Global hotspots in soil moisture-based drought trends

The SoilClim water balance model (Hlavinka et al 2011, Trnka et al 2020, Řehoř et al 2021) considers four soil layers (0.0–0.1, 0.1–0.4, 0.4–1.0 and 1.0–2.0 m) and calculates soil moisture by comparing the inflow (water infiltration) and outflow (evapotranspiration and deep percolation) water-balance components. SoilClim is applied to each grid and accounts not only for the plant available soil water capacity within the grid but also for the type of vegetation cover, seasonal phenological development, leaf area index (LAI) dynamics, root growth and snow cover accumulation, sublimation, and melting (Trnka et al 2010, 2020, Řehoř et al 2021). The input of all meteorological variables (precipitation, temperature at 2 m, dew-point temperature at 2 m, wind speed at 10 m, and incoming shortwave radiation) consists of ERA5-Land (Muñoz-Sabater et al 2021) data with a 0.1° resolution, aggregated into daily values. Surface runoff is based on the curve number approach (Rosa et al 2012a, 2012b), and evapotranspiration is based on the dual crop coefficient approach (Raes et al 2009, Steduto et al 2009, Rosa et al 2012a, 2012b), using the LAI from ERA5-Land. Canopy height and root depth change as a function of LAI (Řehoř et al 2021). Land-use and terrain inputs were derived from ERA5-Land. Plant available water (AWR) capacity of soil was computed as the difference between field capacity and wilting point based on the inputs from SoilGrids database (Hengl et al 2014, 2017). These inputs included soil texture, organic matter and gravel content from which field capacity and the wilting point were computed using the pedotransfer functions by Saxton and Rawls (2006). SoilClim reproduces well changes in the long-term soil moisture dynamics in the topsoil as well as the trends in the reference evapotranspiration proxy (Trnka et al 2015b). The calculated volumetric soil moisture was converted into relative AWR, in which 100% represents field capacity and 0% represents the wilting point (for more details concerning the global version of the SoilClim model, see Řehoř et al 2021). Daily AWR and its anomalies during the 1981–2021 period were calculated for non-glaciated land in Africa, Asia, Australia and Oceania, Europe, North America and South America, excluding latitudes above 72° N and all of Antarctica. This study used data for the 0.0–1.0 m soil layer, which were obtained by aggregating the upper three of the four modeled layers.

The daily gridded AWR data obtained from the SoilClim model were aggregated into monthly, seasonal and annual means. Linear AWR trends were calculated for each individual grid from seasonal/annual series using the nonparametric Theil–Sen regression method (Theil 1992, Sen 2012), which is robust to outliers. The statistical significance of trends was evaluated using the nonparametric Mann–Kendall test (Mann 1945, Kendall 1975) for p < 0.05.

To define soil drought, 5th and 10th percentile AWR thresholds were calculated for each grid and also each day (employing a 21 day window to smooth the annual variation) for the 1981–2010 reference period to represent severe and extreme drought conditions. The data were further aggregated for individual continents using the percentage of grid points fulfilling the 5th percentile (PD5) or 10th percentile (PD10) conditions every day. These daily percentages were further aggregated into annual and seasonal series. Theil–Sen regression with the Mann–Kendall test was then used to evaluate the trends over all continents.

Pearson correlation coefficients between AWR and temperature or precipitation series were based on monthly data aiming to preserve sufficient variability, and the effective sample size parameter (Hu and Feng 2010) based on AWR data was used to mitigate the effect of autocorrelation in AWR series. Statistical significance of correlations was evaluated at p < 0.05.

To calculate Z scores, an annual cycle was removed from the daily AWR continental averages using the mean annual cycle of AWR during the 1981–2000 reference period smoothed by a 21 day window. To determine whether the statistical distribution of Z scores changed significantly, the Kolmogorov‒Smirnov (K–S) test was applied at p < 0.05.

Because precipitation and potential evapotranspiration (closely related to air temperature) are the main drivers of soil moisture, gridded precipitation and temperatures were correlated with AWR for individual seasons. Correlations of AWR with precipitation totals (figure 6(a)) were positive and generally high and significant, except in extremely arid regions (particularly the Sahara Desert) with nonsignificant correlations. This result is probably connected both to the low variation in the AWR with values close to 0 and to the daily time step of the model, in which occasional rain showers will evaporate in a few hours and not cause a change in the resulting AWR. Moreover, very low correlations occurred in temperate and subarctic zones of the Northern Hemisphere in DJF and in subarctic regions in MAM and SON. The snow cover and frozen ground at these latitudes prevent precipitation from immediately infiltrating into the soil, similar to mountain regions. In the tropics, correlations were slightly lower for the wettest regions because AWR was mostly very close to 100% and therefore relatively unaffected by monthly variations in precipitation. Correlations between AWR and mean temperatures (figure 6(b)) were negative and strong, especially in tropical-wet and monsoon climates, while the situation in arid regions was the same as that for precipitation because actual evaporation was mostly limited by precipitation; thus, temperature variations had only a small effect on AWR. Very high and statistically significant negative correlations appeared in JJA in the temperate climate of the Northern Hemisphere, indicating the importance of temperatures for the origin of JJA droughts and providing an explanation of why the temperate climate zone is quickly getting drier as a result of the rapidly rising loss of moisture to the atmosphere.

Zoom In Zoom Out Reset image size

Because precipitation and temperature from ERA5-Land were crucial for this study, inherent biases of the reanalysis must be considered in interpretation of results. According to Cucchi et al (2020), who used observation- and satellite-based datasets to show biases in ERA5, not significantly different from ERA5-Land (see Muñoz-Sabater et al 2021), ERA5 generally indicates underestimation of temperature and overestimation of precipitation in high elevation above sea level, particularly in Himalayas and Andes. Absence of significant trends for those regions in our results could be influenced by this bias. However, since we focused particularly on long-term changes and used percentile-based drought definitions, constant bias in the reanalysis data should not influence results for most of the regions. This can be demonstrated in equatorial Africa and South America, which have according to Cucchi et al (2020) large variability of both positive and negative precipitation biases in the ERA5, while our results indicate large drying hotspots there without comparable level of variability in AWR trends.

The calculated trends in AWR in our study can be compared with the results of other studies dealing with soil moisture or drought trends. For example, Spinoni et al (2019) identified Amazonia, southern South America, the Mediterranean region, Africa and northeastern China as major drought hotspots. Contrary to Spinoni et al (2019), who identified the Mediterranean region as the main European drought hotspot, this study suggests that drying trends are more severe in eastern Europe. This difference could be an effect of the different period analyzed by Spinoni et al (2019), starting 30 years earlier (1951) than our period and ending earlier in 2016, i.e. not including severe droughts from subsequent years until 2021. Additionally, Song et al (2020), analyzing the 1950–2015 period, identified eastern Europe as an area with a later turning point toward drying, i.e. in approximately 1990, while the Mediterranean region became drier earlier. Another reason for the absence of a more significant AWR decrease in the Mediterranean region may be the most rapid warming during JJA in this region (Mariotti and Dell'aquila 2012), when AWR was generally close to 0%, and further drying becomes physically impossible. However, approaches based on meteorological drought indices instead of soil moisture modeling may indicate an increase in drought while using the same input data because these models do not account for this physical limitation. The existence of significant trends in AWR (in general) is therefore more surprising because our approach does not allow for nearly as much drought accumulation compared to the meteorological drought indices.

Fuentes et al (2022) identified the continental temperate climate zone as severely affected by drought intensification along with the whole South America, but contrary to our study, they specified southwest Africa instead of the central Africa as a major hotspot. Very close to the present study are the results by Rolle et al (2022), who investigated changes in water stress and irrigation requirements in agricultural areas during 1970–2019 and determined a major water stress increases in South America, eastern Europe, eastern Asia and central Africa but a water stress decrease in India. Xu et al (2020), particularly analyzing the tropics based on satellite data for 1981–2019, agreed with our findings of rapid drying, particularly in central Africa and partially in South America. Deng et al (2020) combined both remote sensing and older reanalysis-based data for a similar period (1979–2017) to assess soil moisture trends. Their results mostly agreed with those of the present study, with the locations of drying hotspots in Europe, Asia and Africa. However, they indicated less severe drying in South America and more severe drying in North America, but they did not consider the seasonal aspect of the changes. A comparison of identified trends with projected changes in the near future also holds significance. Chiang et al (2021) proved that recent increase in drought frequency, duration, and intensity has been mainly caused by anthropogenic impacts on the climate. They concluded that both Americas, the Mediterranean, western and southern Africa, and eastern Asia have been affected the most, which partially agrees with our results. Vicente-Serrano et al (2022) studied drought trends based on meteorological drought indices in the last 120 years. They stressed South America as a drying hotspot (this agrees with our results), but their reported trends in many regions differed based on used drought index. They confirmed, that increases in drought frequency and intensity are mostly driven by increased evaporative demand and not by decreased precipitation. Liu et al (2023) used inherent soil moisture products from multiple reanalysis to assess trends in 1980–2020 period. Compared to our results, they found similar drying trends in the European and Asian temperate zones, however they did not report drying hotspot in South America or Midwestern US. The most recent IPCC report (IPCC 2021) indicated continued soil drying in South America, while in central Africa, where we have identified rapid soil drying, the report predicts a significant increase in soil moisture during the 21st century. The report also does not anticipate continued soil drying in central Asia.

As typical for any model outputs, uncertainties associated with the nature of the results cannot be ignored. For example, the module for actual evapotranspiration and soil-water content in the SoilClim model simplifies three-dimensional soil-water transport processes by using only four soil layers. SoilClim estimates of soil-moisture content are also affected by the estimated maximum plant AWR capacity for the soil layers in each grid, by considering only the dominant landcover types, together with uncertainties in ERA5-Land meteorological inputs discussed in section 4.1. Therefore, future continuation of our research could use different water balance model, soil data, meteorological inputs or more detailed landcover specific parametrization, which could involve another reanalysis and also observation-based data. Despite these limitations which may propagate into systematic biases in the simulated water fluxes and state variables, this study focuses mainly on percentile-based droughts and their trends. It is therefore likely that the impacts of these inherent systematic biases on the presented results are minimized. The findings in this study have a great potential for the following studies of soil drought drivers with regard to climate oscillations and large-scale circulation patterns, as well as for analysis potential environmental and socio-economic impacts of soil drying within indicated hotspots.