Comparison of different machine learning techniques for downscaling SMAP and NLDAS soil moisture over CONUS.

Abstract

Although many soil moisture products are available, they do not have a high enough spatial resolution for many applications. For example, soil moisture for agriculture applications is best at sub-field scale resolution. The goal of this study is to identify the best approach for downscaling 1-km soil moisture. Two distinct sources of soil moisture data and two units of soil moisture (volumetric water content (VWC) and percentiles (a standard form of soil moisture value for different purpose)) were utilized: satellite-derived soil moisture from NASA's Soil Moisture Active Passive (SMAP) mission (2015-2021) and model-based soil moisture from the North American Land Data Assimilation System (NLDAS) (2001-2021). Three machine learning techniques were applied to generate higher resolution soil moisture over CONUS: random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGB). SHapley Additive exPlanations (SHAP) values were generated to determine which features are most important for downscaling soil moisture. This study found that RF had the best performance for downscaling volumetric water content (VWC) (MAE for SMAP = 0.0816; MAE for NLDAS = 0.0828) and soil moisture percentiles (MAE for SMAP = 0.217; MAE for NLDAS = 0.226). XGB also had good accuracy. The difference in accuracy between RF and XGB is negligible, and XGB was faster to run. This makes it a good choice for downscaling soil moisture. SVM had larger errors for downscaling VWC, and it was slower to run. Elevation and precipitation are the most influential features in the RF downscaling of SMAP and NLDAS soil moisture. Dew point temperature, antecedent precipitation index, elevation, and maximum temperature are the most influential features in the XGB downscaling of SMAP and NLDAS soil moisture.