Because Antarctic surface melt is mostly driven by local processes, its simulation necessitates high-resolution regional climate models (RCMs). However, the current horizontal resolution of RCMs (≈25–30 km) is inadequate for capturing small-scale melt processes. To address this limitation, we present SUPREME (SUPer-REsolution-based Melt Estimation over Antarctica), a deep learning method to downscale surface melt to 5.5 km resolution using a physically-informed super-resolution model. The physical information integrated into the model originates from observations tied to surface melt, specifically remote sensing-derived albedo and elevation. These remote sensing data, in addition to a Regional Atmospheric Climate Model (RACMO) run at 27 km resolution, account for the diverse drivers of surface melt across Antarctica, facilitating effective generalization beyond the training region of the Antarctic Peninsula. A comparison of SUPREME with a dynamically downscaled RACMO run at 5.5 km over the Antarctic Peninsula shows high accuracy, with average yearly RMSE and bias of 5.5 mm w.e. yr⁻¹ and 4.5 mm w.e. yr⁻¹, respectively. Validation at five automatic weather stations reveals SUPREME's marked improvement with substantially lower average RMSE (81 mm w.e.) compared to RACMO 27 km (129 mm w.e.). Beyond the training region, SUPREME aligns more closely with remote sensing products associated with surface melt than super-resolution models lacking physical constraints. While further validation of SUPREME is needed, our study highlights the potential of super-resolution techniques with physical constraints for high-resolution surface melt monitoring in Antarctica, providing insights into the impacts of localized melting on processes affecting ice shelf integrity such as hydrofracturing.

Surface melt plays a vital role in impacting the polar mass balance and global sea level rise. Over the past decades, synthetic aperture radar (SAR) imagery has garnered considerable attention due to its capacity to provide high-precision and long-term information. However, the traditional SAR-based large-scale surface melt detection methods utilizing co-orbit normalization predominantly depend on reference images and the precise spatial registration to mitigate geometric distortions arising from diverse incidence angles. Consequently, both the absence of reference imagery and the movement of ice sheets and shelves present challenges to the method. In this study, we address this issue by developing a reference-free deep learning network integrating the Convolutional Block Attention Module (CBAM) into DeepLabv3+ to automatically detect surface melt and establishing the surface melt dataset based on multi-temporal Sentinel-1 SAR imagery, encompassing diverse surface conditions of the Antarctic. Our model achieves an accuracy of 95.67%, surpassing the reference-based method and an advanced deep learning-based approach by 4.23% and 4.67%, respectively. Moreover, compared to 500 m resolution UMelt product and the kilometer-level results obtained from Advanced Scatterometer (ASCAT) and Special Sensor Microwave Imager Sounder (SSMIS), our model demonstrates the capability to accurately capture the small-scale melting patterns of ice shelves with a higher spatial resolution of 40 m. Notably, our findings underscore the dispensability of reference imagery in traditional methods through the formidable information extraction capabilities of deep learning. We finally applied the proposed method to extract and analyze the spatiotemporal characteristics of surface melt on the Larsen C Ice Shelf during the 2019/2020 period. The corresponding code of this study is at https://github.com/Tangyu35/Surface-melt-detection.