Estimation of Snow Depth from AMSR2 and MODIS Data based on Deep Residual Learning Network

Advanced Microwave Scanning Radiometer 2 (AMSR2) brightness temperature (TB) observations have long been utilized for snow depth (SD) estimation. However, the traditional approaches which are based on ‘point-to-point’ predictions ignore spatial heterogeneity within a AMSR2 pixel and limited by coarse resolution of sensor. To solve these problems, novel deep ‘area-to-point’ SD estimation model, residual learning network combining convolutional neural networks (CNN) blocks, was proposed. The model utilizes all channels TB data along with Moderate-resolution Imaging Spectroradiometer (MODIS) normalized difference index (NDSI) auxiliary geographic information. Taking Qinghai-Tibet Plateau (QTP) as study area, 0.005° over 2019–2020 season is estimated, accuracy validated in situ from 116 stations. results show that: (1) proposed shows desirable root mean square error (RMSE), absolute (MAE), bias (MBE), coefficient determination (R2) method 2.000 cm, 0.656 −0.013 0.847, respectively. (2) slightly larger medium elevation or slope grassland areas, RMSE 2.247 3.084 2.213 (3) has most satisfactory performance low-elevation regions, only 0.523 cm. indicate that through considering cover utilizing high information presented MODIS product, good accuracy, promising application other regions.