Soil moisture droughts have strong impacts on vegetation and agricultural production (Raymond et al., 2019; Tramblay et al., 2020; Vicente-Serrano et al., 2014; Pena-Gallardo et al., 2019). There is a growing interest in simple indicators to monitor drought events at short timescales that could be related to impacts (Li et al., 2020; Noguera et al., 2021). In particular, soil moisture indicators could be more relevant than climatic ones to monitor potential impacts of droughts on agriculture and natural vegetation (Piedallu et al., 2013). Since actual soil moisture measurements remain very scarce, soil moisture simulated from land-surface models is an interesting proxy to develop simplified methodologies that could be applied on data-sparse regions. Land-surface models (LSMs) are valuable tools for a fine-scale monitoring of drought events; however, their implementation requires accurate forcing data and computational resources (Almendra-Martín et al., 2021; Quintana-Seguí et al., 2019; Barella-Ortiz and Quintana-Seguí, 2019). Global implementation also exists but with a coarser resolution and driven by reanalysis data (Rodell et al., 2004; Muñoz Sabater, 2020) that may not be adequate for local-scale applications. Only very few countries have land-surface schemes implemented at the national level to monitor droughts (Habets et al., 2008).

Remote sensing is another option which allows for monitoring soil moisture (Dorigo et al., 2017; Brocca et al., 2019). Microwave sensors allow for the monitoring of surface soil moisture (first 5 cm for L-band-based products and skin for C-band-based products), without the interference of clouds. However, surface soil moisture is not enough for most applications, which require root zone soil moisture, which is the water resource in the soil available to plants. Furthermore, passive L-band products, such as SMOS (Soil Moisture and Ocean Salinity; Martínez-Fernández et al., 2016) or SMAP (Soil Moisture Active Passive; Mishra et al., 2017), have a low resolution, and active C-band products, such as Sentinel-1 (Bauer-Marschallinger et al., 2019), which have a higher resolution, suffer from higher noise and are more sensitive to vegetation. Thus, even though remote sensing is very useful, it still has problems to be surmounted. The resolution of passive L-band products can be increased using optical data (NDVI, normalized difference vegetation index; LST, land-surface temperature), by means of downscaling algorithms (Merlin et al., 2013; Fang et al., 2021), but then the resulting product is sensitive to cloud cover. Also, some progress has been made in deriving root zone soil moisture from surface soil moisture estimations using an exponential filter (Stefan et al., 2021) calibrated using the SURFEX LSM (Surface Externalisée; Masson et al., 2013), but these products are in their early stages and are not operational yet.

Simplified methodologies to estimate and monitor the status of soil moisture are needed in contexts where LSM data are not available and where remote sensing products fall short, such as areas and time periods with dense vegetation or high soil roughness which may affect their accuracy (Escorihuela and Quintana-Seguí, 2016). Different modelling approaches have been proposed, either with conceptual soil moisture accounting models or computational variants of the antecedent precipitation index (Willgoose and Perera, 2001; Javelle et al., 2010; Brocca et al., 2014; Zhao et al., 2019; Li et al., 2020). The general availability of spatial estimates of soil moisture content would help introduce soil moisture into drought monitoring systems, improving their scope and usefulness. Furthermore, this would also facilitate the creation of long-term reanalysis, based on meteorological forcing data, and future climate change studies, without the need for running LSMs. However, to apply this model type at a regional or national scale, there is a need to estimate their parameters over the area of interest. For that purpose, regionalization methods have been employed in hydrology for decades to estimate the parameters of hydrological models in ungauged basins (Blöschl and Sivapalan, 1995; He et al., 2011; Hrachowitz et al., 2013). Several methods exist, based on either catchment similarity or the direct estimation of model parameters using regression techniques with physiographic attributes. For soil moisture modelling, up to now only very few studies have considered these approaches to apply soil moisture accounting models at ungauged locations (Grillakis et al., 2021) or estimate root zone soil moisture using machine learning methods (Carranza et al., 2021).

The goal of the present study is to regionalize a simple soil moisture accounting (SMA) scheme that could be used to monitor soil moisture droughts. The SMA model considered in the present study requires a single parameter, the maximum soil water holding capacity. Two different approaches are compared to estimate this parameter regionally: direct estimation with soil maps or with a machine learning technique, namely random forests.

The study area of this work is the Iberian Peninsula, which is located between the Mediterranean Sea and the Atlantic Ocean and thus is influenced by both synoptic-scale systems, which often come from the Atlantic side, and mesoscale heavy-precipitation events, which often come from the Mediterranean side. The Iberian Peninsula presents a marked relief, with a large and high central plateau and different mountain ranges, which heavily influence the spatial patterns of precipitation, enhancing it windward and decreasing it leeward, generating areas of high precipitation in the west, northwest and north and very dry areas on the central plains and, especially, in the southeast. As a consequence the Iberian Peninsula has a heterogeneous distribution of average annual rainfall, with values ranging from 2000 mm yr-1 to less than 100 mm yr-1. All this has a strong influence on the spatial and temporal variability of soil moisture and soil moisture regimes, having wet regimes in the west and north, where the soil is hardly stressed, and semi-arid areas elsewhere, with a wet (energy-limited) and a dry (water-limited) season, with a dry down that might be interrupted by convective events. All this makes the modelling of soil moisture in Iberia a rather challenging task.

Daily precipitation, temperature and potential evapotranspiration (PET) were retrieved from the SAFRAN-Spain database (Quintana-Seguí et al., 2017). SAFRAN (Durand et al., 1993) is a meteorological reanalysis that produces gridded datasets by combining the outputs of a meteorological model and all available observations using an optimal interpolation algorithm. It has been implemented over France (Quintana-Seguí et al., 2008) and recently over the Iberian Peninsula (Quintana-Seguí et al., 2017) with a 5 km × 5 km spatial resolution. The SAFRAN dataset used in this study includes not only observations from the Spanish part of the Iberian Peninsula but also ingested data from Portugal. The SURFEX LSM (Masson et al., 2013) has been run using SAFRAN-Spain as the meteorological forcing dataset and on the same grid, as was done in Quintana-Seguí et al. (2019). SURFEX uses the ECOCLIMAP2 (Faroux et al., 2013) physiographic database, and it uses the ISBA (Interaction Sol-Biosphère-Atmosphère) scheme (Noilhan and Mahfouf, 1996) for natural surfaces. ISBA has different options; we have used ISBA-DIF, the multi-layer diffusion version (Boone, 2000; Habets et al., 2003). From this simulation, we have extracted the soil moisture of the first 60 cm of the soil by calculating the weighted average of the soil layers that fall within this range. This simulated soil moisture over the Iberian Peninsula is considered herein as the observed reference, in the absence of dense monitoring networks of soil moisture (Martínez-Fernández et al., 2016). From the ECOCLIMAP2 database, elevation and land cover data have also been retrieved and aggregated into the following nine categories: water, bare, ice/snow, urban, forest, grass, dry crops, irrigated crops and wetlands.

We also use the European Soil Database (ESDB) produced by the European Soil Data Centre (Panagos et al., 2012). The ESDB contains information on soil characteristics, including soil depth and texture for topsoil (0–30 cm) and subsoil (30–70 cm) layers at a grid resolution of 1 km. The total available water content (TAWC) is a volumetric parameter describing the water content between the field capacity and permanent wilting point, as a function of the available water content, presence of coarse fragments and depth (Reynolds et al., 2000). In the ESDB, water content at the field capacity and permanent wilting point were determined following the equation from van Genuchten (1980) to estimate the soil water retention curve (Hiederer, 2013). The parameters of the equation are provided by a pedotransfer function (Wösten et al., 1999) for the volumetric soil water content computed from the soil water retention curve. The pedotransfer function uses soil texture, organic carbon content and bulk density to determine the parameters of the soil water retention curve (Hiederer, 2013).

In this study, a simple model allowing for the monitoring of soil moisture conditions was regionalized over the entire Iberian Peninsula, taking as a reference the soil moisture simulated by a high-resolution land-surface model. Two different regionalization methods have been compared, either by the direct estimation of the soil water holding capacity from European soil maps or by random forests, using covariates such as altitude, temperature, precipitation, potential evapotranspiration and land cover. Results have shown that the estimation by random forests is more robust notably to estimate low soil moisture levels. Despite similar average performance between the two methods, the use of soil maps to set the water holding capacity reveals less stable results in some cases, most probably related to the uncertainties in the pedotransfer functions used. While these pedotransfer functions are process-based predictive functions of certain soil properties, random forests are not based on physical processes and are tailored to provide the best estimates in a statistical sense. Therefore, they provide a valuable alternative in contexts where high-resolution soil maps are not available, since they rely on a set of covariates that can be reliably estimated from global databases, such as satellite or reanalysis products (Funk et al., 2015; Hersbach et al., 2020; Muñoz Sabater, 2020).

It should be noted that the results presented herein are highly dependent on the quality of land-surface simulations, in the absence of dense monitoring networks of in situ soil moisture data; thus these results suffer from the same limitations as LSMs, notably, the lack of human process representation in these models (notably irrigation). However, new remote sensing irrigation estimates are being developed (Massari et al., 2021); as a consequence, once the RF model is trained, irrigation estimates could be added to the precipitation forcing data in order to include the human impacts on soil moisture estimations. The results show that this approach could allow for cheaply extending the value of high-resolution LSM simulations to areas where no LSM is implemented (i.e. northern Africa), as long as the climate conditions belong to the range of values used to train the model, mostly in terms of precipitation and potential evapotranspiration ranges. Thus, the model train over the Iberian Peninsula could be applied to other similar areas such as northern Africa, Italy or Greece. As a perspective, other simulations from countries where high-resolution LSM simulations are available, such as France or the USA, could be added to the database in order to expand the coverage over different physiographic and climate contexts (Ma et al., 2021). Consequently, the benefits of LSM simulations of soil moisture could be expanded to other areas, provided those suitable forcing datasets are available. Furthermore, if public meteorological and hydrological organizations were to create soil moisture observation networks, cleverly designed to cover the most relevant climates of their countries, this approach could be used to train the model using these observations and then regionalize the results to the rest of the territory, thus, converting an in situ observation dataset into a gridded dataset with a much greater spatial coverage.