The study site is located southeast of Lille, Province of Antwerp, northern Belgium (Appendix A – Fig. A2) within the Kleine Nete watershed of the Campine Region. The site elevation is about 10 m above sea level. The land use near the site is grassland and arable land (maize and potatoes are the main crops) and the site has been under grasslands since 1970s. The site has a temperate oceanic climate (Peel et al., 2007). Over the years 1979–2023, annual precipitation and potential evapotranspiration values range from 480–890 and 500–800 mm respectively (Toreti, 2014).

The site's soil is loamy sand with anthropogenic humus A horizon (Bogemans, 2005). A compacted subsoil at 40 to 55–60 cm depth is present, which is an iron oxide pan (Bsm horizon) having green color and red-orange concretions (Appendix A – Fig. A3). This pan is also reported to be common on sandy soils found east of Antwerp, Belgium (Vandamme, 1978), where the study site is situated. Other soil types within the site's vicinity are in Appendix A – Fig. A4.

Eight 2 m square plots were constructed at the site in December 2021. Four of them were de-compacted by excavating the soil down to 150 cm depth using a crane to eliminate the compact layer (Appendix A – Figs. A5, A6). To check the effectiveness of de-compaction, penetrometer logs were also used wherein the lack of peaks in pressure values indicate the absence of compacted layer (Fig. 2a). The remaining four plots still have the compact layer intact (Fig. 2b, Appendix A – Fig. A5).

For each plot, the excavated sandy topsoil layer (40 cm thick for compacted plots and 150 cm thick for non-compacted plots, Fig. 2) was loosened and homogenized well using a cement mixer and then returned to that same plot. In four plots, 480 L woodchips were also mixed, mimicking Campine Region's general agricultural practice (Appendix A – Fig. A5). Impermeable plastic boxes were installed to avoid subsurface lateral in or outflow (Appendix A – Fig. A6). Per plot, two capacitance sensors (HOBO Onset S-SMC-M005) monitored the soil moisture content at 10 and 40 cm depths, allowing to capture the hydrological conditions near the surface and right above the compact subsoil layer, respectively (Fig. 2). This monitoring spanned from 8 July 2022 to 8 September 2024.

Each plot had either maize (Zea mays) or mixtures of Italian ryegrass (Lolium multiflorum) and white clover (Trifolium repens) as vegetation cover (Appendix A – Fig. A5), thus replicating Campine Region's general agricultural landscape. No plots were fertilized.

The grass-clover mixtures were mowed to 15 cm stubble height (mowing timeline in Fig. 1b and Appendix A – Fig. A7) whenever they appeared to grow beyond the plot boxes (Fig. 2). As shown in Eq. (1), the clippings' dry weights were converted into LAI using a representative leaf dry mass of 66 g m⁻², based on the mean of the median leaf dry mass of herbs (60 g m⁻², representing white clover) and of graminoids (72 g m⁻², representing grass) (Poorter et al., 2009) 1LAI=dried weights of grass-clover mixturemean leaf dry mass per unit leaf areaplot area Temporal hourly LAI values during grass-clover's growths (Fig. 1b) were linearly interpolated from the LAI after last mowing and LAI before the next mowing. The LAI after last mowing, pertaining to the grass remaining after mowing, was always assumed to be 0.25, significantly lower than the lowest LAI from clippings (0.84–1.01) (Fig. 1b). This LAI after the last mowing was then added with LAI from the next mowing's weighed clippings to obtain the overall LAI before last mowing.

Maize was sown on 26 April 2022; 11 May 2023; and 22 May 2024 and then harvested on the onset of wet season on 17 October 2022; 20 October 2023; and 8 October 2024. Harvesting was done by cutting them down, leaving 10 cm stem above ground level, whereafter the maize plots were left fallow.

Outside the plots, eight non-compacted sand and seven compact sand samples were obtained at 15 and 50 cm depths, respectively. Their grain size compositions were estimated by laser diffraction technique. Their bulk densities were estimated based on their dry masses and ring kit volumes. Saturated hydraulic conductivity (Ks) was estimated from the falling head method (Table 1).

Precipitation, air temperature (Fig. 1a), solar radiation, relative humidity, and average wind velocity were measured on an hourly resolution using a local weather station located 2 km from the site. Measurement gaps were filled using the data from the closest active meteorological station at Herentals (Flemish Environment Agency, 2021), 5.5 km away from the site. Daily mean temperature ranged from -5.2 °C (in December 2022) to 27.5 °C (in July 2022) during the experiments (Fig. 1a). From these meteorological variables, the potential evapotranspiration (ET₀) [mm h⁻¹] was derived using the Food and Agriculture Organization (FAO) 56 Penman-Monteith Equation for a reference crop (hypothetical grass surface) (Allen et al., 1998), shown in Eq. (2). 2ET0=0.408ΔRn-G+γU2(es-ea)(37T+273)Δ+γ(1+0.34U2)G=0.1Rn,Rn>0[daytime] G=0.5Rn,Rn≤0[nighttime]  where Δ is the slope of the vapor pressure curve [kPa °C⁻¹], Rn is the net radiation at the grass' surface [MJ m⁻² h⁻¹], G is the soil heat flux [MJ m⁻² h⁻¹] whose calculation depends on the sign of Rn that indicates if the hourly Rn was measured at daytime or nighttime, γ is the psychrometric constant [kPa °C⁻¹], U2 is the wind speed measured at 2 m height [cm s⁻¹], es-ea is the vapor pressure deficit [kPa] and T is the average air temperature [°C]. The resulting daily potential evapotranspiration and comparison with precipitation are shown in Fig. 1a.

In 2022, intense summer rainfall events that could generate surface runoff and could be more frequent in the future climate (Hodnebrog et al., 2019; Hosseinzadehtalaei et al., 2020) were mimicked by irrigating the plots. Irrigation is performed by sprinkling using 10 L graduated watering cans. Applied irrigation doses are listed and justified in Appendix B – Tables B2 and B3.

At the end of the monitoring period, the roots inside plots with grass-clover mixtures were sampled. For each of these plots, a root length density profile and a dry biomass profile were obtained. Root-soil cores (8 cm diameter, 15 cm height) were extracted every 15 cm throughout the 0–105 cm depths using root augers. These cores were then washed thoroughly to obtain the roots. The root lengths were estimated using the grid-intersection method and Eq. (3) (Freschet et al., 2021; Newman, 1966; Tennant, 1975). These roots were also oven-dried at 70 °C for a week and then weighed. 3Root length=1114×total number of intercepts(horizontal and vertical)×grid unit length The timeline of these experiments is summarized in Appendix A – Fig. A7.

The experimental plots were represented as 1D soil columns, assuming that lateral flow there was insignificant (Fig. 2). The soil water hydrology was modeled using the numerical model HYDRUS-1D (Šimůnek et al., 2008).

The calibrated and validated models were then used to quantify the water budget under the historical (1972–2000) and wet and dry future climate scenarios (2072–2100). Constant LAI based on the temporal means of measured LAI in the experimental setups were used (1.57 for non-compacted and 1.32 for compacted case models) (Fig. 1b). For Vegetation Parameterization 1, the non-compacted case model's LAI was adopted for the compacted case model (Fig. 3a, b). These ensure that the seasonal patterns of hydrological variables are solely brought by the climate forcings. The initial soil moisture profiles were also set at field capacity.

The future climate projection data are in the form of a 30 run ensemble from regional climate models, used to force the soil water flow models. All these runs, set at the Royal Meteorological Institute at Uccle, Brussels, Belgium (Appendix A – Fig. A2), can be considered representative for inland Flanders (Van Schaeybroeck et al., 2021). They were generated based on the EURO-CORDEX regional climate models under the combined four (RCP 2.6, 4.5, 6.0, 8.5) for the future period 2071–2100. Equal weights were given to these four RCP scenarios. The regional climate model outputs were statistically downscaled via the quantile perturbation method of Willems and Vrac (2011) based on the historical daily meteorological observations in Uccle (1 January 1971 to 31 December 2000).

For the models' spin-up, the results from January–November 1971 were discarded under each run. This is to omit undesired deep percolation recorded a few days after the start of model simulations. The year 1972 was thus defined as the calendar months of December 1971 to November 1972 and so on. Thus, each run spans from the years 1972 to 2000, corresponding to the calendar months of December 1971 to November 2000. With this, equal number of data sets was ensured across seasons in the scenario analyses: 29 data sets each under winter (DJF: December–January–February), spring (MAM: March–April–May), summer (JJA: June–July–August), and autumn (SON: September–October–November). This scheme also applies for the future scenarios.

To represent a wet and a dry future scenario, this study used the 95 percentiles of the mean annual hydric excess (precipitation minus potential evapotranspiration or P - ET₀) and deficit (potential evapotranspiration minus precipitation or ET₀ - P), respectively, from the full ensemble of 30 model runs (Table 3) (Christensen and Christensen, 2003; Van Schaeybroeck et al., 2021). These 95 percentile based scenarios can be interpreted as “high-impact” scenarios, where the future is expected to lie with high likelihood between the current climate and that high-impact scenario.

The whole modeling workflow from setup to scenario analyses is summarized in Fig. 4.

Overall, the compacted setup has higher soil moisture than the non-compacted setup, as showed by both observation data and simulations (Fig. 5). At 40 cm depth, these differences are even larger. Moreover, very high soil moisture peaks are more flattened in the compacted setup, reflecting the compacted setup's frequent saturation moments (Fig. 5d). Despite the irrigation at the experiment's beginning, the non-compacted setup was drier than the compact soil. All these reflect the compact layer's role in both experimental plots and simulations.

At 10 cm depth, the models simulate the soil moisture dynamics well. However, at 40 cm depth, simulations underestimate the soil moisture (Fig. 5b, d). Moreover, the compacted case models simulate a faster decline in moisture content in mid-January 2023 to start of March 2023 (Fig. 5d). This might indicate that the compacted layer's estimated hydraulic conductivity from optimizations is still relatively high. Nearer to the soil surface (Fig. 5a, c), models involving different root depths are closer to observations than the ones with same root depths.

At 10 cm depth (Fig. 5a), the deep rooted non-compacted case model improved the moisture content simulations in mid-May 2023 to mid-July compared to its shallow rooted counterpart. With shallow roots, the model overestimates the root water uptake, leading to its simulations' faster decline than both the observations and deeper root simulations. At 40 cm depth (Fig. 5b), the shallow rooted non-compacted case model underestimated the observed decline in moisture content in June–July 2023. Meanwhile, for the compacted case models (Fig. 5c, d), the differences in LAI yielded very little differences in the simulated moisture contents.

Compared to the historical scenario, the wet scenario depicts slightly higher annual precipitation yet similar annual potential evapotranspiration (Fig. 6a, c). Meanwhile, the dry scenario has a similar annual precipitation yet higher summer potential evapotranspiration. Both future scenarios also show stronger precipitation seasonality (lower in summers and higher in winters) than the historical scenario (Fig. 6b, d).

For the hydrological components (actual evapotranspiration, actual evaporation, actual transpiration, soil moisture storage above the compact layer (i.e., 0–40 cm), deep percolation), the actual fluxes are reported in Figs. 7 and 8. Their results were reported in terms of the annual and monthly differences between the compacted and non-compacted case models in Figs. 9 and 10.

With compaction, water storages in the upper soil profile increase during winter for all vegetation parameterizations and climate scenarios (Fig. 10a, b). However, in the dry climate scenario's summers, both the compacted and non-compacted soils' upper layers dry out to similarly low moisture contents. This is most noticeable for vegetation parameterization 3 with deeper roots and higher LAI in the non-compacted model than in the compacted soil model (Fig. 10c).

For vegetation parameterizations 1 and 2 that assumes same shallow root depths and root density profiles (Fig. 9a, b), actual evapotranspiration and evaporation increase, and deep percolation decreases with compaction. The lower LAI for the compacted case model in parameterization 2 leads to a lower transpiration than the non-compacted case. This is almost fully compensated by higher evaporation in the compacted case model. Thus, vegetation parameterizations 1 and 2 do not exhibit much difference on the simulated impact of compaction on deep percolation.

Under vegetation parameterization 3, the non-compacted case model's deeper rooting leads to more water uptake from the deeper subsoil. This further increases the difference in transpiration between the non-compacted and compacted case models, especially under the dry future scenario (Fig. 9c). This deeper rooting also reduced the uptake from the non-compacted case model's upper depths compared to the shallow rooting from parameterization 2 (Fig. 9b). Thus, more water remained available for evaporation in the non-compacted soil, leading to smaller differences in evaporation losses between the compacted and non-compacted soils (Fig. 9c), than in parameterization 2 (Fig. 9b). Higher transpiration and evaporation from the non-compacted case model also led to its simulated less deep percolation than in parameterization 2. This is shown by parameterization 3's smaller difference in simulated deep percolation between the compacted and non-compacted case models (Fig. 9c). For the historical and wet climate scenarios, more deep percolation is simulated in the non-compacted soil, but the opposite occurs for the dry climate scenario. Thus, whether deep percolation increases or decreases with compaction or after de-compaction depends on how vegetation was parameterized and can be opposite in dry or wet future climate scenarios (Fig. 9c).

In terms of seasonal dynamics, when root depth does not change with compaction (Fig. 10a, b), deep percolation is reduced and delayed for the non-compacted setup after the summer season (especially in November and December). However, if root depth is also adjusted after compaction (Fig. 10c), deep percolation is reduced in the compacted setup in March and April. During this period, the compacted setup has water still draining from the subsoil while root water uptake from deeper layer in the non-compacted setup already starts, reducing deep percolation.

All these key findings of water balance changes due to compaction under various vegetation parameterizations are summarized in Fig. 11.

Results from the scenario analysis show the effect of both soil compaction and its interaction with the vegetation on the soil water balance (Fig. 11). As per parameterizations 1 and 2 (Figs. 9a, b, 10a, b), a trade-off seemingly exists between soil water retention for vegetation (by leaving the subsoil compacted, based on high soil moisture for the compacted case model) and maximizing groundwater recharge potential (by de-compaction, based on high deep percolation for the non-compacted case model). Considering that compaction reduces vegetation biomass production (i.e., parameterization 2) (Figs. 9b, 10b), de-compaction could promote both groundwater recharge and vegetation productivity. Finally, if the vegetation is allowed to develop a deeper root system in the de-compacted case model (i.e., parameterization 3), de-compaction will not always guarantee higher recharge (Fig. 9c). These abovementioned inferences drawn from these three vegetation parameterizations provide very different and even contrasting overviews of the agricultural water availability's dynamics and sustainability. From these could arise conflicting water resource management strategies tailored to those inferences in this climate change context.

These insights show that while sandy subsoil compaction directly affects both vegetation growth and water balance, the affected vegetation growth also further influences the water balance. Therefore, in hydrological studies involving (de-)compaction, vegetation growth above- and belowground should be dynamically incorporated. With this, field evidence of vegetation growth, root growth and yield, often far lacking in compaction studies, is crucial.

Climate change would likely negatively impact crop production as it could promote more and longer dry (droughts) (Cotrina Cabello et al., 2023; Li et al., 2009; Shanker et al., 2014) and even very wet periods (waterlogging) (Fischer et al., 2023; Tian et al., 2021; Yang et al., 2024). These crop's long-term responses under these contrasting extreme events could also be considered in the hydrological model simulations of future scenarios. This could refine the current understanding of future agricultural water balance even further, including evaporation losses and groundwater recharge.

Compared to observations at 40 cm depth, the non-compacted case's models severely underestimate the observed soil moisture during wet periods – in two weeks of January 2023 during calibration period (narrow peak) and mid-November to end of December 2023 and mid-February to March 2024 during validation period (wider peak) (Fig. 5b, Appendix A – Fig. A9). The water table could have risen temporarily high enough to saturate these loosened layers at 40 cm depth. The wider soil moisture peak during 2023–2024 is likely due to much higher precipitation (528 mm from 1 October to 31 March) than 2022–2023 (359 mm from 1 October to 31 March) in the site. Unfortunately, the 200 cm deep water table observed from soil profiles during March 2023 and 2024 could have occurred only when the water table had already receded. In flat, low-lying areas, such shallow water table can influence soil hydrological fluxes (Groh et al., 2016) and even vegetation growth (Glanville et al., 2023; Horsnell et al., 2009; Odili et al., 2023; Ridolfi et al., 2006). With this, the feedback loop involving soil compaction, vegetation, and soil water becomes more complex, which, however, is beyond this study's scope.

This study did not also indicate the runoff amounts in the main results. This is because under all scenarios, runoff only occurs for one day just on all compacted case models (i.e., 31 mm on 29 August 1996; 74 and 96 mm in wet and dry August 2096, respectively) whenever an extreme rainfall event occurs under very high antecedent soil moisture conditions. This means that runoff appears to be infrequently simulated, despite reported occurrences of waterlogging in even sandy subsoil compaction (Huang and Hartemink, 2020; Polge De Combret-Champart et al., 2013). This might be due to the difficulty in simulating infiltration-excess runoff as the projected meteorological data have coarse time resolution (i.e., daily), and thus rainfall intensity is not considered (Mertens et al., 2002). This then leads to overestimated infiltration and underestimated runoff in the scenario analyses (Šimůnek and Weihermüller, 2018). Meanwhile, the hourly simulations during the experiments generated runoff for the compacted setups (Appendix B – Table B6), linked to intense irrigation events (40 mm runoff brought by 100 mm irrigation in three hours of 7 October 2022, 27 mm runoff brought by 160 mm irrigation of which 50–60 mm is delivered for six hours every day of 13–15 September 2023; Appendix B – Table B3). Nevertheless, the compaction's role on runoff generation is still clear as only the compacted case models generated runoff in both the experiments and scenario analyses.