HTFU-Net: A Cross-Modal Deep Learning Framework for Antarctic Surface Melt Mapping Using Multisource Microwave Remote Sensing Data

Abstract

Accurate monitoring of Antarctic surface melt is essential for understanding ice-sheet mass balance and assessing future sea-level rise. Existing microwave-based melt products generally suffer from a tradeoff between spatial and temporal resolution, limiting their ability to capture both fine-scale melt patterns and continuous melt evolution. To address this challenge, this article proposes the Hierarchical Texture–Microwave Fusion U-Net (HTFU-Net), a cross-modal deep learning framework that integrates multisource microwave observations and synthetic aperture radar (SAR) texture information for Antarctic surface melt detection. The framework combines Sentinel-1-derived gray-level co-occurrence matrix texture features with advanced scatterometer backscatter observations and special sensor microwave imager/sounder brightness temperature measurements. A dual-branch encoder architecture is designed to separately learn microwave and texture representations, while a directional cross-modal attention module is introduced to intelligently bridge the datasets by dynamically identifying meaningful relationships between texture and microwave features to enhance feature interaction and adaptive fusion across modalities. Experiments conducted over representative Antarctic melt regions, including the Antarctic Peninsula, Queen Maud Land, and East Antarctic ice shelves, demonstrate that the proposed method achieves an overall accuracy of 90.21%, an F1-score of 83.35%, and an intersection-over-union of 71.46%, outperforming conventional U-Net-based approaches. Results indicate that SAR texture features significantly improve melt boundary delineation and spatial consistency, while the proposed cross-modal fusion strategy enhances the detection of weak and heterogeneous melt signals. The resulting framework produces daily Antarctic surface melt maps at 500 m spatial resolution and effectively mitigates the long-standing tradeoff between spatial and temporal resolution in microwave melt monitoring. These findings demonstrate the potential of integrating complementary microwave observations and SAR texture information for large-scale cryosphere monitoring and surface melt characterization.