Prior knowledge-informed semantic segmentation framework for precise glacial lake mapping from multimodal imagery

Variation in size and number of glacial lakes (GLs) is important indicators of climate change in the cryosphere and have attracted increasing research attention. However, publicly annotated datasets suitable for computer vision techniques, especially deep learning-based GL mapping, remain scarce. Moreover, existing datasets often contain noisy labels, which affect evaluation results and subsequently hinder downstream processes such as multimodal remote sensing data fusion. To address these issues, we propose a prior knowledge-informed framework for GL segmentation that integrates a self-training-based correction algorithm for glacial lake segmentation dataset (ST-CAGL), which iteratively refines noisy annotations without manual intervention. We also introduce a dual encoder glacial lake semantic segmentation network (DEGSNet) that has a cross-modal feature rectification module (CM-FRM) to enhance multimodal data fusion. Through comparative and ablation experiments, our method achieves an IoU of 86.26% and a DICE of 92.05% at the patch level, yielding improvements of 3.39% in IoU and 2.15% in DICE over the best-performing CNN-based model (UNet), and 5.92% in IoU and 4.33% in DICE over the best-performing Transformer-based model (SegFormer-B3), when these two models are trained with uncorrected labels. In addition, our framework demonstrates superior performance in extracting small GLs, compared to current works. The source code and dataset are available at https://github.com/tanhuizhi123/GlacierSeg . • A novel self-training based correction algorithm for glacial lake segmentation. • The dual encoder glacial lake segmentation network improves segmentation performance. • A large annotated multimodal glacial lake dataset benefits the community. • The framework improves recognition of small glacial lakes with low accuracy.