Identifying Important Features for Downscaling Soil Moisture to 1-km in the Contiguous United States

Abstract. Soil moisture is a fundamental state variable in climatology, meteorology, and hydrology. Many of the available soil moisture products have a coarse spatial resolution that is not useful for agricultural applications. This study used Random Forest to identify which features are most helpful for accurately downscaling soil moisture to 1-km resolution. Fourteen features were considered: precipitation, antecedent precipitation index, maximum daily air temperature, minimum daily air temperature, mean daily air temperature, diurnal temperature range, dew point temperature, elevation, slope, aspect, normalized difference vegetation index, leaf area index (LAI), soil texture, and land use/land cover. The analysis of variable importance was repeated using two different sources of soil moisture data (e.g., satellite-derived soil moisture from NASA’s Soil Moisture Active Passive (SMAP) and model-derived soil moisture from the North American Land Assimilation System (NLDAS)) and two different ways of representing soil saturation (e.g., volumetric water content (VWC) and percentiles). We found that dew point temperature is the most important variable for downscaling SMAP percentiles (0.18), NLDAS VWC (0.27), and NLDAS percentiles (0.17) over CONUS, while elevation is the most important variable for downscaling SMAP VWC (0.28). Dew point temperature is crucial for downscaling in most regions of the United States, except in the South and WestNorthCentral, where elevation is the most important feature. The accuracy of the downscaling varies by region. In the South, SMAP VWC and NLDAS VWC downscaling are relatively accurate, both have mean absolute errors of ~0.07. The MAE values in the South region are 0.196 for SMAP percentiles and 0.175 for NLDAS percentiles.