Estimation of Moderate-Resolution Snow Depth in Xinjiang With Enhanced-Resolution Passive Microwave and Reanalysis Data by Machine Learning Methods

The integration of multi-source data into the passive microwave retrieval of snow depth (SD) is vital for accurately capturing large-scale distribution of SD. However, since the exiting SD retrieval algorithms overlook the impact of snow characteristics on brightness temperature, leading to inadequate representation of SD in complex regions. Therefore, this study constructs and optimizes SD retrieval models using four machine learning algorithms, including extreme gradient boosting (XGBoost), light gradient-boosting machine (LightGBM), categorical boosting (CatBoost), and random forest (RF) combing enhanced-resolution passive microwave data. A variety of variables are integrated into the models, encompassing geolocation, topographic indices, land cover, and snow characteristics (fractional snow cover, snow density, and snow grain size) sourced from ERA5-Land reanalysis data. This approach focuses on accurately estimating the spatiotemporal distribution of SD over Xinjiang, characterized by dry-cold snow. The results indicate that: (1) Upon incorporating auxiliary variables, the SD from the CatBoost model demonstrated superior performances over the other algorithms (R2 : CatBoost > LightGBM > XGBoost > RF). (2) The SD product from the CatBoost model exhibits interannual fluctuations. It slightly overestimates shallow snow (SD < 20 cm) and underestimates deep snow (SD > 20 cm). (3) The SD product reveals the spatial differentiation. Areas in northern Xinjiang with high-value for SD (SD > 20 cm) are mainly distributed in the Tianshan Mountains and Altai Mountains. In contrast, the southern Xinjiang with high-value for SD (SD > 10 cm) are largely clustered in the high-elevation regions surrounding the Kunlun Mountains. The findings highlight that employing this approach can lead to the establishment of valuable long-term SD datasets for capturing SD distribution