SSH-Net: Spectral-Spatial Enhanced Hierarchical Network for Robust Polymorphic River Extraction From Sentinel-2 Imagery

Abstract

Efficient extraction of rivers from remote sensing imagery is crucial for river environment monitoring, flood disaster warning, and hydrological studies. Although deep learning methods have been widely applied to water body segmentation, accurate extraction of polymorphic rivers in complex backgrounds remains challenging. To this end, this article proposes a spectral-spatial enhanced hierarchical network, improving river extraction robustness and accuracy. The network incorporates three key components: a spectral feature enhancement adapter to enhance spectral discrimination and suppress background interference; a spatial continuity context block that combines convolutional local features with transformer-based global modeling to capture long-range spatial dependencies; and a multiscale context integrator to adaptively represent rivers of varying widths and morphologies through hierarchical dilated convolution and pyramid pooling. To address class imbalance problem of low river pixel share, a joint Tversky-BCE loss function is introduced to balance boundary details and global accuracy. Experiments conducted on a self-built global polymorphic rivers dataset derived from sentinel-2 imagery demonstrate that the proposed model achieves an overall accuracy of 97.8%, intersection over union of 80.4%, F1 score of 87.3%, with an average inference time of 0.243 seconds per image. Comparative and ablation studies validate the effectiveness of the proposed modules and show superior performance over existing benchmarks in complex scenarios. The results provide a robust solution for intelligent river monitoring and dynamic analysis using remote sensing imagery.