Boundary-Guided Semantic Context Network for Water Body Extraction from Remote Sensing Images

Automatically extracting water bodies is a significant task in interpreting remote sensing images (RSIs). Convolutional neural networks (CNNs) have exhibited excellent performance in processing RSIs, which have been widely used for fine-grained extraction of water bodies. However, it is difficult for the extraction accuracy of CNNs to satisfy the requirements in practice due to the limited receptive field and the gradually reduced spatial size during the encoder stage. In complicated scenarios, in particular, the existing methods perform even worse. To address this problem, a novel boundary-guided semantic context network (BGSNet) is proposed to accurately extract water bodies via leveraging boundary features to guide the integration of semantic context. Firstly, a boundary refinement (BR) module is proposed to preserve sufficient boundary distributions from shallow layer features. In addition, abstract semantic information of deep layers is also captured by a semantic context fusion (SCF) module. Based on the results obtained from the aforementioned modules, a boundary-guided semantic context (BGS) module is devised to aggregate semantic context information along the boundaries, thereby enhancing intra-class consistency of water bodies. Extensive experiments were conducted on the Qinghai–Tibet Plateau Lake (QTPL) and the Land-cOVEr Domain Adaptive semantic segmentation (LoveDA) datasets. The results demonstrate that the proposed BGSNet outperforms the mainstream approaches in terms of OA, MIoU, F1-score, and kappa. Specifically, BGSNet achieves an OA of 98.97% on the QTPL dataset and 95.70% on the LoveDA dataset. Additionally, an ablation study was conducted to validate the efficacy of the proposed modules.