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

Area of Science:

• Earth Science
• Remote Sensing
• Hydrology

Background:

• Existing soil moisture products lack the high spatial resolution required for applications like precision agriculture.
• Sub-field scale resolution is ideal for agricultural soil moisture monitoring.

Purpose of the Study:

• To identify the optimal machine learning approach for downscaling 1-km resolution soil moisture data.
• To evaluate Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGB) for downscaling soil moisture.

Main Methods:

• Utilized satellite-derived soil moisture from NASA's Soil Moisture Active Passive (SMAP) and model-based data from the North American Land Data Assimilation System (NLDAS).
• Applied RF, SVM, and XGB machine learning techniques to downscale soil moisture (volumetric water content and percentiles) over CONUS.
• Employed SHapley Additive exPlanations (SHAP) to identify influential features for downscaling.

Main Results:

• Random Forest (RF) demonstrated the highest accuracy for downscaling both volumetric water content (VWC) and soil moisture percentiles.
• Extreme Gradient Boosting (XGB) showed comparable accuracy to RF but offered faster processing times.
• Support Vector Machine (SVM) resulted in larger errors and slower execution compared to RF and XGB.

Conclusions:

• RF is the top-performing model for downscaling soil moisture, with XGB as a viable, faster alternative.
• Elevation and precipitation are key predictors for RF downscaling, while XGB relies on a broader set of meteorological and topographical features.
• The findings support the development of higher-resolution soil moisture products for improved agricultural and hydrological applications.