Estimating soil moisture conditions for drought monitoring with random forests and a simple soil moisture accounting scheme

Soil moisture is a key variable for drought monitoring, but soil moisture measurements networks are very scarce. Land-surface models can provide a valuable alternative for simulating soil moisture dynamics, but only a few countries have such modelling schemes implemented for monitoring soil moisture at high spatial resolution. In this study, a soil moisture accounting model (SMA) was regionalized over the Iberian Peninsula, taking as a reference the soil moisture simulated by a high-resolution land-surface model. To estimate the soil water holding capacity, the sole parameter required to run the SMA model, two approaches were compared: the direct estimation from European soil maps using pedotransfer functions or an indirect estimation by a machine learning approach, random forests, using as predictors altitude, temperature, precipitation, potential evapotranspiration and land use. Results showed that the random forest model estimates are more robust, especially for estimating low soil moisture levels. Consequently, the proposed approach can provide an efficient way to simulate daily soil moisture and therefore monitor soil moisture droughts, in contexts where high-resolution soil maps are not available, as it relies on a set of covariates that can be reliably estimated from global databases.