Improving fractional snow cover retrieval from passive microwave data using a radiative transfer model and machine learning method

Optical sensors are subject to cloud obscuration and sunlight dependence, resulting in large proportions of missing snow cover information. Microwave sensors are a good alternative to snow cover monitoring in all weather conditions. Thus far, few studies in the literature have directly derived the fractional snow cover (FSC) from passive microwave data, and none have considered the relationship between FSC and brightness temperature (TB). This study firstly explores the FSC–TB relationship with a radiation transfer model, exhibiting that no generic function can properly describe the nonlinear and complex FSC–TB relationship. Therefore, a new algorithm based on machine learning method was designed to improve FSC retrieval from TB data, considering other auxiliary information, including soil property, land surface, and geography information. Benchmarked against the MODIS reference FSC, the FSC retrieval model performed well with an average correlation coefficient of 0.70, the mean absolute error ranging from 0.15 to 0.17, and the root-mean-square-error ranging from 0.19 to 0.21. The generated FSC maps effortlessly characterized the seasonal dynamics and spatial distribution patterns of snow cover; time series analysis with three AmeriFlux stations observation indicated effective capture of snowpack evolution process by the generated FSC. In addition, the verification of snow mapping capability using snow depth measurements from 13,521 stations indicate it was relatively stable with overall accuracy greater than 0.88. For precise monitoring of snow cover extent in all weather conditions, particularly for sub-pixel snow cover areas, the development of the FSC estimation scheme with TB data should be extensively encouraged and implemented.