Towards a high spatial resolution soil moisture product for Switzerland: observations, modeling, downscaling, and upscaling

Information about wetness conditions of the soil is beneficial to hazard prediction and monitoring, water resources management, and weather and climate predictions. For most of these applications, it is important not only that the estimates are accurate, but also at high spatial resolution. Soil moisture can be measured in-situ, remotely or estimated by models. While in-situ measurements are considered the most accurate, networks are very expensive to maintain and provide very sparse coverage, allowing to monitor specific locations but not obtain spatially distributed information. Conversely, remote sensing products can provide estimates over large areas (globally), but lack the spatial resolution required for these applications.

Here, we focus on Switzerland, with the goal of exploring different alternatives for obtaining soil moisture information. We compare existing products (satellite products, in situ observations, hydrological models) with two methods developed here: a downscaling approach, downscaling satellite observations (SMAP), and one based on upscaling of in situ observations. The former overcomes classical limitations of downscaling (i.e., being spatially limited to existing soil moisture observations stations and the scale mismatch) by training a Machine Learning (ML) downscaling model on physic-based simulations (ParFlow-CLM). The latter, similarly, takes advantage of physics-based simulations (Tethys-Chloris) to train a ML model able to predict soil moisture at any given location provided local meteorology (precipitation and temperature) as well as observed soil moisture at existing stations.

We compare all these products among each other and with in situ observations, both temporally, comparing timeseries at stations’ locations, and spatially, comparing daily values at the different station locations. This comparison allows to assess the quality of the different products and even to identify issues with stations’ observations. We find the upscaling approach compares best to observations, but it also uses them as inputs. Interestingly, when looking at the spatial standard deviation (std) of the different products at stations, the lack of variability of the satellite product (too small std) is improved in the downscaled version. This demonstrates that while the scale mismatch does not allow direct comparison with stations (250x250m² resolution of the downscaled product, a few centimeters for the in-situ measurements), the downscaling is very beneficial, adding higher resolution spatial variability.