Exploring the Potential of AMSR2 for Fractional Snow Cover Retrieval Using Multivariate Machine Learning in Western China

Passive microwave (PMW) remote sensing emerged as a promising avenue for developing long-term, global, and daily fractional snow cover (FSC) products. Despite its potential for FSC retrieval, the method remains at an early stage of development with limited understanding of the underlying mechanisms and insufficient research on its applicability and reliability, particularly in Western China (WCN). This study systematically investigated the potential of PMW remote sensing for FSC retrieval using AMSR2 brightness temperature (Tb) data to estimate FSC in WCN during snowpack stability. The proposed model adopts nonlinear machine learning to retrieve FSC from PMW measurements and other auxiliary information. Key variables were selected using the simulated annealing algorithm (SA) and recursive feature elimination algorithm (RFE) to improve model performance, and multiple machine learning methods viz. linear regression method, tree method, support vector machine, ensemble algorithm, nearest neighbor regression, MARS and NNET were constructed. Results demonstrated that Passive microwave showed considerable potential and could be effectively applied to FSC retrieval in WCN. Upon incorporating multivariable, the ensemble algorithms including Random Forest (RF), Light Gradient Boosting Machine (LIGHTGBM), and Bayesian Additive Regression Trees (BART) generally demonstrated stronger performance than the other methods (R² = 0.762, RMSE = 0.191, MAE = 0.121), and the RFE-LIGHTGBM yielded satisfactory overall performance (R² = 0.792, RMSE = 0.182, MAE = 0.111). The results were validated by in situ observations and compared with the Moderate-Resolution Imaging Spectroradiometer (MODIS) FSC products, showing strong agreement and reliable retrieval accuracy with both datasets. Error analysis revealed systematic overestimation at low FSC values (<0.2) and underestimation at high FSC values (>0.8) The results also confirmed strong temporal transferability, exhibiting consistent predictive performance across varying training and testing periods. This research supports the use of passive microwave data for long-term snow cover monitoring from the 1980s to the present and highlights its potential in cryospheric and climate-related studies.