A New Method to Simulate the Microwave Effective Snow Grain Size in the Northern Hemisphere Without Using Snow Depth Priors

Abstract

The snow water equivalent (SWE) of a snowpack represents the amount of water it contains. SWE retrieval with passive microwave remote sensing techniques is notably affected by snow metamorphism. The effective snow grain size (effGS) described in the GlobSnow methodology can be retrieved over areas where prior snow depth data are available. However, this dependency on the availability of prior knowledge results in increased retrieval uncertainty in areas where such information is unavailable. In this work, we proposed a new method to predict the effGS using a random forest (RF) model. The results indicated that the geographical location, orographic effects, and seasonal characteristics are important for retrieving the effGS, and they are important components of the kriging interpolation of the daily effGS in the GlobSnow methodology, resulting in spatial and temporal autocorrelation phenomena. We assessed the performances of snow depth priors from either station measurements or gridded products in predicting effGS and found that the RF model is promising for effGS prediction even without snow depth priors. The application of spatially independent verification and 10-fold cross-validation (10-CV) techniques revealed that the effGS estimates generally agreed with GlobSnow reference data over Eurasia, with overall unbiased root mean square error (unRMSE) values of 0.15.0.18 mm, but high errors were observed over North America due to terrain variability and heterogeneous forest cover, with overall unRMSE values ranging from 0.22.0.31 mm. Compared to the GlobSnow methodology, the proposed method does not rely on ground-based prior information, and improved computational efficiency can be achieved.

Keywords

• Effective grain size
• Global Snow Monitoring for Climate Research (GlobSnow)
• machine learning
• snow depth priors
• snow emission model