource monitoring with spectral ambiguity and boundary degradation.

Despite the remarkable progress in remote sensing water-body segmentation, existing methods still struggle to jointly preserve fine boundary fidelity and achieve robust semantic disambiguation under cross-domain shifts, undermining applications like flood risk assessment and urban water monitoring. This limitation largely stems from purely data-driven fitting with little explicit control over aliasing artifacts and spectrally induced physical ambiguities. To address this issue, we propose an Enhanced Anti-aliasing dual-branch Segmentation network (EAAS-Net) that integrates sampling theory and spectral physics. At the signal level, an Edge-Adaptive Anti-aliasing Gate (EAAG) is inserted before downsampling to explicitly suppress boundary aliasing, improving sub-pixel geometric fidelity along narrow rivers and complex shorelines. At the semantic level, a physical-guided branch and a multiscale Adaptive Spectral-Spatial Enhancement Fusion (ASEF) module extract domain-invariant spectral representations, enabling reliable separation of water bodies from spectrally similar shadows. At the decision level, a Dynamic Mutual Fusion Decoder (DMF-Decoder) employs a two-round negotiation scheme to progressively reconcile conflicts between visual features and physical priors across scales. Experiments on the GLW dataset and the challenging cross-domain GLH-water dataset show that EAAS-Net achieves highly stable mean Intersection over Union (mIoU) scores of 0.926 and 0.849, respectively, while maintaining sharp boundaries. Furthermore, with 60.2 M parameters, the model maintains a highly efficient inference speed of 18.1 ms per image, demonstrating an optimal balance between sub-pixel accuracy and computational cost, indicating strong potential for near-real-time, large-scale operational deployment.