A Snow Depth Retrieval Method Based on Super-Resolution Brightness Temperature Reconstruction and Multimodal Feature Synergy

Snow plays a critical role in the hydrological ecosystem of the Tibetan Plateau (TP), influencing both regional water availability and climate dynamics. Passive microwave remote sensing is a powerful tool for large-scale continuous snow depth monitoring, yet the low spatial resolution and saturation limitation of brightness temperature (TB) make it difficult for a single passive microwave data to address the challenges posed by the complex topography and fragmented snow distribution on the TP. To address this issue, an area-to-point downscaled daily SD retrieval method that accounts for spatial heterogeneity was proposed in this study using multi-modal remote sensing data. First, we developed a new algorithm to reconstruct the high-resolution TB from AMSR2 considering the effects of snow and land cover types. The recursive feature elimination (RFE) was used to select the optimal subset of features, including active and passive remote sensing information, meteorological data, and geographic factors, and then a multi-modal feature-driven convolutional neural network (CNN)-long short-term memory network (LSTM) was used to retrieve daily 500 m SD data. The method effectively captures the dynamic spatial variation features of SD. The validation of the proposed TB reconstruction downscaling method using the enhanced-resolution TB data showed the RMSE ranged from 2.65 to 6.81 K at different frequencies. The contribution of multi-modal features was analyzed based on SHapley Additive exPlanations (SHAP) values. Snow cover accumulation, H-pol TB at 89 GHz, and latitude were key features, although the relative contributions of these features varied significantly across different environments. In the site-based independent validation, the RMSE of the proposed method was 2.33 cm, demonstrating better accuracy than existing SD retrieval methods in the TP. The use of multi-modal features improved accuracy by 1.67 cm (41.7%) compared to passive microwave features only. The proposed SD retrieval algorithm can obtain high-resolution and high-precision SD information, thereby supporting water resource management and climate change monitoring over the TP.