High-spatial-resolution snow depth data are critical for hydrological, ecological, and disaster research. However, passive microwave snow depth products (10/25 km) no longer meet modern demands for high precision and resolution. This study integrates newly calibrated Enhanced-Resolution Brightness Temperature (CETB) with optical snow cover fraction (SCF) and snow-covered days (SCD) data. Using a deep learning FT-Transformer model, we inverted daily snow depth data at 5 km spatial resolution during snow seasons (October–April) for the Three-River Source region. Compared to China’s long-term snow depth data (25 km), the 5 km snow depth product demonstrates superior accuracy, with RMSE typically below 8.5 cm, providing a reliable foundation for snow resource monitoring in the region.
（1）Calibrated Enhanced-Resolution Brightness Temperature (CETB): Provided by the National Snow and Ice Data Center (NSIDC, https://nsidc.org/data/NSIDC-0630/versions/1). Covers satellite-observed brightness temperatures since 1978, with 1-day temporal resolution and 6.25 km/3.125 km spatial resolution.
(2)Snow Cover Fraction (SCF): Sourced from https://www.sciencedirect.com/science/article/pii/S0924271624003265, with 1-day temporal and 5 km spatial resolution.
（3）Snow-Covered Days (SCD): Obtained from the National Cryosphere Desert Data Center (http://www.ncdc.ac.cn), with 1-day temporal and 500 m spatial resolution.
（4）DEM: Derived from the Shuttle Radar Topography Mission (SRTM, http://srtm.csi.cgiar.org), 90 m spatial resolution.
（5）Land Use: Based on the MCD12Q1 V061 dataset (https://earthexplorer.usgs.gov/), using IGBP annual land cover classification (500 m spatial, 1-year temporal resolution).
(1) Unified batch processing of various data sources with a temporal and spatial resolution of 5 km day by day using the python platform is used to construct the data input for snow depth inversion;
(2) The snow depth inversion model with multiple data fusion is realized by deep learning model training and parameter optimization;
(3) Inversion of snow depth data from Sanjiangyuan using the training-saved model;
(4) Performing a water body mask and then filling the passive microwave radiometer track gap by averaging before and after the day.
Three standard indicators, root mean square error (RMSE), mean absolute error (MAE) and correlation coefficient (R), were used to represent the snow depth error of the inversion, which was evaluated using the ground-truthed snow depths from 2000 to 2020. The validation results show that the RMSE of 5 km snow depth in the Three-River Source Region is in the range of 8-8.5 cm, the MAE is in the range of 5.6-6.5 cm, and the R is greater than 0.7, which is a better accuracy than that of the long time series of snow depth data (25 km) in China (the RMSE is in the range of 10-11.5 cm, the MAE is in the range of 7.5-8.3 cm, and the R is greater than 0.45). The results indicate that the 5 km snow depth of this inversion has good accuracy in the Three-River Source Region. Therefore, this dataset can be used as a reliable data base for evaluating the snow resources in this region.
| # | number | name | type |
| 1 | 2023YFC3206300 | National key R & D plan |
This work is licensed under a
CC BY 4.0.
CSTR:11738.11.NCDC.NIEER-SNOW.DB6816.2025
10.12072/ncdc.nieer-snow.db6816.2025
