Estimation of snow depth from AMSR-2 based on an AutoML method over the Qinghai-Tibet Plateau

Abstract. Snow depth (SD) is a crucial parameter for describing the spatiotemporal variations of snow cover, and passive microwave SD products (10–25 km) are widely used for monitoring SD changes. However, as one of the three major snow-covered regions in China, the Qinghai-Tibet Plateau (QTP) has complex terrain and rapid change in snow cover with strong spatial heterogeneity, making it difficult for coarse-resolution SD products to accurately describe its spatiotemporal characteristics. This study proposes a high spatial resolution (500 m) SD estimation method based on AMSR-2 brightness temperature (BT) data and an Automated Machine Learning (AutoML). Firstly, using Pearson correlation coefficient, 19 key factors influencing SD, including AMSR-2 BT, slope, and surface roughness, were selected as input data (independent variables) for AutoML. Meanwhile, a passive microwave downscaled SD data and ground-based SD measurements were introduced as dependent variables for AutoML. Then, the AutoML model was trained separately for four different types of snow cover surfaces (forest, grassland, water, and bare land). Finally, through the ten folds cross validation method, the optimal machine learning model for SD estimation under each type of underlying surface coverage was selected, thus sequential SD datasets were obtained for ten-year snow cover periods of the QTP from 2012 to 2021. Results show that the estimated SD values are consistent with ground-based observations (R=0.81), and the accuracy is high with an RMSE of 3.65 cm. Compared with Landsat-8, the estimated SD spatial distribution is consistent with the snow cover extent on optical images, which can provide reliable data for monitoring snow cover changes in mountainous regions.