Improving the accuracy of gridded snow depth estimation through multi-source data and a machine learning fusion model

Snow depth (SD) provides information on the spatial distribution of snow cover, which is critical to assess water resources and global climate change. Currently, SD can be obtained from passive microwave radiometers, reanalysis models, and in-situ observations. However, the SD data produced from different methods have poor performance in completeness and consistency, and it is difficult to meet the needs of related scientific research. In this study, we developed an SD fusion method based on the random forest algorithm (RF) and used it to generate the SD spatial distribution over China from 2014 to 2018. This method can combine the information from multiple sources of SD data (ground-based, satellite-derived, and reanalysis) to improve the representation of the spatiotemporal distribution of SD. Five SD products (WESTDC, ERA-Interim, CMC, GLDAS-NOAH, and MERRA2) were adopted as input data for constructing the model. In addition, the fusion model was built with consideration of ancillary information (e.g., land cover types, forest cover fraction, geographical information, land cover heterogeneity, surface roughness, and snow class). We evaluated the error of merged SD (RF-SD) data and five SD products against in-situ observations in detail under different land cover types, forest cover fractions, land cover heterogeneity, surface roughness, and snow classification. The results showed that RF-SD data improved the accuracy of SD estimates over China, and increased the Kling-Gupta efficiency (KGE) from 0.21 to 0.64 to 0.73 and lower RMSE (5.1 cm) when compared with five original SD datasets. This method can be effective for integrating the advantages of each SD data source, improving the accuracy of SD estimation and reducing inconsistency among multi-source SD datasets.