Random Forest Based Downscaling of SMAP L4 Soil Moisture and Subsequent Use in Landslide Slope Stability Analysis

Soil moisture has a strong influence on the stability of shallow colluvial hillslopes. Thus, it is often necessary to monitor seasonal variations of soil moisture at individual sites of interest. This work demonstrated a workflow showing that satellite-based soil moisture data can be downscaled through machine learning and used as a source of data for landslide stability analyses. In particular, the NASA Soil Moisture Active Passive (SMAP) Level 4 Root Zone Soil Moisture (L4_SM) product provides near-real-time measurements of soil moisture on a gridded spatial resolution of 9 km. Such resolution, however, is too coarse for soil moisture measurements at a single site. Therefore, satellite-based soil moisture data must be downscaled for use in slope stability assessment at individual sites. This study used a machine learning (ML) technique, Random Forest Regression, to downscale the SMAP Level 4 Root Zone Soil Moisture product and applied the results to landslide slope stability analyses. We modeled the L4_SM data based on 1 km NASA Moderate Resolution Imaging Spectroradiometer (MODIS) products. First, the MODIS products were upscaled from 1 km to 9 km and used as features, combining with L4_SM data as labels to train the Random Forest model. Then, the L4_SM data were downscaled from 9 km to 1 km through the trained model and locally calibrated to that of ground-based data by linear regression. The processed SMAP data were served as the input of an infinite-slope stability model, which was used to investigate incipient slope failure conditions. The slope stability results were verified using known landslide occurrences in the Commonwealth of Kentucky. Results of this work showed that the stability models constructed using downscaled L4_SM data functioned well at detecting incipient conditions at the investigated sites. The purpose behind this study was to develop a process flow routine through which SMAP L4_SM data can be downscaled through machine learning as well as investigating the potential efficacy of utilizing downscaled L4_SM to conduct landslide stability analyses.