Introduction
The ongoing biodiversity crisis is now widely recognised as one of the most pressing global environmental challenges (IPBES 2019). Amongst the main drivers of this crisis, land-take, defined as a process which causes a modification of land use and of the characteristics of the soil (European Commission 2025), is identified as one of the direct drivers with a large global impact (IPBES 2019). Through processes, such as soil sealing and other forms of soil artificialisation, land-take contributes to soil degradation, which can negatively affect soil properties and the ecosystem services (ES) provided, including atmospheric composition and conditions and hydrological cycle and water flow regulation. All EU Member States and associated countries face growing pressure on soil and land resources, with over 60% of EU soils considered unhealthy (European Commission 2023). This challenge is addressed in the new Soil Monitoring Law (European Commission 2025), which establishes a harmonised EU framework for soil health monitoring and explicitly include land-take through indicators related to soil sealing and soil removal. The Directive also clarifies the definition of land-take, responding to previous criticism that the "No Net Land-Take (NNLT) by 2050" objective was insufficiently defined and inconsistently interpreted (Marquard et al. 2020). Under the Soil Monitoring Law, land-take encompasses two complementary dimensions. First, it refers to the conversion of natural and semi-natural land uses into settlement areas. Second, it includes soil artificialisation, defined as the durable alteration of soil components and characteristics, leading to a reduced capacity of soils to deliver ES. Recent research has underlined the importance of recognising and distinguishing these complementary dimensions (Hanna et al. 2024, Vieillard et al. 2024), showing that the capacity of a land to provide ES varies according to the degree of soil alteration and naturalness. A purely binary or surface-based understanding of land-take (land taken versus not) fails to capture the ecological functions or prior quality of the soils involved (Chéry et al. 2014). In response, several studies have highlighted the importance of integrating soil quality indicators and ES assessments into the broader discourse on NNLT (Decoville and Feltgen 2023, ESPON 2024b, Evers 2025), underscoring the need for more comprehensive and nuanced land-use evaluation frameworks. Urban planning and land development strategies have, however, traditionally approached the issue of land-take primarily through the lenses of economic growth, demographic expansion and urban sprawl, largely overlooking the pedological characteristics of soils and the associated losses in soil quality and their capacity to provide ES (Pauleit et al. 2005, Xiao et al. 2013, Ceccarelli et al. 2014, Ferrara et al. 2014). Consequently, land-use change and particularly the conversion of agricultural land and the sealing of soil, are assessed mainly through shifts in land-use land-cover (LULC) categories, with ES impacts inferred from differences in ES capacity associated with each LULC class (Lauf et al. 2014, Perennes et al. 2020). This is closely linked to the widespread assumption underlying many ES assessment methods, particularly the Tier 1 approach of the widely used ES matrix model of Burkhard et al. (2009), that LULC classes can serve as proxies for ecosystems and their service-providing capacity (Van Wijnen et al. 2012, Van der Biest et al. 2014).
The ES matrix method has long been described as having numerous advantages, as it is technically simple and quickly provides understandable and “mapable” ES data (Burkhard et al. 2012a, Burkhard et al. 2012b, Jacobs et al. 2015, Campagne et al. 2020). Such approaches can be adequate for certain ES or highly managed ecosystems (Schröter et al. 2015). However, relying solely on LULC classes (Tier 1 approach) has important limitations, particularly at finer spatial scales or for regulating ES, as it assumes fixed relationships between LULC classes and ES capacity derived from expert judgement or previous studies. This approach does not capture spatial variability within LULC types or the more detailed indicator-based relationships or models used in higher tier assessments (Tiers 2 and 3) (Jacobs et al. 2015, Campagne et al. 2020). Over time, various refinements have improved the application of the matrix method (Burkhard and Maes 2017), particularly by refining the characteristics of the geospatial units (i.e. the “service providing units” (SPUs); Syrbe and Walz (2012)), used in the matrix by incorporating information on ecosystem condition and spatial heterogeneity (Burkhard and Maes 2017). This can, for example, be done through the integration of geobiophysical data derived from soil maps (Kroll et al. 2012) or by systematically considering both abiotic factors and human management types to define the SPUs (Maebe et al. 2019).
In the context of the Mapping and Assessment of Ecosystems and their Services (MAES) framework, ecosystem condition is defined as the physical, chemical and biological state of an ecosystem at a given point in time (Burkhard et al. 2018). Indicators describing ecosystem condition, therefore, seek to capture both abiotic and biotic components, including soil, water and vegetation characteristics. Amongst these components, soils play a particularly important role (Willemen et al. 2012, Smith et al. 2017). While ES are not exclusively provided by soils, soils make a critical contribution that largely determines the extent to which many ES can be delivered (Paul et al. 2021). Soil-related ES depend on soil properties and their interactions and are strongly influenced by land use and management (Adhikari and Hartemink 2016). Integrating soil properties as proxies of ecosystem condition into refined matrix approaches can, therefore, improve the representation of ecosystem condition and increase the reliability of ES mapping and assessment (Campagne et al. 2020).
While some studies have attempted to quantify and qualify the loss of agricultural land due to artificialisation by overlaying LULC change data with soil quality maps (Kurucu and Chiristina 2007, Xiao et al. 2013), the joint consideration of both LULC characteristics and changes in soil condition resulting from land-take remains largely unexplored. As a result, the consequences of land change on ES delivery are seldom assessed in an integrated manner. Our study addresses this gap by:
analysing not only the original LULC classes affected by land-take, but also the change in ecosystem condition (based on soil and topographic information) resulting from this land-take;
enabling a more accurate estimation of the resulting impact on ES capacity.
This allows not only the identification of where land-take occurs, but also the assessment of the magnitude and qualitative nature of the ES losses associated with land-use change and the resulting changes in soil conditions. Such an approach is essential if we are to move beyond LULC classes and adequately account for the ecological value of soils and their multifunctionality in land-use planning.
This gap is particularly relevant in Wallonia (southern Belgium), where NNLT policies have been translated into a territorial development strategy (SPW Territoire 2024) through the concept of spatial optimisation (ESPON 2024b). Two trajectories are defined: one aiming at reducing net land-take and another aiming at limiting urban sprawl. It also includes a monitoring framework to track progress towards the goal of NNLT by 2050. However, neither the trajectories nor the monitoring framework account for soil properties or ecological functionality, relying solely on a quantitative surface assessment. The region is, however, characterised by considerable ecological heterogeneity, where short-distance variations in soil depth, texture or hydromorphy create a fine-grained mosaic of natural and semi-natural habitats (Claessens et al. 2023). These areas include not only forests and natural grasslands, but also a range of ecologically rich or sensitive agricultural systems. Understanding where land-take has occurred, how soil conditions have been altered and how these changes affect ES capacity is crucial for providing information for future land-use policies aligned with the NNLT goal of reducing soil degradation. This article responds to this need by proposing a spatially-explicit methodology that aims to:
Quantify artificialised areas in Wallonia between 2007 and 2018 and qualify the change in LULC classes for these areas;
Characterise the resulting change in ecosystem condition by developing a map of abiotic properties of soils and topography;
Build an ES capacity matrix using both LULC classes and the abiotic properties of soils and topography to refine the SPUs and complete the matrix by expert judgements to calibrate capacity scores;
Estimate losses in ES capacity resulting from land-take over the 2007–2018 period.
Material and methods
This study builds on a conceptual framework (Fig. 1) designed to assess how land-take affects ES capacity within the investigated period. The framework links changes in LULC with variations in ecosystem condition and their implications for ES capacity. In particular, we consider ecosystem condition through a set of abiotic proxies describing soil properties, together with topography context as an additional environmental descriptor. Based on this approach, we first characterise ecosystem conditions and topography across the study area to define SPUs and identify ecologically constrained areas where environmental factors play a key role in shaping ecosystem functioning. We then combine this information with LULC in a matrix of ES capacity to evaluate how land-take alters the capacity of ecosystems to support multiple ES. The following subsections describe each step of this framework.
Step 1 : Identification of newly-artificialised areas in Wallonia between 2007 and 2018 and assigning their LULC classes
Artificialisation is a process that refers to a change in land use in favour of so-called “urban” functions (such as housing, economic activities etc.) (Godart and Ruelle 2019). In this study, we used the concept of artificialisation as defined in the Walloon Spatial Development Scheme (SPW Territoire 2024), which aligns with the definition of land-take in the Soil Monitoring Law (namely, the conversion of natural, forested or agricultural states to urban functions, resulting in a durable alteration of soil properties and ES capacity) whether or not it is built upon and whether or not it is sealed (e.g. a parking lot vs. a suburban garden). The distinction follows the typology adopted in 2021 by the European Environment Agency (European Environment Agency 2023), whereby, for example, parks and private gardens are classified as artificialised (e.g. vegetated unsealed surfaces), while agricultural and forest land are classified as non-artificialised. The identification of newly-artificialised areas in Wallonia between 2007 and 2018, along with the attribution of their LULC classes, involved data preparation and standardisation, as well as a comparison between 2007 and 2018. All cartographic processing was performed using ArcGIS Pro© (ESRI Inc. 2023).
Data preparation and standardisation
To identify areas that were non-artificialised in 2007 and became artificialised by 2018 and to characterise their LULC classes, we relied on four official land-use or land-cover datasets produced by regional or national administrations. The 2007 baseline of non-artificialised areas in Wallonia was established using two official datasets: the Walloon LULC (COSW; SPW (2008)) and the national LULC map (Top10vector; IGN (2008)). In both datasets, land‑use and land‑cover information cannot be fully distinguished; they must, therefore, be interpreted as mixed LULC products rather than strictly land‑cover or land‑use layers. For 2018, we used the Walloon land-use map (WalUTS dataset -SPW et al. (2020a)), which delivers parcel-level information on 2018 current land use. As WalUTS contains only land-use information, it was complemented with the Walloon land-cover map (WalOCS - SPW et al. (2020b)), produced at 1 m resolution and simplified to 5 m for processing. Both COSW and WalUTS are grounded in cadastral information, ensuring methodological consistency for diachronic comparison.
In the 2007 baseline dataset, all polygons labelled as “artificialised” in the native classification (Level-1 class) were considered artificialised. Non‑cadastral parcels (Level-9 class) were added to the "artificialised" native classification to account for under‑represented linear infrastructures (e.g. roads), acknowledging that this choice also introduced limited misclassification (e.g. occasional watercourses incorrectly labelled as artificialised). Minor misclassifications (e.g. agricultural land adjacent to farm buildings) were deemed negligible, as our analysis focused on areas that shifted from non-artificialised to artificialised, not the reverse. To further improve reliability, COSW 2007 “non-artificialised” areas were cross-checked against national LULC map (TOP10vector). This correction reclassified approximately 126 km² (0.74 % of the study area) from non-artificialised to artificialised, thereby strengthening the baseline.
For 2018, artificialised areas were defined using both WalOCS and WalUTS datasets. All sealed surfaces and built-up areas from WalOCS were considered artificialised, regardless of their land use. In addition, parcels classified in WalUTS as residential, secondary or tertiary production areas, infrastructure, recreational facilities or gardens (levels 2 to 6 of the native UTS classes) were also considered artificialised. Waterbodies and watercourses were excluded for both 2007 and 2018 due to uncertainty regarding their artificialisation status.
The LULC class was determined for both time steps (2007 and 2018) using a harmonised LULC typology defined for this study. This typology was obtained by reclassifying the original COSW and WalUTS/WalOCS categories into a common set of LULC classes. For 2007, COSW categories were reclassified to match the study typology. For 2018, WalUTS land-use data were rasterised at 5 m and complemented with land-cover information from WalOCS; both layers were then combined to reconstruct the corresponding LULC classes. This harmonisation allowed us to reconstruct a complete LULC typology for 2018, consistent with the mixed LULC information available in the 2007 baseline.
Comparison between 2007 and 2018
To identify pixels that shifted from non-artificialised in 2007 to artificialised in 2018, COSW 2007 and WalUTS 2018 vector layers were first reclassified into binary classes (0 = non‑artificialised, 1 = artificialised) according to the classification rules described in the data preparation and standardisation subsection and then rasterised at 5 m resolution. A raster calculation was then performed to extract pixels that shifted from 0 in 2007 to 1 in 2018. Only these newly-artificialised pixels were retained for further analysis. LULC classes were then assigned to these pixels for both 2007 and 2018 using the harmonised LULC typology defined for this study, by applying a mask of newly-artificialised pixels to the LULC rasters.
A visual comparison using orthophotos from 2007 and 2018 was thoroughly conducted over the city of Namur to assess the accuracy of this method. Two major types of classification error were identified: firstly, artificialised areas of quarries were frequently overestimated in 2007, including wooded areas likely reforested more than 20 years ago. Conversely, some extensions of these quarries were missing in the 2018 dataset. Secondly, urban parks and urban woodlands were inconsistently classified in the source datasets, sometimes appearing as parks (i.e. artificialised) and sometimes as forests (i.e. non-artificialised). These inconsistencies stem from the original mapping rules of the datasets used — often influenced by ownership, cadastral categories or adjacent land‑use designations — rather than from our own processing. Based on these findings, we systematically corrected the most substantial misclassifications across Wallonia. This included a targeted review of large parcels that could have skewed the data if incorrectly classified in 2018. Parcels over 5 ha classified as “residential use” in WalUTS dataset were inspected visually. We also examined parcels under the “extractive industry” use class and large institutional or infrastructural properties. All secondary and tertiary land-use parcels over 15 ha intersecting new artificialisation pixels were also checked. In terms of transport infrastructure, only airport areas were reviewed, as these are most likely to span overestimated areas. Parcels found to be non-artificialised by these visual checks were corrected accordingly. Finally, we performed a last round of visual verification at a scale of 1:25,000 to ensure that large artificialised parcels were indeed valid. These corrections substantially improved the reliability of the land‑conversion dataset, as 286 large parcels were reviewed visually across Wallonia and 91 of them (≈ 32%) were found to be falsely classified as artificialised and subsequently corrected.
Changes in LULC between 2007 and 2018 were visualised using a Sankey diagram produced in R (v. 4.4.2, R Core Team (2024)) with the ggalluvial package (Brunson and Read 2026).
Step 2: Ecosystem condition assessment, based on key abiotic soil and topography properties
To account for the influence of soil and topographic constraints on ES capacity, we developed a map of ecologically constrained areas as a proxy for ecosystem condition in Wallonia. These areas correspond to zones where biophysical limitations restrict intensive land use. Although these constraints limit agricultural productivity or infrastructural development, they also tend to reduce human disturbance. These areas, therefore, retain semi‑natural vegetation and support higher habitat heterogeneity. The map identifies distinct categories of ecologically constrained areas defined by soil moisture conditions, slope characteristics (steepness and aspect) and shallow substrate depth.
The classification was derived from three geospatial datasets: (1) the Digital Soil Map of Wallonia (SPW 2015b) provided information on soil properties, such as natural drainage, peat occurrence, alluvial and colluvial deposits, hydromorphic features, soil depth and texture; (2) the Digital Terrain Model (SPW 2015a) was used to extract slopes exceeding 20° and aspect (north and south-facing slopes); and (3) a flood hazard map (SPW 2021) provided information for valley-bottom soils and hydrological constraints with hazard levels categorised as very low, low, moderate or high. For areas where soil information was missing, primarily within large urban centres, only topographic and flood‑hazard data were available. In these cases, the flood hazard map was used to refine the delineation of valley‑bottom areas and ensure spatial consistency across polygons. The resulting map (Fig. 2) delineates the nine main categories of ecologically constrained areas across Wallonia and allowed us to integrate soil‑ and terrain‑based limitations into the assessment of ecosystem condition and the ES assessment.
Each newly-artificialised pixel already characterised by its 2007 and 2018 LULC classes was subsequently assigned to its corresponding category of ecologically constrained area. This additional classification allowed us to link each pixel to the appropriate SPU in the capacity matrix.
Step 3 : Capacity matrix
Capacity Matrix Design
To assess the ability of ecosystems to deliver ES, a ES capacity matrix was used, with SPUs listed in rows and ES in columns. Unlike earlier approaches relying solely on LULC (e.g. Burkhard et al. (2009)), our matrix uses a refined typology (Suppl. material 1) of SPUs (matrix rows) inspired from Maebe et al. (2019) combining LULC data with soil and topographic constraints. Ecosystems were first classified according to the typology for ecosystem classification following MAES (Maes et al. 2013). Some ecosystems were grouped to reduce matrix dimension. The matrix was deliberately simplified for artificial LULC. These categories were primarily defined, based on the presence or absence of sealed surfaces, rather than the specific land-use function (e.g. industrial, residential, commercial). A distinction was, therefore, made between sealed artificial surfaces (such as transport infrastructure, all building types and sealed areas) and non-sealed artificial surfaces (including gardens, parks, sports facilities and other wooded or herbaceous covers considered artificialised; i.e. removed from their original agricultural or forestry use according to the definition applied).
Key abiotic soil and topographic properties were grouped into five classes for natural habitats, pastures and forests. These include four types of ecologically constrained conditions: peat soils and permanently waterlogged soils, temporarily waterlogged soils, steep slopes (> 20°) and shallow soils, plus one additional class representing non- or less ecologically constrained soils. For annual croplands, key abiotic soil and topographic properties were grouped into three classes, corresponding to two ecologically constrained types (peat soils and permanently or temporarily waterlogged soils and shallow soils and steep slopes (> 20°)), along with one class of non- or less ecologically constrained soils. The abiotic soil and topographic properties were not taken into account where not pertinent (e.g. artificial LULC classes). This cross-referencing allowed for a more nuanced characterisation of ecosystem types, while maintaining a balance between ecological relevance and operational feasibility in matrix design. Suppl. material 1 provides an overview of the matrix lines along with their alignment with the EU ecosystem classification of the environmental accounting framework (Eurostat 2025).
Priorisation of ES
To determine which ES would be included in the ES capacity matrix, we carried out a structured prioritisation process, based on expert judgement. The initial list of ES was drawn from the Walloon adaptation of the Common International Classification of Ecosystem Services (CICES) version 5.2 classification (Haines-Young 2023). Members of a multidisciplinary steering committee, comprising 15 experts from the relevant departments of the regional administration, were invited to evaluate each ES. They assigned a score from 0 (“not important”) to 3 (“essential to assess”), based on its ecological relevance, societal value and policy importance. In parallel, we assessed whether each ES was methodologically compatible with the matrix approach, particularly regarding its ability to reflect differences in ES capacity linked to soil and topographic characteristics. Only ES with an average score of at least 2 and deemed suitable for the matrix method were retained, resulting in a final set of 16 ES. Suppl. material 2 provides an overview of the matrix columns along with their alignment to the European CICES version 5.2.
Matrix Completion and Expert Input
The ES matrix was completed through an expert-based scoring approach. A total of 43 experts, representing a diverse range of organisations, including civil society, universities (researchers) and government administrations, participated in the ES scoring process. To manage workload, experts were grouped by field of expertise into five thematic groups, each focusing on a subset of ES (i.e. provisioning, water regulation, pollution control, climate, lifecycle maintenance, habitat protection and cultural ES). Group sizes ranged from 6 - 12 participants. Each cell of the matrix, corresponding to a specific SPU–ES pair, was scored on a scale from 0 (not relevant) to 5 (very high capacity), reflecting the perceived ability of the ecosystem to deliver that ES. Experts drew on their own knowledge and, where available, on pre-filled values derived from previous studies (i.e. Raquez and Dendoncker (2013), Maebe et al. (2019)). For each score, they were also asked to indicate their level of confidence, ranging from 1 (low) to 3 (high), to account for variations in expertise and data availability as advised in Jacobs et al. (2015). The scoring process was carried out over two workshop days in October 2021, and was preceded by an individual scoring phase guided by a shared tutorial. Each session included presentations, discussion of disagreements and efforts to reach consensus values. The standard deviation for each matrix entry was calculated from the complete pre‑filled matrices submitted prior to the workshop. When a consensus score was reached during the workshop discussions, this consensus value was adopted and its associated standard deviation was set to zero. When full agreement was not achieved, the initial standard deviation was retained to reflect the residual divergence amongst experts. Confidence scores were used as metadata describing the robustness of expert judgement rather than as inputs to the score calculation. The full capacity matrix, including the detailed list of SPUs (rows) and ES (columns), is available in Suppl. material 3, providing a comprehensive view of the typology and scoring system used. Standard deviations are provided in Suppl. material 4.
To quantify the relative contributions of LULC and abiotic soil and ecosystem condition to ES variation, a linear model was fitted using LULC and classes of key abiotic soil and topographic properties as explanatory variables (score ~ LULC + abiotic_properties). An analysis of variance (ANOVA) was performed and the sums of squares for each factor were used to calculate the percentage of variance explained. To further examine variability in ES capacity within and between SPUs, coefficients of variation (CV) were calculated for each ES across SPU classes. SPUs were grouped into major categories, based on their land-use type (e.g. artificial, cropland, natural and sylvicultural systems). The CV was computed as the ratio of the standard deviation to the mean of expert-assigned scores. The resulting values were visualised using a heatmap to highlight differences in variability across SPU classes and scored ES capacity. Only regulating and cultural ES were included in this analysis, as provisioning ES were not consistently represented across all SPU classes. All analyses and visualisations were conducted in R (v. 4.4.2; R Core Team (2024)).
Step 4 : Evaluating losses from land-take
The methodology for evaluating losses due to soil artificialisation is adapted from Campagne and Roche (2021). First, the ecosystems impacted by artificialisation are mapped by identifying each pixel that transitioned from non-artificial in 2007 to artificial in 2018. Each of these pixels is characterised both in terms of its initial and final SPU classification, based on LULC, soil data and elevation.
Next, for both the initial (2007) and final (2018) situations, a surface-weighted capacity matrix is calculated by considering the area of each SPU according to equation 1:
\documentclass[12pt]{standalone}
\usepackage{varwidth}
\usepackage[utf8x]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage{amsmath, amssymb, graphics, setspace}
\newcommand{\mathsym}[1]{{}}
\newcommand{\unicode}[1]{{}}
\newcounter{mathematicapage}
\begin{document}
\begin{varwidth}{50in}
\begin{equation*}
WS_{ij} = S_{ij} \times \frac{A_i}{\sum_{i=1}^{n} A_i}
\end{equation*}
\end{varwidth}
\end{document}
\(\) [1]
where: WSij = Weighted score for matrix cell i,j; i = single providing unit i (matrix line i); j = ES j (matrix column j); n = total number of single providing unit (SPU; total number of matrix lines).
This allows the computation of an overall capacity score for each ES (equation 2) before and after artificialisation took place:
\documentclass[12pt]{standalone}
\usepackage{varwidth}
\usepackage[utf8x]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage{amsmath, amssymb, graphics, setspace}
\newcommand{\mathsym}[1]{{}}
\newcommand{\unicode}[1]{{}}
\newcounter{mathematicapage}
\begin{document}
\begin{varwidth}{50in}
\begin{equation*}
FS_{j} = \sum_{i=1}^{n} WS_{ij}
\end{equation*}
\end{varwidth}
\end{document}
\(\) [2]
where: FSj = Final weighted score for ES j; WSij = Weighted score for matrix cell i,j; i = single providing unit i (matrix line i); j = ES j (matrix column j); n = total number of single providing unit (SPU; total number of matrix lines).
The final weighted scores range from very low to very high capacity, with the following thresholds: ≤ 1: very low capacity, > 1 and ≤ 2: low capacity, > 2 and ≤ 3: medium capacity, > 3 and ≤ 4: high capacity, > 4: very high capacity according to the qualitative expert‑based scoring scale. The capacity scores before and after artificialisation are then compared to quantify changes in the capacity of ecosystems to deliver ES. In order to interpret the magnitude of the differences obtained, we defined significance levels, based on the statistical distribution of ES value estimates, following the approach proposed by Campagne and Roche (2021). For each ES, we constructed a statistical distribution, based on the standard deviation derived from the pre‑workshop individual matrices (Suppl. material 4). The mean standard deviation of all service‑specific SPUs was then used to test whether the difference between initial and final capacity scores exceeded what could be expected from expert uncertainty alone. Statistical significance thresholds were derived using a Student’s t‑distribution with n–1 degrees of freedom (where n is the average number of complete matrices available, here n = 8). We report significance levels at α = 5%, 1%, 0.1% and 0.01%.
Results are synthesised using circular bar plots of the final weighted score of ES capacity, allowing for a holistic understanding of the cumulative impacts of land artificialisation. ES circular bar plots were generated using R (v. 4.4.2, R Core Team (2024)): packages ggplot2 (Wickham 2016) and patchwork (Pedersen 2024).
Results
LULC classes transitions of newly-artificialised land parcels
Fig. 3 illustrates the study area and the spatial distribution of land taken. Between 2007 and 2018, a total of 235.76 km² were converted to artificial uses (LULC classes) in Wallonia (corrected data), representing an average artificialisation rate of 21.43 km² per year, equivalent to about 5.87 hectares per day. Fig. 4 presents a Sankey diagram of LULC transitions between 2007 (left) and 2018 (right).
Agricultural areas were the primary contributors to land artificialisation: together, the three agricultural aggregated LULC classes (sown pasture, annual croplands and intensive orchards and nuts) accounted for 76.10% of the total newly-artificialised surfaces over the study period. Additionally, 16.42% of the newly-artificialised surfaces resulted from the conversion of forest areas (both deciduous and coniferous), while 0.2% resulted from the conversion of areas designated for nature conservation in 2007.
Most artificialisation occurred through the expansion of built-up and impervious surfaces (buildings and transport infrastructures (43.30%)), as well as residential gardens and parks (37.85%). The third most frequent transition corresponds to areas converted from agricultural or forestry uses to other human-dominated land uses that are classified as artificial in the LULC typology, but still retain a vegetated cover. These areas are either grouped under the category "Other artificial LU (secondary or tertiary sectors)" including both tree cover (5.33%) and herbaceous cover (9.21%) or represent "Sparsely vegetated ecosystems, quarries and waste lands" (1.63%) or "Sports and outdoor recreation facilities" (2.68%). Suppl. material 5 presents a map showing the spatial distribution of LULC classes between 2007 and 2018.
Soil and topographic characteristics of converted areas
Beyond the classes of LULC affected by land-take, it is also important to consider the physical characteristics and abiotic conditions of the impacted soils. Table 1 summarises the distribution of areas affected by land-take in Wallonia (2007–2018) according to key abiotic soil and topographic properties and compares these figures with the overall regional availability of such soils.
Overall, 12.31% of land take occurred on ecologically constrained areas. Given that these soils account for 15.93% of Wallonia’s total area, artificialisation has affected them slightly less than their regional representativeness. The bulk of artificialisation was still concentrated on less or non-ecologically constrained areas (84.07%), which is consistent with their broader availability and the fact that they generally pose fewer limitations to development, such as steep slopes or high soil moisture. Amongst ecologically constrained areas, hydromorphic alluvial and colluvial soils (4.10%) and non-alluvial hydromorphic soils (3.56%) were the most affected subtypes.
Matrix production
Analysis of the ES matrix revealed that LULC classes were the primary driver of ES capacity. Variation partitioning showed that LULC explained approximately 19.1% of the variance in ES scores, while key abiotic soil and topographic classes accounted for only 0.43%. The ANOVA confirmed that LULC had a highly significant effect on ES scores (F = 16.4, p < 2e-16), while soil and topographic classes showed no statistically significant effect (F = 0.73, p = 0.625). These results suggest that the expert judgements underlying the matrix give greater weight to differences between LULC types than to soil and topographic variability within each class. As our analysis focuses on transitions between LULC classes from 2007 to 2018 due to artificialisation, these changes are, therefore, expected to be the primary drivers of variation in ES capacity across the study area. However, ecosystem condition still influenced ES capacity within several LULC classes. For example, expert scores for timber production in coniferous forests increased along the soil gradient from 1 on peat soils and permanently waterlogged soils to 4.8 on less or non-constrained ecological areas.
The analysis of coefficients of variation across SPU classes revealed substantial differences in the variability of ES capacity evaluated by the experts (Fig. 5). The heatmap in Fig. 5 highlights that several regulating and cultural ES display higher variability within artificial (ART) SPU class compared with natural (NAT) or forested systems (SYL). In particular, ES such as water regulation, pollination, global climate regulation via greenhouse gas sequestration and visual screening showed strong contrasts amongst SPUs of the artificial class, reflecting the heterogeneity of artificialised SPUs. These results indicate that artificialised land cannot be considered as a homogeneous category, as sealed surfaces differ markedly from vegetated urban areas, such as parks, gardens or sports facilities, in their capacity to provide ES. The high variability observed for some ES within the agricultural (CRO) SPU class reflects the heterogeneity of land uses grouped in this category. For example, conventional cereal croplands generally shows a very low capacity for flood regulation, whereas orchards or mixed crops exhibit substantially higher capacities, resulting in larger coefficients of variation.
Loss of ES capacity from land-take and relative importance across ES
Land-take between 2007 and 2018 led to a systematic decline in the capacity of ecosystems to provide nearly all evaluated ES (Table 2, Fig. 6). Provisioning ES were the most severely impacted, with five out of five experiencing a 98–100% reduction in capacity. Notably, ES such as fodder production, livestock production and food crops showed complete capacity losses, all statistically highly significant. Amongst regulating ES, those related to hydrological regulation, such as soil erosion control (–64%), flood risk mitigation (–70%) and surface water quality regulation (–63%), showed major declines. These losses are closely related to the findings presented in the previous section, which showed that land-take mostly affected hydromorphic alluvial and colluvial soil subtypes (4.10%) and non-alluvial hydromorphic (3.56%) soil subtypes. These losses were backed by very strong statistical evidence.
Pollination (–63%) and global climate regulation through carbon sequestration (–61%) also showed substantial declines. In the cultural ES category, reductions ranged from –42% (recreation) to –61% (cultural and spiritual inspiration), with most losses supported by strong or very strong statistical evidence. Only the visual screening ES showed a minor and statistically insignificant decline (–23%). Overall, these results indicate a widespread reduction in ES capacity due to land-take.
Discussion
Land-take estimates and loss of agricultural land
The land-take per year land artificialisation rate estimated in our study (21.43 km² per year) for the period 2007 - 2018 is higher than the values reported in regional statistics from IWEPS data, which report an increase of at least 613 km² between 1985 and 2025, corresponding to an average of 15.35 km² per year in Wallonia (IWEPS 2025). This difference is likely related to methodological discrepancies between datasets. In particular, IWEPS reports net artificialisation balances, that account for both newly-artificialised and de-artificialised land. By contrast, our assessment considers only gross artificialisation, i.e. the total area newly converted without substracting areas that may have undergone re-naturalisation. Both IWEPS data (calculated over a longer time period) and the results obtained in our study are considerably higher than those published by the European Environment Agency (EEA 2023), which reports a total land-take of 104 km² in Belgium between 2000 and 2018 or approximately 5.7 km² per year. Urbanisation rates in Belgium as assessed at the European level are several times lower than those derived from high-resolution regional cadastral data. This discrepancy between national and European estimates has been discussed in literature (ESPON 2024a) and is largely explained by the fundamental difference in data sources: European figures are based on satellite-derived land-cover data (CORINE land-cover), which has a minimum mapping unit that limits its detection of small-scale changes, whereas national statistics typically rely on cadastral land-use data. As satellite methods often struggle to capture net land-take accurately (Decoville and Schneider 2015, ESPON 2020, Nicolau and Condessa 2022), they may diverge from administrative records (Fina et al. 2023), particularly when the distinction between land cover (physical traits) and land use (functional occupation) leads to misinterpretations. This highlights the importance of using more precise regional or national land-use data, which better reflects the diversity of artificialised areas.
Despite methodological differences between regional and European data sources, both point to the same pattern: agricultural land is the most commonly converted land type through artificialisation. In Belgium, according to European data (EEA 2023), a significant share of newly-artificialised land originates from agricultural areas representing 78.28 km² (respectively 42.99 km² and 35.29 km² for arable land & permanent crops and pastures & mosaic farmland) or 75.04% of total land-take between 2000 and 2018. Regional statistics (IWEPS 2025) also show a substantial decline in agricultural land over the past decades, with cadastral data showing a surface loss of 645 km² between 1985 and 2025, highlighting the steady loss of productive farmland to urban uses. This can be explained by the fact that, for the past three decades, urbanisation in Wallonia has largely followed the land-use zoning established by the regional statutory land-use zoning plan (i.e. plan de secteur, Charlier and Reginster (2024)). This zoning framework designates buildable areas near existing urban and village centres, which are historically located either on fertile agricultural plateaus or in the valleys of Haute Belgique (Defer et al. 2022). As a result, nearly 90% of land currently available for residential development still consists of agricultural areas (Charlier and Reginster 2024). The conversion of agricultural land is especially concerning because it affects areas that are not only agriculturally viable, but also often amongst the most fertile and accessible for food production. As this dynamic continues, it poses increasing challenges for food security (Ceccarelli et al. 2014, Gardi et al. 2015, Bren d’Amour et al. 2017, Van Vliet et al. 2017).
Landtake affects ecologically constrained areas
To better understand which environments are most affected by land-take, we characterised the ecologically constrained areas that are being converted or lost. Rather than assessing soil processes or soil functioning directly, we used a set of key abiotic soil and topographic properties to describe the ecosystem condition of these areas. These variables were, therefore, not intended to measure soil functions themselves, but to capture the environmental context in which ecosystems operate and, consequently, their potential capacity to provide ES. This perspective differs from studies that explicitly analyse how soil properties control ecosystem functioning, soil-related ES or the supply of soil functions across landscapes (e.g. Tóth et al. (2013), Greiner et al. (2017), Su et al. (2018), Wadoux et al. (2025)). Such studies typically focus on detailed soil characteristics — such as soil organic carbon content, available water capacity, clay and silt content, soil type, soil depth, bulk density or hydromorphic features related to soil–water dynamics — to explain processes including primary productivity, water and climate regulation, nutrient cycling or habitat provision. Our results indicate that, although land-take primarily affects less or non-ecologically constrained areas, substantial losses also occur on hydromorphic soils. This pattern is consistent with previous studies showing that urban development has often been associated with the widespread degradation and loss of wetlands and other moisture-dependent environments (Ehrenfeld 2000).
The impact on land-take on ES capacity
Across the study area, we observe a net reduction in overall ES capacity over 2007–2018. This finding is consistent with prior evidence from diverse contexts showing ES reductions associated with land‑use change (Pauleit et al. 2005, Haines-Young et al. 2012, Ferrara et al. 2014, Lauf et al. 2014, Kindu et al. 2016, Gao et al. 2017, Song et al. 2018, Liu et al. 2019). The expansion of built‑up areas has been shown to negatively affect ES not only within urbanised zones, but also in adjacent natural landscapes (Alberti 2005, Seto et al. 2012). A comprehensive review by Hasan et al. (2020) further indicates that the magnitude and direction of ES responses vary according to ecosystem type and the specific land‑use transitions considered. In our case, the aggregate decline aligns with the dominance of conversions from agricultural land (17,942.74 ha; 76.10% of all land‑take surfaces), which mechanically depresses the capacity of provisioning ES (Jiang et al. 2015). More broadly, the conversion of cultivated land to construction land has repeatedly been identified as a major driver of ES capacity loss associated with land‑use change (Guo et al. 2022). However, the pattern for regulating ES warrants closer examination because newly-developed land comprises a mix of sealed and non‑sealed surfaces: in our dataset, 43.30% of newly-taken land is sealed, while 37.85% corresponds to parks and gardens. Non‑sealed urban green can score higher capacities than some agricultural uses for specific regulating ES (Evans et al. 2022). This is reflected in our expert‑informed matrix for erosion control, for instance, where impervious surfaces scored 0, parks & gardens 3.0, cereal cropland on good soils 1.8 and permanent grassland 4.8. Yet, despite the comparatively better performance of urban green relative to arable fields, the net effect at regional scale remained negative for the capacities of regulating ES. Several interacting factors help explain this outcome. A large proportion of the converted land consists of sealed surfaces, which systematically receive a score of 0 in the matrix. At the same time, although parks and gardens perform better than sealed areas, the area‑weighted balance of conversions means that the gains they offer do not compensate for the losses associated with both sealing and the disappearance of high‑performing categories, such as grasslands.
Methodological considerations on the matrix used
The approach adopted in this study for the evaluation of the ES, integrating LULC information (e.g. to distinguish between primary, residential and nature conservation uses) with data on ecologically constrained areas within the matrix framework, has enhanced the capacity matrix methodology (see also Maebe et al. (2019)). To better interpret these results, we examine key methodological elements of the capacity matrix, including the definition of SPUs, the selection of ES and the review of the expert-based scoring process together with the associated confidence levels. These elements are known to be linked to both technical and thematic sources of uncertainty in ES assessments (Hou et al. 2013).
Relevance of SPUs (LULC categories and ecologically constrained areas)
A first issue that deserves attention is the definition of the rows in the ES capacity matrix, which reflect the SPUs used in our analysis. In this study, SPUs were defined by combining LULC classes with ecologically constrained areas derived from key abiotic soil and topographic properties. While the use of land-cover typologies is common (many studies reviewed by Campagne et al. (2020) used the European Corine Land Cover typology), we opted for a combination of LULC information, based on Walloon typologies. This choice was motivated by the need to better distinguish between different types of artificialised areas, notably by separating built-up, sealed surfaces from non-sealed artificial features, such as urban green spaces. In practice, this required integrating land-use information (e.g. nature conservation, primary uses considered as non-artificial and residential or tertiary uses considered as artificial) with land-cover data for the study area. Finally, complementing LULC data with additional geobiophysical information derived from soil maps or hydrological datasets is not unprecedented (e.g. Kroll et al. (2012), Nedkov and Burkhard (2012)). It has indeed been argued that LULC alone lacks information regarding important components of ecosystem conditions that support ES capacities, such as soil type and quality, water availability, geomorphology or overall ecosystem integrity (Campagne et al. 2020). In our study, the integration of soil abiotic and topographic properties was specifically designed to capture variations in ecosystem condition associated with soil contexts and topography. This perspective is particularly relevant in the context of land-take policies, where attention should not only be given to the amount of land converted, but also to the ecological quality of the areas affected, as territories with higher ecological value are likely to sustain a broader range and capacity of ES provision (Senes et al. 2023).
ES prioritisation
A second methodological element relates to the selection and prioritisation of the ES included in the columns of the matrix. Previous reviews of the ES matrix approach have assessed an average of 15.6 ES per matrix (Campagne et al. 2020), meaning that the number of ES considered here (16) is consistent with literature. The choice of the ES included in this study was not primarily driven by their direct relevance to soil-related functions. Rather, it was guided by their importance in the Walloon context and by the ability of the matrix design to capture variations in ES capacity across different land-use types and ecological conditions. Overall, this selection seeks to ensure consistency with previous studies conducted in the area (Raquez and Dendoncker 2013, Maebe et al. 2019), while ensuring its relevance for analysing the impacts of land-take.
Nevertheless, a comparison with literature (Paul et al. 2021) shows that most of the selected ES correspond to ES identified as soil-related or soil management-related. In particular, several provisioning services included in our assessment (e.g. food crops, timber and wood fibre, fodder, bioenergy production) correspond to the CICES classes related to cultivated terrestrial plants that are consistently identified as strongly linked to soils. Similarly, key regulating services considered in this study, including flood risk mitigation, water quality regulation, climate regulation and soil erosion control, align with CICES classes frequently highlighted in literature as directly influenced by soil properties and management (Schwilch et al. 2016, Bartkowski et al. 2020, Steinhoff‐Knopp et al. 2021, Notte et al. 2022). A number of cultural ES in our set, such as recreation, scientific and educational inspiration and cultural or aesthetic values, also correspond to classes identified in the review as having a soil-related dimension (Paul et al. 2021). By contrast, some ES prioritised in our analysis (e.g. filtration of airborne pollutants, visual screening or pollination) are less explicitly represented amongst the most frequently cited soil-related CICES classes, although their delivery can still be indirectly influenced by soil condition and land management. During the scoring process, experts found it particularly difficult to differentiate scores within a given LULC category for these three ES. As a result and following the discussions held during the workshop, it was decided to apply identical values across all abotic soil and topographic classes for most LULC types for filtration of airborne pollutants, visual screening and pollination.
Expert-based scoring
A third important aspect concerns the reliance on expert judgement to assign ES capacity scores. While expert-based approaches are commonly used in ES assessments, they inherently introduce uncertainty related to subjective perceptions and varying levels of confidence (Jacobs et al. 2015). To address this, several measures were implemented to improve the robustness of our results. First, consensus rounds were conducted: experts were individually consulted prior to the workshop, median scores were calculated from these initial inputs and the workshop was then used to discuss discrepancies and reach a consensus, allowing each expert the opportunity to reconsider their initial judgement (as suggested in Campagne et al. (2020). Although some participants found the matrix demanding — particularly due to the inclusion of ecologically constrained areas — facilitated discussions helped clarify its rationale. Second, confidence reporting was implemented, as recommended by Jacobs et al. (2015), to enable quality comparisons. Confidence scores were expressed by experts together with their scores for each ES and for each SPU as in Campagne et al. (2017) and Elliott et al. (2019). Reported confidence was lower for non‑agricultural or urban SPUs (mean confidence score of 2.7 for artificial SPUs versus 1.6 for agricultural and forest SPUs on a 1 = “comfortable” to 3 = “uncomfortable” scale). Several factors likely contributed to this pattern: the intrinsic heterogeneity of artificial SPUs complicates consistent scoring; in addition, the expert panel was predominantly composed of specialists in agricultural and forest systems, which reduced perceived confidence when dealing with urban contexts. Third, beyond median values, we explicitly examined the variability of expert scores by computing standard deviations across expert scores. On average, the standard deviation for non‑artificial SPUs across all services was 0.6, whereas artificial SPUs exhibited a higher mean standard deviation of 0.9, with numerous cases reaching or exceeding 1.0, notably for abandoned areas (wastelands), gardens and parks and all SPUs grouped under "Other artificial land uses (secondary or tertiary sectors)" classes. This greater dispersion for artificial contexts reflects both the intrinsic heterogeneity of these environments and the wider range of plausible interpretations regarding their ES capacity. It also points to a broader methodological issue: the level of detail used to represent artificialised areas in the matrix. Although the approach distinguishes several types of artificial land uses, it does not include specific classes for urban green spaces or, more broadly, for non-sealed artificial surfaces. This may partly limit the ability of the analysis to fully capture the diversity of ES within urban environments and make the expert scoring assessment more complex. Recent research has highlighted the importance of differentiating between types of urban green infrastructure when assessing ES, as artificial spaces are far from homogeneous and can exhibit widely differing levels of land-use intensity and ecological potential (Decoville and Feltgen 2023). For instance, Evans et al. (2022) showed that ES delivery varies substantially depending on the type of urban growing space, with community gardens, parks, allotments and other urban green spaces being particularly conducive to a diverse range of ES. This may have contributed to more conservative assessments for urban green areas as Kaiser et al. (2013) showed that less well-known or indirectly used ecosystems or ES are often underestimated by the local experts.