Belgium is located in Western Europe covering an area of 30 528 km², varying in topography from sea level along the North Sea coast to 700 m in the Ardennes-Eifel massif in the south eastern parts (Fig. ) . The country experiences a warm temperate maritime climate (Köppen-Geiger Cfb) strongly modulated by the warming effect of the North Atlantic Drift . Data from the Royal Meteorological Institute of Belgium (RMI) shows that mean annual temperature ranges between 13 and 17 °C, varying spatially with elevation and distance inland. Winters are generally mild, with December–January lows dipping under 5 °C but rarely below freezing conditions for long periods. Winters are colder in the Ardennes region due to a weaker maritime influence and higher elevation. Summers are moderately warm with July highs peaking around 18 °C although extremes above 30 °C have occurred in recent years. The country receives an annual average precipitation of about 800 mm which varies between 700 mm in the western low lying regions, up to 1400 mm in the Ardennes where precipitation is enhanced by orographic effects . Temporally, rainfall is fairly evenly distributed throughout the year (Fig. ), with seasonal patterns dominated by summer convective storms and winter frontal systems .

Land cover in the country is predominantly agricultural (44 %), dominated by croplands and animal husbandry. Cultivated areas dominate the central loamy belt and the northwest of the country while the coastal polders typified by heavy soils, are more suited for animal-based farming . Forests cover about 23 % of the territory (just over 700 000 ha) with 79.8 % in the Walloon region, 19.9 % in Flanders and 0.3 % in the Brussels-Capital . Most of the lowland forests are dominated by broad-leaved tree species with clusters of coniferous forest plantations in the northeast. In the Ardennes, forests form a mixed broadleaved–coniferous complex in the foothills, gradually transitioning to conifer-dominated stands at higher elevations . Built-up and urbanized areas account for about 20 % of the land with most cities dating back to the Middle Ages. The average population density of the country is 385 inhabitants per km² .

In our study, we used the mesoscale Hydrologic Model mHM; (version v-5.13.2-dev0) to simulate domain-wide root-zone (0–2 m) soil moisture conditions and streamflow, which we used as an additional hydrologic constraint for validating basin-scale hydrology at major outlets.

mHM is a spatially distributed hydrological model based on numerical representations of dominant hydrological processes. The model is driven by hourly to daily meteorological forcings, which include precipitation, air temperature (henceforth simply temperature), and potential evapotranspiration, and accounts for major hydrological processes like snowmelt and accumulation, canopy storage, evapotranspiration, surface runoff and flood routing, three-layer soil moisture content, and subsurface storage. To represent spatial variability of inputs and state variables, the model uses three different spatial resolutions, namely (in order of fine to coarse resolution) Level-0 (L0: small-scale morphology) to represent the main terrain features, geological features, land cover, and soil properties; Level-1 (L1: mesoscale hydrology) to represent the dominant hydrological processes; and Level-2 (L2: large-scale meteorology) to describe the variability of meteorological forcings. The model harmonizes the data internally using the multiscale parameter regionalization MPR;. MPR links model parameters at L1 to their corresponding ones at L0 using non-linear transfer functions that couple catchment characteristics with global (calibration) parameters to regionalize model hydrologic parameters at L0 and link them to their corresponding values at L1 using upscaling operators such as arithmetic mean, geometric mean, and harmonic mean MPR;. With this technique, mHM achieves quasi scale-invariant parameters that enable the model to preserve the spatial variability of state variables and conserve mass balance . mHM has been successfully used in multiple studies at scales ranging from river basins , country level up to continental-scale and global studies .

To characterize soil moisture droughts, we use a monthly soil moisture index (SMI), following , considering the total soil water content of the root zone up to a depth of 0.5 m (we limit our analysis to this depth since groundwater in some regions is shallower than 0.5 m). For each month, grid cell soil moisture is expressed as a percentile relative to that month's historical soil moisture and scaled to a range between 0 and 1.

The computation of SMI in this study is based on the methodology of , which proceeds as follows. Firstly, the monthly soil moisture averaged over the root-zone depth (0.5 m for this study) is extracted and used to compute a probability distribution function (PDF) ft(x) for each grid cell as; 7ft(x)=1nh∑k=1nKx-xkh Where, x is the soil moisture value at which the PDF is evaluated x1, …, xk represent the simulated monthly soil moisture values for the month t over the simulation period. n is the number of calendar months within a given period (e.g., 50 for a 50-year period. Note that this conversion is done for each calendar month separately to account for inherent seasonality in SM simulations. K is a Gaussian kernel function and h is the bandwidth that controls the smoothness of the kernel (Eq. ). The optimal value of h is computed using a cross-validation criterion. 8K(x,xk)=12πh2exp⁡(x-xk)22h2 The monthly grid cell SMI is then derived by integrating ft(x) and the resulting SMI values are classified into percentiles. Drought-affected grid cells are identified using a threshold percentile τ, which is commonly set at 0.2 (e.g., ). This means that for a given month, a grid cell is experiencing drought if the soil moisture value falls below the 20th percentile of values for that month. According to , this percentile represents the threshold at which the magnitude of drought begins to damage crops, cause water shortages and present high risks of fire. Next, adjacent cells where SMI ≤ τ (henceforth denoted as SMIτ) at each timestep are consolidated to form drought clusters, which are defined by a minimum threshold area. Spatial clusters which share a minimum overlapping area at consecutive time steps are then joined to form multi-temporal clusters, each with a unique identity. For each cluster, the mean duration (months), areal extent from the onset to termination, and the total drought magnitude, which is the spatiotemporal integral of SMIτ over the area affected, are computed. Following , the magnitude of each event is computed as the space-time integral of the drought duration in months over the area under drought. This is represented mathematically as; 9TDM=∑t=t0t1∫At[τ-SMIi(t)]+ t0 and t1 represent the onset and termination month of a multi-temporal drought event, At is the area under drought at timestep t expressed as a percent of the total domain area, and + means the magnitude is computed only for the positive part of the function. To avoid detecting small, isolated and short-lived dry spells as droughts, we specified a minimum threshold area of 640 km² (about 2 % of total domain area), based on for an event to be considered as a drought and an overlap area of the same size for two drought events at successive time steps to be considered as a single multi-temporal drought cluster.

The daily standardized anomalies of mHM-simulated soil moisture evaluated against in situ observations from the ISMN are shown in Fig. . Of the 48 stations where in situ data was retrieved, 21 sites passed quality control checks and were retained for validating the model outputs. The resulting comparison showed that the two datasets are highly temporally correlated, with a mean Pearson r‾ = 0.86 (back-transformed averages from the Fisher z-scale), although the strength of the correlation varied with sensor depth and type. The correlation is lowest for the top 50 mm of the soil profile (r‾ = 0.81 for all networks) and increases to 0.86 for the profile depths greater than 150 mm.

Even for the selected in situ sites, some still exhibited spurious spikes outside of random noise (shown by the red scatter points in Fig. ). We chose not to discard these points so as to preserve an adequate number of validation stations and to highlight the practical difficulty of obtaining perfectly reliable reference soil moisture data for validating model outputs. Despite such outliers, the model simulations and ISMN observation showed similar temporal variability in soil wetness and dryness. The difference mainly occurred in the top 50 mm layer during very dry episodes when mHM produced more extreme negative anomalies than most sensors (Fig. a–d). This explains why the correlation between the datasets is the lowest at this depth. We attribute this divergence partly to a flooring effect of capacitive sensors, which tend to plateau at very low volumetric water contents, whereas the model continues to resolve further drying. For deeper layers, the intensity and duration of dryness were more consistent between both datasets.

To evaluate how well the model simulates drought conditions, we investigated the drought-day detection skill (when the observed standardized anomaly fell below its 20th percentile) by counting hits (H; days when both model and observations indicate drought), misses (M; observed drought days not flagged by the model), false alarms (F; days flagged as drought by the model but not by the observations), and correct negatives (C; days when both indicate non-drought). (The methodology is described in more detail in Sect. S1 in the Supplement). From this analysis, we found that the model shows high skill in reproducing observed drought conditions, as it was able to detect 74 % of observed drought days from the 21 stations. The false alarm rate was also only 5 %, while the mean F1 score (which summarizes the balance between misses and false alarms as 2H/(2H+F+M)) was 75 %. We attribute the differences in detecting droughts to the scale mismatch between mHM soil moisture, which represents average conditions over a grid cell, and the highly localized nature of point in situ measurements. Nevertheless, these metrics indicate that the model can be applied to study droughts.

Regarding streamflow performance, the model shows good and spatially consistent skill across the entire modelling domain and thus provides a reliable basis for analysing soil moisture dynamics. We evaluated daily discharge at 168 gauging stations. During calibration, the mean Nash-Sutcliffe Efficiency (NSE) across stations was 0.62, with 80 % of stations achieving NSE ≥ 0.5 (a commonly used benchmark for satisfactory streamflow simulation). Model performance during the validation period was also comparatively good, with a mean NSE of 0.63 and 83 % of stations recording NSE ≥ 0.5. The full details of the streamflow evaluation, including the NSE definition, are provided in Sect. S2 in the Supplement (Fig. S1).

To summarize how soil-moisture drought behaviour evolves across decades, we use three complementary metrics. First, we quantify the magnitude of each event using the Total Drought Magnitude (TDM), which integrates drought severity over space and time and thus allows drought events to be ranked consistently (Sect. ). Second, in Sect. , we describe how drought severity is distributed by quantifying the fraction of drought-affected area falling into different severity classes (moderate, severe, extreme, and exceptional) within each decade. These classes capture shifts in the composition of drought conditions beyond just the total magnitude. Third, in Sect. , we quantify cumulative drought exposure as the total number of months in which each grid cell experiences drought per decade (months need not be consecutive). This metric summarizes how frequently drought conditions recur at a given location over a decade. For decadal summaries, we defined decades starting from 1971 (i.e., 1971–1980, 1981–1990, etc.) since SPEI construction requires accumulated water-balance anomalies over preceding months (January 1970 will thus not have SPEI-1 values, while January–March 1970 lacks SPEI-3 values. The first year with complete SPEI values is 1971).

To examine how precipitation-based drought indicators reflect land-surface moisture stress, we compared SMI and SPEI patterns during the three most severe soil moisture drought events ranked by total drought months (TDM): 1975–1977, 2016–2017, and 2018–2019. Because SMI is computed on a monthly timescale, we derived the climatic water balance (precipitation minus potential evapotranspiration) from E-OBS and calculated SPEI at one- and three-month accumulation periods. Pixel-wise SPEI-1 and SPEI-3 were computed using the SPEI package developed by . We limited the accumulation period to three months because this timescale is currently used in operational drought monitoring in Belgium. Since SPEI is anomaly-based rather than percentile-based, we associated SPEI =-1.0 with SMI =0.2 to represent the threshold for at least moderate drought, following the drought severity guidelines of . We evaluated the differences between indices in terms of (i) anomaly magnitude, (ii) drought persistence (maximum number of consecutive months under at least moderate drought within an event), and (iii) cumulative drought exposure (total number of months, not necessarily consecutive, under at least moderate drought within the event window).

In terms of anomaly magnitude, SMI generally indicated stronger and longer-lasting deficits than SPEI-1 and, to a lesser extent, SPEI-3 (Fig. ). SPEI-1 responds strongly to short-lived precipitation anomalies that may not immediately translate into changes in root-zone storage. By design, SPEI-3 smooths some of the short-term variability in SPEI-1 and more closely resembles the temporal evolution of soil-moisture anomalies, but still tends to underestimate deficit magnitude relative to SMI over our domain (Fig. ). Among the three events, SMI indicated the strongest soil moisture deficits during 2016–2017 (with SMI approaching zero), which is not reflected in either SPEI-1 or SPEI-3. Although the 2016–2017 drought was interrupted by intermediate wet conditions during March and April 2017, leading to partial recovery, this wet spell did not split the event because the month-to-month overlap in drought area remained above the 640 km² merging threshold, and the drought therefore remained a single multi-temporal event.

We also found that SMI-based droughts exhibited higher persistence than SPEI-based droughts. Median persistence for SMI was 9 months in 1975–1977, 6 months in 2016–2017, and 7 months in 2018–2019 (Table ). In comparison, SPEI-1 shows much shorter median persistence (3 months in 1975–1977 and 2 months in both 2016–2017 and 2018–2019), while SPEI-3 is closer to SMI but remains lower (7 months in 1975–1977 and 5 months in both 2016–2017 and 2018–2019).

The same pattern is evident for cumulative drought exposure. Median cumulative exposure for SMI was 10 months in 1975–1977 and 2016–2017 and 8 months in 2018–2019, compared with 4, 6, and 3 months for SPEI-1 and 7, 8, and 6 months for SPEI-3 (Table , additional maps in Sect. S3 in the Supplement, Figs. S3 and S4).

This systematically longer persistence and exposure shown by the SMI indicates that soil moisture droughts last longer than precipitation-based droughts because transient precipitation events do not necessarily translate into root-zone recovery. Additional description on the patterns of SMI and SPEI recovery is presented in Sect. S3 in the Supplement.

The differences presented herein do not imply that one indicator is necessarily better; rather, they are all useful for demonstrating how a drought shock progressively propagates through different components of the hydrological system. Precipitation-based indices like SPEI reflect short-term meteorological inputs that may still be agriculturally meaningful. As Fig. shows, rainfall events during dry periods may not replenish deeper soil moisture due to immediate losses through evapotranspiration, yet these events can still temporarily alleviate plant water stress, especially for fast-responding, shallow-rooted crops or annual crops. The recovery of SPEI out of drought conditions may thus signal “relief” that is real, albeit short-lived and limited in scope. On the other hand, SMI-based drought analysis better captures the persistence of land surface water deficits and the residual moisture stresses that continue to affect the dependent ecosystems (e.g., perennial deep-rooted vegetation) long after meteorological conditions have normalized.