The research concentrated on two sites within the ICOS (Integrated Carbon Observation System) monitoring network, which are also part of the IPCC Network (Intergovernmental Panel on Climate Change), selected for contrasting soil characteristics and their well-established infrastructure and suitability for comprehensive data collection (Table 1).

Water balance models or Soil–vegetation–atmosphere transfer (SVAT) models are valuable tools for estimating detailed atmosphere–plant–soil water exchange as well as waterfluxes within terrestrial ecosystems using a 1D simplification representing of evapotranspiration and vertical soil water movement processes (Olioso et al., 2005). LWF-BROOK90.jl, a process-based, one-dimensional SVAT model, developed to simulate daily vertical water fluxes within soil–plant–atmosphere systems at the point scale, was applied in this study. In addition to simulating water fluxes and storage dynamics, LWF-BROOK90.jl also includes an isotope module for modelling stable water isotopes; however, isotope simulations were beyond the scope of the present study. Owing to its vertical structure, the model is particularly suited for forest hydrology studies and ecohydrological process analysis, rather than fully distributed catchment-scale simulations. LWF-BROOK90.jl represents a complete reimplementation of the LWF-BROOK90 model in the Julia programming language (Bernhard et al., 2020). It builds upon the R package LWF-BROOK90R, its underlying Fortran source code and the original BROOK90 (v4.8) implementation (Federer et al., 2003; Schmidt-Walter et al., 2020), while preserving the conceptual structure and process formulations of its predecessors. Water movement is represented as a vertically coupled soil–plant–atmosphere continuum consisting of a single-layer (“big-leaf”) canopy, a snowpack when present, a multi-layer soil column with user-defined thickness and hydraulic properties and a conceptual groundwater store. The modeled storage compartments comprise interception (separately for liquid precipitation and snow), snowpack, vertically discretized soil water, plant water storage and groundwater. A more detailed model description including the underlying processes and fluxes is provided in the Supplement, Section Model description.

Initially, the visual examination of soil moisture time series across the 11 soil profiles each, measured at four depths (10, 30, 50, and 80 cm) at the Kienhorst (Fig. 4) and Tharandt (Fig. 5) plot, revealed distinct patterns of variability. Seasonal effects were evident at both study locations across nearly all soil depths due to the fluctuating intensity and timing of precipitation events and the changing consumption of water by vegetation.

However, the redistribution and consumption of precipitation water on its way through the unsaturated soil zone resulted in a diminished manifestation of seasonal patterns in the lower soil horizons. Notably, certain soil profiles at the Tharandt plot exhibited an almost uniform low moisture level throughout the year in the lower horizons, indicating limited seasonal variability in these depths. Additionally, rapid and pronounced increases in soil moisture following heavy rainfall events were observed in several, but not all, soil profiles at the Tharandt plot due to waterlogging following heavy rainfall events caused by poorly permeable soil layers. These conditions are consistent with field observations, where it was possible to observe this phenomenon. At the Kienhorst plot, the range of observed soil moisture values and the variability between individual profiles were consistently low both in dry and wet periods with evident uniformity of soil moisture across all depths. Soil moisture exhibited a consistent pattern across all depths, with a range of approximately 15 vol % between minimum and maximum values. Conversely, the Tharandt plot was characterized by significant fluctuations in soil moisture content, coupled with considerable variability between individual profiles, particularly during and following heavy precipitation events occurring here more often. At this plot, the range between minimum and maximum moisture content reached nearly 50 vol %, highlighting the pronounced heterogeneity of soil water dynamics.

The daily groundwater recharge time series revealed pronounced fluctuations throughout the study periods, with seasonal and vegetation-period effects distinctly evident (Fig. 11). At both the Kienhorst and Tharandt plots, groundwater recharge predominantly occurs outside the vegetation period, when the consumption of precipitation water by processes such as evaporation, root water uptake and transpiration is reduced to a minimum. While groundwater recharge at the Tharandt site remains predominantly winter-dominated, at the Kienhorst site it shows a tendency toward an earlier onset and a partial extension into the vegetation period. Conversely, in the vegetation period between may and October, these processes utilize nearly all available precipitation water, effectively limiting groundwater recharge during the vegetation period. Despite these shared temporal patterns, the magnitude, variability and temporal distribution of recharge events differed substantially between the two contrasting soil environments.

At Kienhorst plot, characterized by a homogeneous sandy soil, groundwater rechargeshowed smooth, continuous and regulated patterns with only a few distinct spikes following after major precipitation events. The high hydraulic conductivity and low water-holding capacity of the sand, facilitated rapid infiltration and percolation. Consequently, recharge responded almost immediately to rainfall inputs, resulting in uniform recharge pulses. The results displayed relatively small differences in the magnitude and timing of recharge events between the 11 soil profiles, reflecting the overall homogeneity of the site. Variability among profiles was primarily expressed in the height of individual peaks, while the temporal structure remained largely consistent across simulations.

In contrast, at Tharandt plot, underlain by a heterogeneous loamy soil with a substantial stone content, groundwater recharge occurred in irregular, pulse-like events rather than in a continuously attenuated manner. Peak magnitudes were comparable to those observed at the Kienhorst site; however, recharge was highly variable across soil profiles, with several profiles exhibiting little to no recharge over the simulated period. This indicates a strongly heterogeneous hydrological response, likely reflecting spatial variability in hydraulic properties and preferential flow pathways. The reduced hydraulic conductivity and higher field capacity of the loamy matrix delayed percolation, causing water to be retained in the upper soil layers before contributing to recharge. Sensitivity to precipitation events was also lower, recharge increases occurred only after sustained wet periods or when soil moisture approached saturation. Moreover, heterogeneity of soil matrix including bypass flow contributed substantially to the variability among the 11 soil profiles, which is much more pronounced at the Tharandt plot than at the Kienhorst plot.

Considering the range of the 30 simulated time series of groundwater recharge per profile, reflecting model uncertainty, it becomes evident that the largest discrepancies between the individual simulations occur predominantly during periods outside the growing season and following heavy precipitation events (Fig. 12).

The uncertainty band is particularly wide during periods of high recharge, indicating that peak events are more sensitive to parameter choices than low-flow periods. During periods of low recharge, the uncertainty band narrows considerably, indicating that the model is more constrained under dry conditions. Moreover, the spread between the simulations is slightly more pronounced at the Tharandt plot compared to the Kienhorst plot. This indicates that groundwater recharge at Tharandt site is more sensitive to parameter variability and structural assumptions within the model.

An analysis of cumulative groundwater recharge highlights substantial spatial variability across the 11 soil profiles at both plots separated for the years 2023 and 2024 (Fig. 13). At the Kienhorst plot, annual recharge among the 11 profiles ranged from 35 to 72 in 2023 and from 125 to 175 mm in 2024. In contrast, the Tharandt plot showed a wider range of recharge rates, with values between 0 and 145 in 2023 and 0 to 142 mm in 2024, while exhibiting a higher degree of profile-to-profile variability with several profiles showing little to no groundwater recharge over the simulated period. These differences primarily reflect the contrasting soil hydraulic environments: the sandy, homogeneous substrate at the Kienhorst plot promotes efficient percolation and thus higher recharge amounts, whereas the stone-rich, loamy soil at the Tharandt plot restricts percolation and reduces overall recharge.

At Kienhorst site, the groundwater recharge calculated using the model parameterisation based on the mean value of the 11 soil moisture time series lies at the edge of the ensemble in 2023 (72 mm), while it is closed to the centre of the ensemble in 2024 (161 mm), suggesting that the representativeness of the spatially averaged soil moisture signal varies interannually, ranging from a marginal to a central position within the ensemble distribution at this relatively homogeneous site. At the Tharandt plot, the mean-based values fall within the center of the ensemble (74 in 2023 and 69 mm in 2024). However, this apparent agreement masks a pronounced bimodal response among the individual soil profiles: while approximately half of the profiles exhibited little to no groundwater recharge, the remaining profiles generated substantially higher recharge. Consequently, the ensemble mean provides a representative estimate of the average annual recharge but does not capture the marked spatial heterogeneity arising from preferential flow and spatially variable hydraulic properties.

To further analyse the recharge dynamics and to check the plausibility of calculated values, annual groundwater recharge was expressed as a percentage of annual precipitation. At Kienhorst plot, recharge amounted to 5 % to 10 % of precipitation in 2023 and 20 % to 29 % in 2024. At Tharandt plot, the corresponding values were 0 % to 17 % in 2023 and 0 % to 19 % in 2024. These percentages highlight the strong influence of soil hydraulic properties on the partitioning of precipitation: despite receiving substantially more rainfall, Tharandt plot exhibits to some extent lower recharge efficiency due to higher water retention, the buffering effects of the loamy matrix and preferential pathways promoting localized bypass flow. In contrast, the sandy soil at Kienhorst plot promotes rapid drainage and limited water storage, resulting in a much larger fraction of precipitation contributing to groundwater recharge.

To assess the plausibility of the simulated groundwater recharge and to ensure that the simulated amounts are physically realistic in relation to the available water input and the site-specific evaporative demand, full annual water balance evaluation for the years 2023 and 2024 was carried out based on the mean model parameterisation of the individual 11 profiles (Fig. 14, Tables S13 and S14 in the Supplement).

The results of the water balance indicate that the dominant processes consuming precipitation input differ between the two sites, reflecting both their contrasting soil hydraulic properties and the structure of the model. At the Kienhorst plot, precipitation is primarily partitioned into interception, transpiration and substantial vertical drainage to groundwater, consistent with the rapid percolation and limited storage capacity of the sandy substrate. At the Tharandt plot, in contrast, precipitation is consumed mainly by interception, transpiration and bypass flow, with groundwater recharge remaining comparatively small due to the strong retention and reduced matrix conductivity of the loamy, stone-rich soil as well as the stone induced preferential pathways. The negative change of soil water storage at both plots in 2024 indicate a net depletion of soil water storage, which is plausible given the dry conditions in 2024 and the stronger evapotranspirative demand during the growing season. Overall, the balance closes well and the relative magnitudes of the fluxes align with the expected soil hydraulic behaviour at both sites.

The results of the performed bootstrapping Monte Carlo–based subsampling procedure show that groundwater recharge estimates vary substantially when only a small number of soil profiles is used and they become statistically stable only once a sufficient sample size is reached (Fig. 15). A comparison of the two study sites highlights slight pronounced differences in the sampling effort required to achieve the level of statistical reliability. At Kienhorst plot, only three soil profiles are sufficient for 95 % of all subset means to remain within ±10 % of the overall mean, indicating a relatively homogeneous system with low spatial variability. In contrast, the Tharandt plot all eleven soil profiles are required to achieve the same accuracy threshold. This reflects the pronounced spatial heterogeneity of the site, where approximately half of the profiles exhibit little to no groundwater recharge while the remaining profiles produce substantially higher recharge. Consequently, small subsets frequently over- or underestimate the site-scale mean depending on the combination of profiles sampled. These findings demonstrate that the number of profiles needed to obtain robust groundwater recharge estimates is strongly site-dependent and controlled by the spatial variability of recharge processes.

Observed soil moisture at the investigated forest plots revealed pronounced differences, both laterally across the investigated area and vertically with soil depth with larger fluctuations and absolute values at the Tharandt plot. These differences are primarily driven by site-specific factors like rainfall and throughfall patterns, soil hydraulics as well as subordinated by vegetation properties. The relatively uniform temporal and absolute progression of soil moisture observed at the Kienhorst plot reflects the homogeneous characteristics of its soil matrix. The sandy soil, characterized by evenly distributed pore spaces and low water-holding capacity over the entire soil profile, contributes to the consistent distribution of soil moisture with moderate volumetric content across all depths. In contrast, likewise to the soil moisture patterns found by Berthelin et al. (2023) the soil moisture dynamics at the Tharandt plot exhibit significant variability, attributable to its heterogeneous soil characteristics. This includes alternating layers of silty soil with variable permeability, higher stone content, and underlying geological features such as impermeable layers of clayey material beneath the upper soil horizons. The observations align with the findings of Vereecken et al. (2016), who highlighted the influence of soil texture and hydraulic properties on spatial variability in soil moisture. At Tharandt plot, accumulation of stagnant water during periods of heavy rainfall at some locations underscores the site's sensitivity to rainfall. This phenomenon is consistent with prior studies by Famiglietti and Wood (1994), which emphasize the role of reduced soil permeability and hydraulic conductivity in limiting infiltration rates and enhancing water retention within heterogeneous soil profiles.

Although the uncalibrated model reproduces the general temporal patterns of soil-moisture dynamics, it systematically overestimates moisture levels and fails to capture the observed variability reflected in the poor variability and mean bias components of KGE. This is indicating structural and parametric deficiencies in the representation of soil hydraulic and vegetation properties. For example, overestimated saturated hydraulic conductivity (Ksat) or mis-scaled van Genuchten parameters may lead to overly rapid drainage and exaggerated soil moisture variability. Likewise, an underestimation of soil water storage capacity due to low porosity or limited rooting depth can further contribute to the observed biases. Such shortcomings may lead to unrealistic percolation fluxes and potentially inflated groundwater recharge estimates, underscoring the need for targeted model calibration at the two sites of investigation.

Following the identified performance limitations of the uncalibrated model, the subsequent calibration substantially improved the simulation of soil moisture dynamics and emphasises the need for site-specific calibration. Despite the spatial heterogeneities in soil characteristics, the SVAT model, when calibrated with site-specific soil moisture observations, was capable of capturing the observed dynamics and variability for most profiles at both plots with acceptable precision. The applied calibration substantially enhances confidence in the model's representation of subsurface water storage dynamics.

Nevertheless, discrepancies between observed and simulated soil moisture remain, particularly slight mean bias and variability bias tendencies that point to residual structural or parametric deficiencies. The model results tended to underestimate the rapid rise in soil moisture and struggles to replicate the absolute magnitude of the observed response during high-intensity precipitation events. This mismatch is likely partly related to the use of daily aggregated precipitation input, which smooths short-term rainfall intensities and may therefore attenuate rapid soil moisture responses to individual events. Consequently, short-term peak dynamics and associated rapid percolation pulses may be underrepresented. Because groundwater recharge estimates are directly derived from the simulated percolation response, the inability to represent these rapid transitions implies that recharge may be underestimated during such conditions. These discrepancies suggest limitations in the parameterization of the model for heterogeneous soil profiles and highlight inherent simplifications in the 1D modeling approach when preferential or fast flow processes dominate. Processes that have been identified as important drivers of focused recharge in heterogeneous soils (Berthelin et al., 2023; Ries et al., 2015) are not explicitly resolved in the model. This may restrict the model's capacity to fully capture the complexities of plot-specific conditions and processes such as spatial soil heterogeneity and the lateral water flux between neighbouring soil compartments (McDonnell, 1990).

In addition, weak or absent soil moisture peaks in the topsoil observed at the Tharandt site may partly be related to water storage and buffering effects in the litter layer, which is not explicitly represented in the model. Instead, litter processes are implicitly incorporated into the uppermost soil layer. This simplification may further contribute to dampened near-surface soil moisture dynamics, particularly during and shortly after precipitation events.

Part of the discrepancy between simulated and observed soil moisture may partly stem from the measurement uncertainty that originates from the accuracy of the soil moisture sensors used in this study, which is ±3 % of soil moisture (Truebner Company, 2026) and the effective uncertainty of field soil moisture measurements inherent to all soil moisture measurements (Jackisch et al., 2020). This uncertainty becomes particularly relevant during the substantial fraction of the observation period characterized by very dry conditions (<5 vol %), where the relative influence of the sensor error is large and the numerical dominance of dry states in the calibration dataset may influence parameter optimization. Although groundwater recharge predominantly occurs under wetter conditions, where percolation is active and sensor uncertainty is comparatively less critical, indirect calibration effects cannot be fully excluded. There is a risk that the calibration procedure, applied to the full time series, may favour parameter sets that reproduce the prolonged dry states well, while comparatively underrepresenting the wetter periods that are more relevant for groundwater-recharge generation. To assess the potential influence of this bias more systematically, future work should compare calibrations based on the full time series with calibrations restricted to periods exceeding θ>10 vol %, allowing an evaluation of how strongly the representation of dry conditions affects recharge estimates.

Further, the analysis showed that incorporating multi-depth soil moisture observations significantly improves the representativeness of the simulated soil water distribution, especially in heterogeneous soil systems as at Tharandt plot. Multiple observation points along the soil profile allow for a more nuanced assessment of the sensitivity of individual model parameters at different depths by calibrating the model against a higher density of data that captures both surface and subsurface processes. This was also observed by Houska et al. (2014), who demonstrated that the inclusion of soil moisture data at different depths increases the representativeness of the simulated soil water distribution and thus increases the identifiability of model parameters.

Building on the calibration results, the subsequent validation using the calibrated parameter sets revealed a mixed model performance, with acceptable temporal robustness at the Kienhorst plot but clear limitations at Tharandt plot, where the calibrated model struggled to reproduce wet-dry changes and soil moisture variability. These shortcomings are plausible given that the hydrological conditions for the validation period differed substantially from the calibration period. Consequently, the application of the Tharandt parameter set for estimating groundwater recharge in 2025 is associated with increased uncertainty and a site- and period specific recalibration would likely improve model reliability under such altered conditions. In general, it can be stated that the use of the calibrated parameters is unproblematic as long as hydrological conditions do not deviate substantially from those of the calibration period, whereas markedly altered conditions warrant a careful reassessment of their suitability for example through recalibration or by extending the validation period. Despite this limitations at the Tharandt plot, the calibrated parameter sets were retained because they provide a consistent and physically plausible representation of system dynamics during the calibration period and thus form a reliable basis for subsequent analyses for this period.

An analysis calibrated model parameter sets for both plots revealed a higher degree of variation in area and depth at the Tharandt plot compared to Kienhorst plot, as indicated by the coefficient of variation (Figs. S5 and S6 in the Supplement). This reflects the greater heterogeneity in soil physical properties observed at Tharandt plot, whereas Kienhorst plot is characterized by more homogeneous plot conditions and is a direct result of the spatially variable observed soil moisture dynamics found within the plots. Despite this broader range of plausible parameter combinations at Tharandt plot, the overall model performance remained at an acceptable level. However, a slightly reduced agreement between observed and simulated soil moisture values was observed compared to Kienhorst plot, suggesting a more complex interplay between parameter uncertainty and model behaviour in heterogeneous environments. The analysis further reveals that key soil hydraulic parameters and soil process parameters show a greater variation compared to vegetation parameters. This underscores their key role in soil water dynamics and highlights their contribution to overall model uncertainty. These insights are supported by the findings of Kreye and Meon (2016), who highlighted the significant impact of sub-scale spatial variability in soil hydraulic properties on hydrological process modelling. Similarly, Scharnagl et al. (2011) pointed out the value of incorporating prior knowledge of soil hydraulic parameters to enhance parameter identifiability in inverse modelling approaches. Complementing these perspectives, Baroni et al. (2010) demonstrated that uncertainties in the determination of soil hydraulic properties can substantially affect the overall performance of hydrological models.

Furthermore, analysis of the 30 best parameter combinations for each individual soil profile revealed that model-based uncertainty (the variation within the 30 best-performing parameter combinations) exceeds plot-based variability (differences between the 11 soil profiles). Both at the Kienhorst plot, where the limited variation in measured soil properties resulted in only minor differences in profile-specific parameterisation, as well as at the Tharandt plot, where input parameters varied more strongly, parameter non identifiability emerged as a key contributor to predictive uncertainty. The wide range of equally well-performing parameter sets for individual profiles reflects the issue of equifinality, where multiple parameter combinations can yield similar simulation outcomes. This suggests that model parameter ambiguity dominates the overall uncertainty. For several parameters, both sources of uncertainty, natural spatial heterogeneity and model-based calibration uncertainty, contributed similarly to the overall variation. At sites with pronounced heterogeneity in soil properties, such as Tharandt, this can amplify the spatial component of uncertainty and partly mask the effects of parameter equifinality. In particular, these findings show that the dominant source of uncertainty can vary significantly depending on plot characteristics. At homogeneous plots, model structural uncertainty and equifinality may outweigh physical variation, while at heterogeneous sites, spatial variability in input data may dominate. Therefore, parameter selection should be guided not only by sensitivity analysis but also by an underploting of the plot-specific balance between model and data-driven uncertainty.

Based on the findings, the calibration of soil hydraulic parameters should remain plot-specific, as their variability and influence on model outcomes is both large and highly context-dependent. Conversely, parameters with consistently low variability across profiles, such as many vegetation related parameters, may be suitable for regionalization or transfer between plots, potentially improving scalability and reliability of water balance simulations.

Despite differences, both sites displayed a clear alignment of major recharge periods with seasonal climatic conditions, indicating that atmospheric forcing is the dominant driver of recharge dynamics. In addition to seasonal effects, the main differences in the time series that can be attributed to the different soil properties of both sites. Groundwater recharge at the Kienhorst plot is characterized by smooth and sustained recharge dynamics reflecting the sandy soil texture, associated with high permeability and low field capacity, promoting continuous infiltration and a steady recharge response to precipitation events. Limited water retention in these soils allows for minor recharge even during the vegetation period, albeit at the expense of reduced water availability for plant uptake (Hillel, 2003). In contrast, the loamy and heterogeneous soil at the Tharandt plot, characterised by higher field capacities and lower matrix permeability, substantially delays the onset of percolation. The presence of discontinuous macropores and stone induced preferential pathways promotes rapid but spatially limited bypassing of the matrix, which reduces effective infiltration and prevents deep percolation until the upper soil layers approach their storage capacity. As a result, the recharge response is delayed and occurs irregularly in intermittent, pulse-like events, with groundwater recharge absent in some cases. The restricted matrix infiltration further contributes to the low recharge efficiency despite substantial precipitation inputs. Moreover, the increased water retention at Tharandt plot enhanced soil water availability for vegetation, effectively suppressing groundwater recharge during the vegetation period (Hillel, 2003). These findings underscore the critical role of soil hydraulic properties, particularly conductivity and retention capacity, in regulating the temporal dynamics of groundwater recharge, consistent with observations by Vereecken et al. (2016) as well as the studies of Beven and Binley (1992) and Zhao et al. (2018). This is emphasizing the complexity of water fluxes in heterogeneous soil systems and their limiting effect on recharge efficiency.

The bypass flow at Tharandt plot further influenced the recharge dynamics. Because a portion of the infiltrating water bypassed the fine-textured matrix and percolated directly through preferential pathways, the matrix response remained comparatively sluggish. This process prevented full replenishment of the soil water storage before deeper percolation occurred, reinforcing the damped recharge signal.

Beyond the general attenuation of the recharge signal, Tharandt plot exhibited variability among the 11 simulated soil profiles, exceeding the profile-to-profile differences observed at Kienhorst plot, with some profiles showing no groundwater recharge at all. The combined effects of matrix heterogeneity and bypass flow contributed substantially to the variability among the 11 soil profiles, reflected in differences in both the timing and magnitude of recharge events. Profiles with stronger bypass flow exhibited earlier and occasionally sharper recharge impulses, whereas profiles with weaker bypass flow showed more delayed and smoother recharge trajectories. This divergence reflects the structural heterogeneity of the site, including differences in macroporosity, stone content and the hydraulic properties of preferential flow domains.

Regarding to groundwater recharge and evapotranspiration quantities, the results for both sites fall within the range of values reported in the literature, derived from measurements and obtained from other site-specific simulations. For the Tharandt site, annual recharge rates of 80–150 mm for Tharandt in the period from 1968 to 1999 are suggested by Goldberg and Bernhofer, 2007, while the Groundwater Recharge Viewer from Saxony indicates values of approximately 25–50 mm for this location (Saxon State Office for Environment, Agriculture and Geology, 2025). Furthermore, calculated mean evapotranspiration across all profiles amounts to 566 in 2023 and 510 mm in 2024, which compares well with eddy covariance-based measurements of 479 and 482 mm for the respective years (ICOS Ecosystem Station Tharandt (DE-Tha), 2025), thereby supporting the plausibility of the model results. At the Kienhorst site, groundwater recharge rates derived from site-specific simulations are in the range from 30 to 296 mm in the period 1961 to 2019 (Ziche and Riek, 2024). Especially for the years 2023 and 2024, simulations from the Forest Environmental Monitoring Program of Brandenburg indicate recharge rates of 77 and 172 mm, respectively, alongside evapotranspiration estimates of 457 and 473 mm (Brandenburg State Forestry Enterprise, 2026). These values are in good agreement with the results obtained in this study, further supporting the robustness and plausibility of the simulated water fluxes.

At the Tharandt plot, the heterogeneous soil conditions and the resulting greater variability in parameterisation led to a much higher spread in the estimates of cumulative groundwater recharge across profiles compared with the more homogeneous Kienhorst plot. But neither site heterogeneity nor the range of input parameters alone fully explains the variation in recharge estimates. Rather, other factors, such as model structure, process representation and calibration uncertainty, appear to play a decisive role in shaping recharge variability at the catchment scale. This observation is consistent with findings by Maxwell and Condon (2016), who emphasised the complex interplay between soil water fluxes, landscape features and recharge processes as well as stressed that heterogeneity does not always lead to higher variability in groundwater recharge results. It further supports the notion that model behaviour can be dominated by structural and parametric uncertainty rather than physical input variability alone. A key contributor to this phenomenon is the concept of equifinality, as extensively discussed by Beven and Freer (2001) and Beven (2006). While the model is able to reproduce the observed soil moisture dynamics with reasonable accuracy using different parameter sets, the existence of multiple parameter sets with comparable performance indicates a high degree of parameter non-uniqueness. This not only increases predictive uncertainty, but also suggests that the model structure, particularly in its one-dimensional form, may not fully capture the spatially distributed and lateral processes that influence soil water movement leading to uncertainties in groundwater recharge estimation.

The results further show that the calculated groundwater recharge based on the parameter sets of the 30 best simulations for the mean value of the 11 soil profiles at the Kienhorst plot provides a generally representative recharge estimate. This averaging over serval soil moisture profiles in a forest was also proven to be useful by Berthelin et al. (2023). However, its representativeness varies interannually: in one year, the mean lies close to the edge of the ensemble of profile-specific simulations, whereas in the other year it is located near the centre of the ensemble distribution. This indicates that even in relatively homogeneous settings, the robustness of spatially averaged soil moisture signals is not fully constant over time. In contrast, at the Tharandt plot the mean-based recharge estimate also falls close to the centre of the ensemble; however, this agreement masks a pronounced bimodal distribution among the individual soil profiles. Approximately half of the profiles exhibit little to no groundwater recharge, while the remaining profiles generate comparatively high recharge rates. As a result, the mean provides a numerically representative value but does not reflect the underlying process heterogeneity of the system. This highlights the importance of a differentiated consideration of plot-specific soil profiles and moisture distributions in modelling, in order to ensure realistic results. Solely relying on mean values cannot always adequately capture the system's heterogeneity and poses the risk of systematic biases in the water balance, especially in heterogeneous settings, where nonlinear percolation processes and preferential flow introduce substantial variability.

Hence, in order to cope with uncertainties in obtaining representative recharge estimates, it is advisable to define an adequate number of soil profiles equipped with soil moisture sensors that serve as critical calibration points. The performed analysis (Fig. 14) indicated that below three randomly selected profiles at the Kienhorst plot, spatial variability as well as model uncertainty can still bias the representativeness of recharge estimates. The results demonstrate that at this relatively homogeneous site, the use of at least three soil profiles is recommended to ensure for 95 % of all groundwater recharge subset means to remain within ±10 % of the overall groundwater recharge mean at the 20×20 m plot scale when using the LWFBrook90 model calibrated with soil moisture data. In contrast, at the Tharandt plot, the strong heterogeneity of recharge responses, characterized by a pronounced bimodal distribution with approximately half of the profiles showing little to no groundwater recharge and the other half exhibiting comparatively high recharge, requires the full set of 11 profiles to achieve the same statistical robustness. Smaller subsets are unable to reliably capture this contrasting behaviour and may therefore lead to substantial bias depending on the selected profiles. In conclusion, the number of soil moisture profiles required for robust groundwater recharge estimation is strongly site-dependent. While three profiles may be sufficient in relatively homogeneous systems such as Kienhorst, highly heterogeneous sites such as Tharandt require full spatial coverage to adequately represent the underlying variability. A targeted monitoring approach should therefore consider site-specific variability structures to balance model reliability with fieldwork feasibility when applying the LWFBrook90 model.

Although the experimental setup is inherently site specific and relies on detailed soil characterization, vegetation information and profile-specific soil moisture observations, the study provides insights that extend beyond the two investigated plots. The results should therefore be interpreted as a proof of concept for high-resolution recharge modelling under realistic field conditions rather than as a universally transferable framework. Importantly, plot-scale analyses remain essential even in the context of multi-site or gridded recharge estimations. Large-scale approaches depend on upscaled parameterisations that may overlook local heterogeneities and nonlinearities in infiltration and percolation, whereas plot-scale studies provide the process understanding and high-resolution data needed to improve these models and to guide the selection of representative parameters. Focusing on plot-scale groundwater recharge allows model performance to be evaluated under well-controlled conditions using detailed soil moisture observations, thereby capturing site-specific processes – such as root-zone dynamics, canopy interception and soil hydraulic behaviour – that are often averaged out or simplified in large-scale frameworks. At the same time, the contrasting environmental conditions across the two forest plots, representing typical combinations of dominant tree species and soil types in Central European forests, suggest that the underlying process relationships are not unique to these sites but may be indicative of similarly structured and heterogeneous forest soils. Nevertheless, the generalisability of the results remains constrained, and future work should include multi-site comparisons, simplified proxy approaches and scaling strategies to more robustly assess transferability across broader environmental gradients. Bridging the gap between plot-scale understanding and landscape-scale application is therefore essential, and we see this study as a step toward that goal.