Soil moisture downscaling based on physics-based hydrological simulations: a downscaling playground and a novel high-resolution downscaling product for the continental USA

High-resolution soil moisture is key for a wide range of applications such as water resources management, agriculture, and natural hazard monitoring and prediction. In the last few decades, the prominent approach for deriving high resolution soil moisture fields over large areas has been downscaling of satellite observations. These approaches, which are mainly machine learning based, use the coarse resolution estimate from the satellite, combine it with other variables which impact soil moisture distribution (e.g., landcover, topography, soil characteristics) and/or other remote sensed products with higher spatial resolution, and use the in-situ soil moisture observations for training/testing.

Here, we follow a different approach, which takes advantage of physics-based hydrological simulations. First, we create a downscaling playground by using historical simulations of the hydrological model ParFlow-CLM over the continental USA (CONUS). We use the 1 km² run as our high-resolution estimate, and an upscaled version (averages over 10x10 km² gridcells) as representative of the coarse resolution estimate. By doing this, we remove two of the biggest issues when downscaling soil moisture: first, we know there’s a perfect match between high- and low-resolution soil moisture and second, we can train and test the model freely over the entire domain, as information is available for every gridcell, without being constrained by the number/locations of in-situ stations. In terms of downscaling approach, we use a random forest model, trained on coarse resolution soil moisture, drainage area, slope, elevation, hydraulic conductivity, porosity, and landcover. We carry out several experiments changing the locations and timing of both the training and testing sets. These experiments allow us, for example, to test whether the in-situ stations available are adequate in number and representative of the entire domain for reliable downscaled products.

Finally, we take advantage of this playground to develop a new downscaling product. We train the same random forest model, but over the CONUS domain, using all gridcells. This results into a model that has learned the spatial scaling of soil moisture between the two resolutions and can predict the 1 km² over CONUS, fed by a 10x10 km² estimate in addition to static predictors. We then use the model in a predictive mode, feeding it the coarse resolution estimate from Soil Moisture Active Passive (SMAP) satellite, creating a high-resolution (1 km²) version of SMAP soil moisture.