Best Siting for Small Hill Reservoirs and the Challenge of Sedimentation: A Case Study in the Umbria Region (Central Italy)

Abstract

This study presents a GIS-based Weighted Overlay Process (WOP) for Small Hill Reservoir Best Siting (SHRBS) in the Umbria region (central Italy), with a focus on supporting regional-scale planning rather than site-specific engineering design. The WOP incorporated commonly adopted SHRBS criteria, with suitability scores defined through two approaches: Model A, based on scoring scales from the literature, and Model B, which assigns scores based on the frequency distribution of the various attributes observed in a database of over 3000 existing SHRs in the region. The comparison between the models revealed significant differences, particularly in the scores assigned to texture, precipitation, and contributing area. Models A and B, tested on the existing SHRs, indicated quite different average suitability values (2.68 and 3.30, respectively, on a 5-point scale) and only a slight agreement (weighted Cohen’s kappa Kw ≤ 0.13). Both models also showed poor agreement (Kw < 0) when compared with a third suitability model based solely on sedimentation risk, which was developed using the Sediment Delivery Ratio from the InVEST suite. This indicates that many sites considered highly suitable by models A and B were also highly susceptible to sedimentation. Given the economic and environmental implications of sedimentation, this study recommends explicitly incorporating sedimentation risk criteria into SHRBS methodologies to enhance the effectiveness of siting decisions.

1. Introduction

Small hill reservoirs (SHRs) are crucial in water resource management in upland and hilly regions [1]. These small- to medium-sized reservoirs are essential for agricultural irrigation, livestock watering, aquaculture, domestic use, and ecosystem restoration. They help to regulate water availability, mitigate drought and flood impacts, and reduce dependency on external water sources [2]. SHRs are widely used in many regions worldwide, especially in areas where topography and water scarcity necessitate localized water storage solutions [1,3,4]. These reservoirs can capture and store surface runoff during rainy seasons, providing a stable water supply for various purposes during dry periods. In many rural and semi-arid regions, they are often the backbone of local water security [5,6,7]. The success of these water bodies depends mainly on their location. Poor siting can lead to structural instability, insufficient water collection, and increased vulnerability to sedimentation [8]. Thus, identifying the best site is critical in planning and constructing SHRs. The SHR Best Siting (SHRBS) involves a multi-disciplinary approach that considers physical, ecological, and socio-economic factors. The hydrology of the catchment area is the most important factor in SHRBS. The selected location should have sufficient runoff potential to fill the reservoir while minimizing water losses. Key assessment parameters include rainfall patterns, catchment size, and soil infiltration rates. Areas with impermeable or low-permeability soils, such as clay soils, are preferred for minimizing seepage [9]. The underlying bedrock should also be stable to avoid potential structural failures. Additionally, the site should not be prone to landslides or excessive erosion, which could compromise the reservoir’s safety and longevity. Other aspects to be considered are potential ecological and environmental concerns and socio-economic factors (economic feasibility, proximity to agricultural areas, etc.).

Detailed and specific surveys capable of accounting for all these aspects can be carried out when the goal is the best siting and the subsequent construction of a specific structure. For large-scale SHRBS studies, it is necessary to use rapid procedures to classify the territory based on various relevant criteria, preferably already available as georeferenced information or easily derived through simple processing. Such studies are often required in planning assessments to provide a preliminary evaluation of a territory’s potential for a specific use and to identify sub-areas where suitable conditions are more common. Numerous examples of SHRBS studies can be found in the literature. A commonly used technique is multi-criteria analysis, which is known as the Weighted Overlay Process (WOP) and uses georeferenced information layers. For example, Adham et al. [3,9] defined and applied different types of multi-criteria analysis within a Geographic Information System (GIS) to test the suitability for constructing rainwater harvesting structures in Tunisia and Iraq, respectively. Similarly, Dile et al. [10] assessed the potential for water harvesting in the Upper Blue Nile Basin using two GIS-based multi-criteria evaluation methods. Forzini et al. [5] conducted SHRBS analyses for the specific case of “pockari” in the mountain rural communities of the Himalayas. Mugo and Odera [11] integrated the use of GIS and runoff production models to assess the suitability for constructing various rainwater harvesting structures in Kenya. Shadmehri Toosi et al. [12] demonstrated the importance of including socio-economic and practical aspects in multi-criteria analyses aimed at identifying the most suitable areas for the implementation of rainwater harvesting structures in Iran. In a study focused on northern Ghana, Umuzika et al. [13] showed how planning studies based on multi-criteria decision-making techniques can help optimise the placement of dams and significantly increase stored water volumes by as much as 30–60%. Multi-criteria analyses also demonstrate high versatility in supporting specific planning objectives. For example, in a recent study, Arabatzis et al. [14] mapped optimal reservoir locations for wildfire prevention, whereas Hosseinzadeh et al. [15] identified suitable sites for detention ponds in an urban context.

One of the major challenges associated with SHRs is sedimentation, i.e., the gradual accumulation of soil particles, organic matter, and debris in the reservoir. This process, driven by erosion in the catchment area, can significantly reduce the storage capacity of a reservoir over time. Sedimentation undermines the reservoir’s functionality, increases maintenance costs, and shortens its lifespan [16]. The development of methodologies or studies that help identify suitable sites for reservoir construction—while also accounting for sedimentation risk—is of primary interest not only to private landowners (e.g., farmers), but also to policymakers. On one hand, landowners benefit from selecting locations where sedimentation is less likely to undermine the investment. On the other hand, legislators could use such insights to encourage reservoir development only in areas with lower erosion risk or to require the adoption of soil conservation practices as a prerequisite for construction. Moreover, from a broader perspective related to water resource management, sedimentation introduces significant uncertainty in the assessment of a territory’s actual water availability [17].

The sedimentation rate depends on various factors, including the geology and vegetation cover of the catchment area, catchment size, land use practices, and the intensity of rainfall events [18]. Poorly sited SHRs, especially in catchments with high erosion potential, are particularly vulnerable to rapid sedimentation.

Some of the variables (e.g., slope, catchment area, land use, and land cover) commonly used in multi-criteria analyses aimed at SHRBS are inherently capable of providing indications that should help minimize the risk of identifying them as suitable areas highly prone to sediment production and, consequently, sedimentation. However, a specific assessment of this aspect is usually not present in the literature [19].

In light of these considerations, the present study analyzed and compared the results of two different types of SHR multi-criteria analyses for an Umbria region (Central Italy) case study. Both approaches are based on criteria commonly used in the literature for this purpose, with the key difference being that in one case, scores are assigned based on scales derived from the literature, while in the other, scores for each criterion are proportional to the frequency with which attributes or values appear in existing reservoirs.

Finally, both suitability classifications are compared with a suitability assessment specifically focused on sedimentation susceptibility, derived from estimates of the Sediment Delivery Ratio model from the InVEST^® suite, version 3.16.0 [20].

Therefore, the main objectives of this study are (i) to assess the effectiveness of commonly used multi-criteria analyses for SHRBS in hilly areas of Central Italy and (ii) to compare the suitability maps deriving from such analyses and those derived from analyses focused only on the sedimentation risk criterion.

2. Materials and Methods

2.1. Study Area

The analysis focuses on Umbria, a central Italian region of about 8456 km², characterized by a complex morphology (Figure 1).

Most of the territory is hilly, with altitudes between 200 and 800 m asl. The plains are limited to the valleys crossed by the main watercourses and the areas bordering Lake Trasimeno. The mountainous areas are concentrated in the eastern part of the region along the ridge of the Apennine mountains, with reliefs that reach a maximum value of 2414 m asl. Figure 1 shows the spatial location of the region, the distribution of the main watercourses and reliefs. Umbria has a Mediterranean climate, with dry and warm summers and mild winters. Based on the E-OBS gridded data at a 0.1 × 0.1° resolution (version 29.0e) [21], the mean annual rainfall amounts vary between 705 and 1348 mm, while the average annual temperature is between 6.4 °C and 15.8 °C.

2.2. Data

2.2.1. Topographic, Meteorological, Soil and Land Use Data

Various types of quantitative and qualitative data were retrieved to characterize the study area. The meteorological data used are daily values of cumulative precipitation and minimum and maximum temperatures. To derive this information at the regional scale, the E-OBS gridded dataset provided by Copernicus was utilized. The E-OBS dataset is based on the processing of ground-measured data provided by various national meteorological centres across Europe. It offers different types of daily data starting from 1950, with resolutions of 0.1° × 0.1° and 0.25° × 0.25°. For this study, the 0.1° resolution and version 29.0e were used, covering the period 1950–2022 (accessed in August 2024). The data, available globally, were clipped to the Umbria region and processed using R software. The maximum and minimum temperatures were used in the Hargreaves and Samani model [22] to obtain an estimate of the reference evapotranspiration (ET0), with the same spatial resolution as the original E-OBS data. The maps in Figure 2a,b show the spatial variability of annual precipitation and ET0, respectively.

The digital elevation model (DEM) of the region (Figure 1), with a 10 m resolution, was obtained from [23]. From this DEM, slope and flow accumulation layers (i.e., upslope contributing area) were derived using QGIS (version 3.32.3) and Whitebox (version 1.3.0) software.

Soils were characterized in terms of soil texture (Figure 3). This information was retrieved from the ESDAC (European Soil Data Center) by accessing the raster soil data at a 500 m resolution [24].

Land use and land cover information was obtained from the CORINE programme (Coordination of Information on the Environment). The data used for the study refers to 2018, the most recent and reliable version to date, created using both the European Sentinel-2 satellite system, part of the Copernicus programme, and the American Landsat-8 system [25]. The data is characterized by particularly accurate spatial resolution, with a precision of 100 m, and distinguishes 44 land use-land cover classes, 29 of which are found in the Umbria region, as shown in Figure 4.

All input layers were used at their native spatial resolution without resampling. This choice was made to preserve the integrity of the original data and avoid introducing artificial spatial detail, especially for coarser datasets such as soil texture. Conversely, high-resolution layers such as the DEM were maintained at 10 m to ensure accurate derivation of terrain-dependent variables (e.g., slope).

2.2.2. Database of Existing Reservoirs

The georeferenced database of small reservoirs used in this study consists of 3206 water bodies and was developed by Casadei et al. [7]. The dataset is based on information originally provided by the Umbria Region and primarily includes reservoirs located in the Province of Perugia, which covers approximately 71% of the regional territory.

As explained in [7], the initial dataset underwent extensive verification and enhancement. First, all available data were analyzed and georeferenced for the entire Province of Perugia. Since direct volume measurements were missing in many cases, reservoir capacities were estimated using empirical correlations between surface area and volume. Surface areas were also cross-checked and validated using aerial imagery. For further details, the reader is referred to the original publication in which the dataset was developed and presented.

As shown in Figure 5, most of the reservoirs (about 73%) have a surface area smaller than 2500 m², and the information on storage capacity remains uncertain for a significant portion of the dataset. Moreover, the database does not distinguish between lowland (i.e., groundwater-fed) and hillside (i.e., stream-dammed) reservoirs, which is a known limitation. Despite these uncertainties, the dataset represents the most comprehensive and consistent inventory currently available for small reservoirs in the study area and has been previously used in hydrological and water resource studies at the regional scale [7].

2.3. Methods

2.3.1. Weighted Overlay Process

The method used for the spatial characterization of SHRBS was based on a multi-criteria decision-making (MCDM) analysis. MCDM represents a key challenge in decision analysis, focusing on identifying the optimal alternative by evaluating multiple criteria during the selection process. The technique can be applied using different algorithms and methods [26]. MCDM techniques that rely on spatial information and make use GIS tools are known as the Weighted Overlay Process (WOP). Generally, the steps required to develop a WOP analysis include the following:

• Defining the criteria, i, considered relevant for determining suitability;
• Defining the classes, c, within each criterion;
• Establishing a scale of normalized suitability scores s_i,c for each class (e.g., from 1 to 5 or from 1 to 10) within each criterion;
• Optionally assigning different weights, w_i, to each criterion based on their relative importance;
• Computing the suitability score (S_su) for each cell to produce the final suitability map as follows:

$$S_{su}={\sum }_{i=1}^N{w}_i{s}_{i,c}$$

(1)

where N is the number of criteria; w_i is the weight assigned to the i-th criterion (with ${\sum }_{i=1}^N{w}_i=1$)) and s_i,c is the score corresponding to class c of criterion i to which the cell belongs according to the established scale.

2.3.2. Model A

The first WOP method (Method A) was defined based on criteria and corresponding scores solely derived from the literature in studies with the same objective. For some criteria (e.g., precipitation), considering that the studies in the literature refer to various areas across the globe, the rationale was maintained, but the value ranges were rescaled based on the specific characteristics of the study area.

Table A1. List of criteria considered in the WOP analysis and their suitability scores (Model A).

Criterion	Suitability Score
	5 (Highly Suitable)	4	3	2	1 (Low Suitability)
Texture, T	Clay	Silty-clay	Sandy-clay	Sandy-clay-loam, Sandy-loam	Clay-loam, Silt, Silt-loam, Loam, Sand, Loamy-sand, Silty-clay-loam
Contributing Area, CA (km2)	0.005–0.5	0.5–5	5–50	<0.005	>50
Slope, Sl (%)	1.5–2.5	2.5–4.5	<1.5	4.5–7.5	>7.5
Precipitation, Pr (mm/year)	962–1091	1091–1219	833–962	>1219	<833
LULC	Sparsely vegetated areas, inland marshes, pastures, natural grasslands, non-irrigated arable land, predominantly agricultural land with significant natural vegetation areas, water bodies	Fruit trees and berry plantations, annual crops associated with permanent crops, complex cultivation patterns, olive groves, sclerophyllous vegetation, vineyards	Burnt areas, transitional shrubland-forest	Broad-leaved forest, mixed forest, coniferous forest	Bare rocks, mining sites, beaches, dunes, sands
Simplified Water Balance, WB (mm/year)	<−200	−200–0	0–200	200–400	>400

presents the six criteria considered and the corresponding scores, which were established based on reference studies and normalized within a scale from 1 (low suitability) to 5 (high suitability). Below, a summary of the basis for the assigned scores is provided.

For the texture (T) criterion, a fine granulometry, predominantly clayey, is preferred due to the excellent physical properties of clay regarding water retention, allowing for better resource storage. Additionally, clayey soils tend to promote runoff processes rather than infiltration, facilitating water flow and accumulation in reservoirs [5]. Fine-grained soils are also suitable for the construction of the dam’s embankment, reducing construction costs and making the project more feasible. As the clay content in the soil decreases, greater importance is usually given to the silt component rather than the sand component, as the latter has a strongly negative impact on water retention. The scoring scales reported in Table A1 are exactly those presented in [9].

Regarding the size of the contributing watershed area (CA), both excessively small basins (as they are unable to provide sufficient runoff) and excessively large basins (which would lead to oversized ancillary structures, e.g., spillways) are generally considered unsuitable. Therefore, adapting the values proposed for this criterion in the study by [5], the highest score of 5 was assigned to basins between 0.005 and 0.5 km², while lower scores were given to excessively small contributing areas (<0.005 km²) and excessively large ones (>50 km²).

For slope (Sl), sloping terrains are considered to hinder runoff, while excessively steep slopes would accelerate the silting process of the reservoir due to erosion and solid transport downstream. Moreover, overly steep basins and tributary streams reduce the impoundable volume at equal dam height. The slope scoring classes in Table A1 are exactly the same as those reported in [3].

Regarding annual precipitation (Pr), the scores were assigned following the scheme proposed by [10], adapted based on the statistical analysis of annual precipitation in the region. The maximum and minimum annual precipitation values for the Umbria region, 1348 mm and 704 mm, respectively, were extracted from the initial dataset. Starting from the minimum value, five intervals were created, each with a class increment of 129 mm. The central class, 962–1091 mm, was given greater importance due to the dual interpretation of the criterion: on one hand, an artificial reservoir may be more useful in areas with lower water availability (i.e., less rainfall), but on the other hand, if an area does not receive enough precipitation to fill the reservoir, constructing one there would not be practical.

Land use (LULC), as texture, represents a qualitative criterion, where classification is based on attributes rather than value ranges. To ensure the most accurate analysis, the classification of this criterion incorporated information from three studies [5,27,28]. The aim was to prioritize areas that are more “modifiable,” favouring, for example, croplands over forests and entirely excluding all urbanized and cemented soils by assigning them a score of 0 (see Equations (2) and (3)), the only such case among all criteria used in the analysis.

Finally, although less common in rainwater harvesting studies, the average annual value of the simplified climatic water balance (WB = Pr–ET0) was also considered among the criteria. Specific scoring scales were defined for the case study, given the excessive variability and occasional inconsistency of those found in the literature. Based on the WB range observed in the region (approximately from −400 to +600), five intervals of 200 units each were identified. A decreasing suitability score from 5 to 1 was assigned to these intervals, starting from the class between <−200, which corresponds to the most favourable (i.e., most negative) water balance conditions. This was performed based on the assumption that these areas have the greatest need for reservoirs to store water resources.

Within the framework of Model A, the generic Equation (1) was formulated in two different ways, generating models A1 and A2. Model A1 combines all the criteria presented in Table A1, with the exception of WB.

$$S_{su}={\displaystyle \frac{\left(Pr+T+LULC+CA+Sl\right)\ast USC}{5}}$$

(2)

where Pr, T, LULC, CA, Sl are the scores obtained by the single cell for each criterion; USC stands for Urban Soil Cover with USC = 0 for any urban area and USC = 1 for any other LULC. Model A2, on the other hand, combines all the criteria considered in Table A1:

$$S_{su}={\displaystyle \frac{\left(Pr+WB+T+LULC+CA+Sl\right)\ast USC}{6}}$$

(3)

As evident from Equations (2) and (3), both models make no distinction in terms of weight between the different criteria (i.e., w_i = 1/N for all criteria). The effect of the USC factor is to assign a suitability score of S_su = 0 to any portion of land that is urbanized or occupied by infrastructure, regardless of the values of other criteria. Conversely, in all other land conditions, this factor has no influence on the final suitability score.

2.3.3. Model B

Model B is based on the rational assumption that most existing reservoirs have been created in contexts where the combination of different key criteria is favourable. This model is inspired by a similar approach recently proposed and applied in Tuscany (central Italy) [29]. The following step-by-step approach was adopted to assign scores to the classes of the various quantitative (i.e., contributing area, slope, precipitation, simplified water balance) and qualitative (i.e., texture and LULC) criteria. To ensure methodological consistency, the same classes employed in Model A (Table A1) were used.

1.

Assignment of a unique class per criterion to each reservoir

Each reservoir was assigned a representative class c for each criterion, based on spatial overlays with the various criterion GIS layers. For slope, precipitation, and simplified water balance, the value assigned to each reservoir corresponds to the area-weighted average of the cell values from the respective raster layers within the reservoir’s surface. For the contributing area, the reservoir was assigned the value of the cell with the maximum value inside the reservoir’s extent, assuming it corresponds to the catchment outlet.

For qualitative criteria (texture and LULC), if a reservoir extended across multiple classes, it was assigned the dominant class (i.e., the one covering the largest proportion of the reservoir area).

2.

Computation of the weighted frequency (${WF}_c$) for each class c

$${WF}_c={\displaystyle \frac{f_{c,p}}{f_{c,R}}}$$

(4)

where f_c,p is the ratio of the areas of all reservoirs belonging to class c to the total reservoir area and f_c,R is the frequency of the same class c within the entire region. This ratio reflects how overrepresented or underrepresented a given class is within actual reservoir locations, relative to its occurrence in the whole region.

3.

Ranking and scoring

The computed ${WF}_c$ values were used to derive suitability scores, based on the assumption that higher ${WF}_c$ indicates greater suitability. For quantitative criteria, all classes were ordered by descending ${WF}_c$. Scores from 5 (most suitable) to 1 (least suitable) were then assigned following this ranking.

For qualitative criteria, ${WF}_c$ values were first converted into percentiles across the class distribution. The scores were assigned as follows: a score of 5 for WF_c values above the 81st percentile, a score of 4 for WF_c values between the 61st and 80th percentiles, a score of 3 for WF_c values between the 41st and 60th percentiles, a score of 2 for WF_c values between the 21st and 40th percentiles, and a score of 1 for WF_c values below the 20th percentile. As with Model A, the application of Model B was carried out by considering Equations (2) and (3), referring to the respective models as B1 and B2.

2.3.4. Suitability Evaluation Based on Sediment Yield (Model C)

Numerous models, varying in complexity, allow for the estimation of sediment yield at basin scale. This study used the Sediment Delivery Ratio model included in InVEST [20], a suite of models developed by Stanford University for the assessment of ecosystem services in a given territory. InVEST-SDR model was preferred over more complex tools such as SWAT [30] and SEDD [31] primarily due to its user-friendly graphical interface and ease of implementation, which make it particularly suitable for preliminary assessments and for use by practitioners without advanced modelling expertise. Details on inputs, the model, and procedures are briefly summarized hereafter.

For each pixel k of the watershed, the model initially estimates the annual soil loss A_k and subsequently calculates the sediment delivery ratio (SDR_k), representing the fraction of eroded soil that is effectively delivered to the stream network. The total sediment exported (t yr⁻¹) from the watershed under examination will therefore be given by the following:

$$Y={\displaystyle {\sum }_k}{SDR}_k{A}_k={\displaystyle {\sum }_k}{SDR}_k{\left(RKLSCP\right)}_k$$

(5)

where R (MJ mm h⁻¹ ha⁻¹), K (t ha h ha⁻¹ MJ⁻¹ mm⁻¹), L (unitless), S (unitless), C (unitless), and P (unitless) represent, respectively, the annual mean rainfall erosivity factor, the soil erodibility factor, the slope factor, the length factor, the cover-management factor and the support practice factor to of the RUSLE model [32]. SDR_k is estimated with the following equation [33]:

$${SDR}_k={\displaystyle \frac{{SDR}_{max}}{1+e^{\left(\frac{{IC}_0-{IC}_k}{k}\right)}}}$$

(6)

where K, IC₀, and SDR_max are parameters that should ideally be calibrated based on observations. In the absence of such data, default values suggested by InVEST were used—2, 0.5, and 0.8, respectively. IC_k is the Connectivity Index [34,35] of the cell k, computed as follows:

$${IC}_i={\mathit{log}}_{10}\left({\displaystyle \frac{\overline{C}\overline{S}\sqrt{A}}{\sum \frac{d_j}{C_j{S}_j}}}\right)$$

(7)

where $\overline{C}$ and $\overline{S}$ represent the average values of the RUSLE crop and slope factors for the upstream drainage area flowing into cell k, and A (m²) is the extent of that area. C_j and S_j represent the values of the RUSLE cover-management (C) and slope steepness (S) factors for each cell j that sediment must pass through after leaving cell k, following the steepest descent path toward the nearest segment of the hydrographic network. d_j is the length, in metres, of the hydraulic path within each downstream cell j.

In general, sediment production efficiency increases with topographic factors such as slope and in conditions where sediment encounters fewer transport barriers (e.g., lower surface roughness, less vegetation, etc.).

The input data required by the InVEST programme for this analysis are the DEM of the watershed, the RUSLE factor layers (rainfall erosivity factor R and soil erodibility factor K) and the LULC layer associated with a biophysical table containing the RUSLE C and P factor values for each land use type. A high-resolution spatial characterization of factor R in the region was recently carried out in [36]. However, for the other required factors, no specific high-resolution regional layers are available. Therefore, to ensure greater homogeneity of the layers used and facilitate the replicability of the method in other contexts, all the layers required for the application of RUSLE were sourced from the European Joint Research Centre (JRC) and are freely downloadable. In particular, the R and K layers, with a 500 m resolution, were obtained from data available on the ESDAC website [37,38]. For LULC, the CORINE layer (same of the suitability analysis) was used. The USLE C-factor values for each LULC class were assigned by overlaying the LULC map with the pan-European C-factor map [39], and calculating an area-weighted average of the C-factor for each LULC category. Finally, the layers of the watersheds on which to perform the sediment yield analysis must be specified. The following parameters were also specified in the InVEST simulation, choosing values recommended in the literature [20] in the absence of specifically calibrated ones: the maximum slope length factor L was set to 122 and the threshold flow accumulation, representing the minimum number of upstream pixels that must flow into the analyzed pixel for it to be classified as part of the hydrographic network, was set to 1000 pixels.

Since this process, along with subsequent SDR model computations, is quite time-consuming, the sediment yield analysis was conducted on a sample rather than on the entire set of 3206 reservoirs. The initial selection aimed to obtain a random sample of approximately 10% of the total population. Reservoirs were randomly drawn while ensuring that the sample reproduced the overall distribution of catchment areas in the full dataset, in terms of both range and average value. This approach was intended to preserve the statistical representativeness of the sample with respect to one of the most relevant factors affecting Y (i.e., catchment area).

Two additional reservoirs were manually included in the sample, as they are currently the subject of specific studies on sediment production (see Section 4). A further refinement of the sample was then carried out by excluding reservoirs for which it was not possible to clearly identify the outlet point of the contributing catchment—an essential requirement for the proper application of the InVEST-SDR model.

The final sample consists of 285 small reservoirs randomly selected across the entire region. Their catchment areas range from a minimum of 0.1 km² to a maximum of 4.7 km². The average catchment area is approximately 0.38 km², which is consistent with that of the full dataset. The selected reservoirs are distributed across a broad area of the region, ensuring that other relevant factors influencing Y (e.g., land use/land cover, slope, precipitation) are also adequately represented.

In order to compare the results of this evaluation (hereafter referred to as model C) with those of Models A and B, the 285 Y values obtained for the reservoir sample were transformed into percentiles and assigned corresponding scores from 5 to 1 with a step of 1. Obviously, from the perspective of sediment production, the situations with the lowest yields are to be considered highly suitable. Therefore, a score of 5 (high suitability) was assigned to Y values falling within the 0th to 20th percentiles, 4 to those between the 21st and 40th percentiles, and progressively lower scores down to 1 (low suitability) for Y values between the 81st and 100th percentiles.

2.3.5. Evaluation of Agreement Between Suitability Models

In order to evaluate the level of agreement between different SHRBS models, the weighted (K_W) version of Cohen’s Kappa coefficient was applied [40,41,42]. The weighted Kappa coefficient quantifies the level of agreement between two raters (or models) by taking into account not only whether their classifications match, but also the extent of their disagreement. This is particularly useful when the classification categories are ordinal. The calculation is based on a contingency table (or agreement matrix), where rows represent the classifications of one model and columns represent those of the other. Each cell in the matrix contains the frequency with which a given pair of ratings occurred. While the unweighted Kappa considers only exact matches along the main diagonal (perfect agreement), the weighted Kappa assigns a weight to each cell, reflecting the degree of disagreement between the corresponding categories. Cells closer to the diagonal (minor disagreements) are given higher weights (i.e., considered less severe), while cells further from the diagonal (larger disagreements) receive lower weights, penalizing more severe mismatches. Quadratic weights, commonly used in this context, increase the penalty exponentially with the distance from the diagonal.

In this study, the K_w coefficient was used to comparatively assess the level of agreement between models in evaluating the suitability of areas where existing reservoirs are located (the closer the value is to 1, the higher the level of agreement). Interpretative scales such as the one proposed by Landis and Koch [43] are available (

Table 1. Ranges considered for the interpretation of the Kappa statistic, K_w, and corresponding strength of agreement. Each level of agreement is associated with a background color, ranging from red (poor agreement) to blue (perfect agreement), to allow for a quicker assessment of the degree of agreement.

Kw	Agreement Level
<0.00	Poor
0.00–0.20	Slight
0.21–0.40	Fair
0.41–0.60	Moderate
0.61–0.80	Substantial
0.81–1.00	Almost perfect

).

3. Results

3.1. Definition of the Suitability Scores for Model B

First of all, the frequencies of the classes, c, established for each criterion were calculated for the cells across the entire regional territory (Table A1). The results are summarized in the frequency histograms shown in Figure 6. Besides being of descriptive interest, this information is essential for estimating f_c,R (Equation (4)), and thus assigning scores in model B.

Prior to analysis, the suitability scores obtained from the different methods were rounded to the nearest multiple of 0.5. The calculations were performed using the R software (version 4.4.1), specifically the ‘irr’ package [44].

Regarding texture (Figure 6a), the most common classes are clay-loam (69%), loam (17%), and silty-clay-loam (12%), with negligible or entirely absent percentages for the other classes. For the contributing area (Figure 6b), the vast majority of the territory falls within the <0.005 km² (86.6%) and 0.005–0.5 km² (12.4%) classes. Regarding slope (Figure 6c), the most represented class is above 7.5% (66%), consistent with the region’s predominantly hilly and mountainous nature. The frequency distribution for precipitation classes (Figure 6d) is more uniform than other criteria, except for the >1219 mm class, which accounts for only 4.2% of the territory. For LULC (Figure 6e), two predominant classes emerge (i.e., broad-leaved forest and non-irrigable arable land), together accounting for more than 60% of the territory. Finally, regarding the simplified water balance (Figure 6f), the highest frequency occurs in the <−200 mm class, gradually decreasing as the SWB value increases. The procedure outlined in Section 2.3.3 has made it possible to define the scoring scales of Model B, which are detailed in

Table A2. List of criteria considered in the WOP analysis and their suitability scores (Model B).

Criterion	Suitability Score
	5 (Highly Suitable)	4	3	2	1 (Low Suitability)
Texture, T	Silty Clay-Loam, Loam	Clay-Loam	Silty-Clay, Silt-Loam	-	Clay, Sandy-Clay-Loam, Sandy-Loam
Contributing Area, CA (km2)	0.5–5	5–50	>50	0.005–0.5	<0.005
Slope, Sl (%)	1.5–2.5	2.5–4.5	4.5–7.5	<1.5	>7.5
Precipitation, Pr (mm/year)	<883	833–962	962–1091	1091–1219	>1219
LULC	Inland marshes, Mineral extraction sites, Fruit trees and berry plantations, Water bodies, Non-irrigated arable land	Water courses, Land principally occupied by agriculture with significant areas of natural vegetation, Complex cultivation patterns, Vineyards	Mixed forest, Sclerophyllous vegetation, Pastures, Olive groves, Broad-leaved forest	Natural grasslands, Transitional woodland-shrub	Annual crops associated with permanent crops, Bare rocks, Beaches, dunes, sands, Burnt areas, Coniferous forest, Construction sites, Sparsely vegetated areas
Simplified Water Balance, WB (mm/year)	<−200	−200–0	200–400	0–200	>400

.

3.2. Suitability Maps Based on Model A and Model B

Before producing the SHRBS maps based on Models A and B, a preliminary analysis was carried out to assess the extent of the differences in the scores assigned to each criterion, according to the rating scales shown in Table A1 and Table A2. For this purpose, at the regional scale, the mean absolute differences between the raster layers scored according to Table A1 and Table A2 were calculated for each criterion. The results show mean absolute differences of 3.3, 2.1, 1.2, 0.7, 0.6, and 0.0 for texture, precipitation, contributing area, LULC, slope, and simplified water balance, respectively. This simple comparison clearly indicates that the differences in the suitability maps produced by Models A and B are mainly driven by variations in texture and precipitation scores.

The regional-scale SHRBS maps obtained by applying Model A with the formulations of Equations (2) and (3) are shown in Figure 7a and Figure 7b, respectively. The same type of information, but based on Model B, is shown in Figure 8.

The maps obtained using Model A (Figure 7) highlight a high spatial variability of suitability, with predominantly low to medium values. Higher values (with a greenish colouration) are observed near the plain areas (see for comparison Figure 1). The maps produced with Model B (Figure 8) also show high spatial variability. However, in this case, areas with medium-high suitability are more prevalent compared to what was observed with Model A (as expected, since this evaluation is based on existing reservoirs).

The pattern of these suitability maps appears to be well correlated with precipitation (Figure 2a) and evapotranspiration (Figure 2b) spatial variability. In general, for both Model A and Model B, the two formulations, Equations (1) and (2), show a fair agreement. A more detailed analysis of the results at the regional scale is presented in Figure 9, which shows the percentage of the regional territory falling into different suitability classes, depending on the model considered.

To evaluate the suitability of the areas where the existing reservoirs are located, the suitability maps (Figure 7 and Figure 8) were used to calculate the average pixel value within each reservoir in the available dataset. Figure 10 presents the percentage of reservoirs falling within each suitability range, according to the model used. As shown in Figure 10, the highest frequencies for all models are found in the suitability classes between 2 and 4. As expected, Model B yields higher suitability values than Model A. Additionally, for both models, the suitability scores are higher when using Equation (3) instead of Equation (2). The average suitability scores for the existing reservoirs are 2.54, 2.82, 3.22, and 3.38 for A1, A2, B1, and B2, respectively.

The degree of agreement among the different suitability models in evaluating the location of existing reservoirs was assessed using the K_W coefficient, and the results are presented in

Table 2. Coefficient of agreement K_W between suitability models. The background color represents the level of agreement based on the legend in Table 1.

	A1	A2	B1	B2
A1	-
A2	0.41	-
B1	0.03	0.13	-
B2	−0.03	0.10	0.67	-

. The colours correspond to the interpretation scale shown in Table 1.

The level of agreement between models of the same type (A1–A2 and B1–B2) ranges from moderate to substantial, while it is much lower for the cross-comparisons between A and B models.

3.3. Potential Sediment Production and Siltation Risk on Existing Reservoirs

The sediment yield (Y) values estimated using the model in Equation (5) for the sample of 285 reservoirs ranged from a minimum of 0.1 to a maximum of 623.5 t yr⁻¹ with a mean value of 63.9 t yr⁻¹ a median of 31.8 t yr⁻¹, and a standard deviation of 84.8 t yr⁻¹. The notable difference between the mean and the median was due to the presence of approximately ten outlier values, as identified through a boxplot analysis. No clear spatial patterns were observed in the distribution of sediment yield values among the 285 reservoirs. After expressing these values on a score scale ranging from 1 to 5 (i.e., 1 = low suitability due to a high risk of sedimentation, and 5 = high suitability due to a moderate or negligible risk of sedimentation), a comparison was made with the scores assigned to the same 285 reservoirs based on models A and B. To simplify the analysis—and considering the substantial agreement (Table 2) between the evaluations obtained from the two formulations of Equations (2) and (3)—the average suitability scores of the respective sub-models were considered for both model A and model B (i.e., A = (A1 + A2)/2 and B = (B1 + B2)/2). The comparison of the scores from Model A and Model B with those derived from the criterion based solely on the sediment yield (Model C) showed K_w values of −0.12 and −0.03, respectively. Therefore, the agreement between Model A and Model C, as well as between Model B and Model C, is notably weak.

4. Discussion

The initial considerations regarding the results obtained concern the reliability and overall trustworthiness of the procedure used to assess SHRBS based on Models A and B in the region. As previously explained, these models are based on criteria commonly found in the existing literature on the subject. In the case of Model A, even the scoring scales for each criterion follow literature-based guidelines, whereas in Model B, the scores are calibrated according to the frequency with which different value classes appear in the regional reservoir database. The two models tend to differ primarily in the scores assigned to soil texture and precipitation. However, differences are also observed for the other criteria, with the sole exception of the simplified water balance (Table A1 and Table A2). For example, for the contributing area (CA), Model B tends to favour medium-to-large sizes compared to Model A (specifically, the 0.005–0.5 class shifts from a score of 5 in Model A to a score of 2 in Model B). For precipitation (Pr), Model B shows scores with a perfect inverse correlation to the annual rainfall amounts. This result lends itself to an interesting interpretation. Based on the location of existing reservoirs (Model B), the need for reservoirs appears to be particularly pressing in areas with lower rainfall. While this is entirely reasonable, it may also result in lower storage capacities (an aspect addressed by the scoring scale used for Pr in Model A, where areas with average rainfall are given the highest suitability scores). On the other hand, Model B tends to assign high suitability to areas with medium-to-large contributing basins, which can ensure higher runoff volumes. This appears to be a strategic approach in the siting of existing reservoirs, aimed at compensating for the expected lower rainfall inputs.

Textures such as silty clay loam and loam shift from a score of 1 in Model A to a score of 5 in Model B, while clay loam shifts from 1 to 4. This outcome is likely linked to the dominant textural classes in the region (Figure 6a), which makes the presence of reservoirs on clay loam and silty clay loam soils almost inevitable. However, it is worth noting that the classification of these textures in the low-suitability category in Model A (as derived from [9]) is somewhat questionable. Finally, the differences between Model A and B in terms of LULC, though present, are difficult to interpret due to the high number of categories present.

Overall, despite various weaknesses and uncertainties—which will be discussed in more detail later—it can be said that the procedure demonstrates a certain degree of reliability. In fact, the suitability scores assigned to existing reservoirs using Model A show a slight degree of agreement (Table 2) with those obtained using Model B. Further confirmation comes from the comparison between Figure 9 (frequency distribution of suitability scores across the regional territory) and Figure 10 (frequency distribution of suitability scores for existing reservoirs). As shown, for all models, the frequency distributions in Figure 10 are shifted to the right (indicating higher suitability) compared to those in Figure 9. This suggests that, as one would indeed expect, both Model A and Model B indicate that existing reservoirs were built in territorial contexts more favourable than the regional average.

The suitability values obtained from existing reservoirs based on Model A, which are 2.54 (Model A1) and 2.82 (Model A2), are lower than the average score of the considered scale (i.e., 3 out of 5). This seemingly unsatisfactory result should be interpreted in light of several considerations. First, it should be noted that in a multicriteria analysis with numerous criteria, the maximum score (5 in this case) should be considered almost theoretical [26]. A proof of this is that even with Model B (optimized for the existing reservoirs), maximum scores do not exceed 4.4 (Figure 10).

This study, which has an exploratory character, includes several aspects of both the methodology and the data used that can certainly be further refined. Regarding the methodology, it is known that multicriteria analyses can be characterized by a high degree of subjectivity in the way scoring scales and weights are assigned [45]. Even when employing rational considerations (such as those illustrated in Section 2.3.2), subjectivity remains. For example, for soil texture, apart from the textural categories in the classes with a predominant clay or sandy component (which logically fall at the two extremes of the suitability scoring scale), for all others, it is not easy to identify a ranking. The attribution of differentiated weights was avoided in this primarily exploratory study to prevent the introduction of subjective bias in the absence of empirical evidence. However, we acknowledge that certain criteria may play a more decisive role in determining the actual suitability for SHR implementation. A promising direction for future research could involve analyzing a representative sample of existing SHRs—especially those deemed successful in supporting or maintaining agricultural activity—in order to derive more accurate weights. This could be achieved through statistical analysis or machine learning techniques [46], provided that sufficiently detailed data on SHR performance are available.

Sources of uncertainty can be found in the characteristics of the geospatial layers used. For soils, for instance, the spatial resolution of 500 m, while adequate for the purpose of characterizing SHRBS at a regional scale, may have led to incorrect suitability score assessments during the evaluation of individual reservoirs. It is well known that properties such as soil texture can vary significantly over distances much smaller than 500 m. In numerical terms, an incorrect assessment of soil texture could impact the overall suitability score calculated using Equations (2) or (3) by as much as ±0.8 points. For LULC, a potential source of error, again during the evaluation of existing reservoirs, could stem from land use and land cover changes that occurred between the time the reservoirs were created and the time the mapping used refers to.

Other limitations arise from the characteristics of the available reservoir database, which, as explained, does not specifically refer to SHRs but includes all water bodies, including those in plains. For these, some criteria and scoring scales remain valid, but for others (e.g., CA and Sl), low scores are expected, thereby lowering the overall average.

Finally, the study considered only physical criteria. As demonstrated by numerous studies (e.g., [47], social, economic, and ecological factors can be equally crucial to the success of SHRs. As an example, an area classified as suitable or highly suitable (i.e., S_su ≥ 4) based on Equations (2) and (3) might in fact not be of interest for the construction of SHRs due to the presence of efficient and widespread irrigation distribution networks. Similarly, in some cases, it might be economically and technically more feasible to develop pipeline systems from existing reservoirs rather than to construct a new one. Incorporating socio-economic or infrastructural factors into the suitability analysis is beyond the scope of the present study and would require more advanced MCDM approaches—such as the Analytic Hierarchy Process (AHP) or the Analytic Network Process (ANP). These methods allow for the structured integration of both quantitative and qualitative criteria and can accommodate stakeholder input to derive context-specific weighting schemes [12,27].

Another key point concerns the approach underlying Model B. Assigning suitability scores based on how frequently certain criteria occur in existing reservoirs is certainly innovative and helps reduce subjectivity. However, this method may sometimes lead to distortions, especially in cases of small sample sizes or when certain classes are over- or underrepresented. For instance, the “clay” soil texture, which is generally considered highly suitable for SHRs, appears as unsuitable (Table A2) simply because very few existing SHRs are located on clay soils due to their limited presence in the region (Table A2). Therefore, for the practical application of Model B, such distortions should be identified and corrected.

The comparison of the evaluation of existing reservoirs based on Model A (or B) with that of Model C clearly shows that the most common multicriteria methodologies fail to adequately account for the criterion of limited sediment production. In other words, the application of Model A (or B) could lead, in many cases, to classifying as suitable or highly suitable territorial contexts where high sediment production would result in rapid sedimentation of the reservoir. This result is partly expected, as both Model A and Model B are mainly focused on the conditions at the site where the reservoir is to be located, whereas sediment formation is influenced by the conditions and factors affecting the contributing catchments (e.g., slope and LULC). Moreover, the evaluation of factors such as LULC and slope is primarily oriented toward the technical and economic aspects of constructing the structure, rather than the hydrological processes involved. On the other hand, a suitability analysis capable of integrating assessments related both to the construction site and to the upstream catchment would require computationally intensive modelling, which is hardly compatible with the broad-scale and rapid nature of WOP analyses.

The assessment carried out within the framework of Model C is also subject to relevant uncertainties, as no calibration of the sediment production estimation model was performed. However, it can be noted that the sediment yield values produced by the model were not used in absolute terms. Instead, they were transformed into a qualitative score ranging from 1 to 5, based on their percentile value. Furthermore, the use of a large and diverse sample of reservoirs contributes to mitigating potential biases introduced by default parameter values. While individual estimates may be uncertain, the broader dataset allows for more reliable comparative insights. Together, these elements—qualitative scoring and a large sample—help ensure that the findings remain meaningful in an exploratory and planning-oriented context, even in the absence of detailed model calibration. To complement this, a more concrete investigation was carried out considering the specific case of two SHRs named Spina and Martin Pescatore, where investigations on the sedimentation level are being conducted about 50 years after their construction. These two reservoirs are among the case studies included in the SIGHTING project (see funding section), which aims, among other objectives, to calibrate and test models (in addition to InVEST, SWAT, and SEDD) for estimating the sedimentation levels in SHRs. The project also seeks to identify simplified and rapid assessment procedures based on climatic, land use, and topographic characteristics. However, the results are still preliminary due to the lack of sufficient calibration data; therefore, only the initial qualitative and quantitative findings from field surveys are referenced here. The investigations were carried out both through visual surveys (during periods of minimum water levels) and bathymetric surveys. For Spina, the investigation revealed a high level of sedimentation, with a loss of the usable water volume certainly exceeding 50% of the initial volume. For Martin Pescatore, however, the bathymetric data still show a volume close to the maximum. The two reservoirs are located about 2 km apart and are therefore subject to the same climatic conditions. The contributing basin of Spina is about 2.3 km², with land use classified as non-irrigated arable land, while for Martin Pescatore, the basin is about 0.33 km², predominantly covered by broad-leaved forest. The sediment production simulation using Model C indicates approximately 386 and 7 t yr⁻¹ for Spina and Martin Pescatore, respectively, corresponding to suitability scores (S_su) of 1 and 5 (

Table 3. Suitability scores for the location of two reservoirs based on different evaluation models.

	Suitability Score
Reservoir	Model A1	Model A2	Model B1	Model B2	Model C
Martin Pescatore	2.20	2.66	3.41	3.68	5
Spina	2.34	2.79	3.95	4.13	1

).

The evaluation of these reservoirs based on Models A, B, and C is shown in Table 3.

As seen in Table 3, based on Model A, the suitability is around average for both reservoirs, while it is medium-high according to Model B. Spina scores are slightly higher than Martin Pescatore mainly due to the LULC of the contributing basin, which, for both Models A and B, is considered more suitable compared to Martin Pescatore’s. On the other hand, in terms of sediment production (Model C), the Spina Pond is found to be very unsuitable. The high sediment production results from a number of factors, but it is primarily due to the land use, which leads to USLE C-factor values about 15 times higher in Spina Pond compared to Martin Pescatore.

5. Conclusions

In light of the results and evaluations carried out, it is evident that the methodologies and criteria typically used for SHRBS studies are valid and sufficiently indicative for preliminary planning studies at the regional spatial scale. It is worth emphasizing that this result was achieved using freely available georeferenced data layers and open-source software, ensuring that the analysis is fully replicable and efficient. Analyzing the locations of existing reservoirs, as achieved using Model B in this study, is helpful in more realistically and objectively defining the scoring scales to be assigned to the various criteria. However, it was also noted that this method may lead to biassed evaluations when applied to limited datasets or in cases where certain criteria are either too rare or overly common across the territory under assessment. The actual suitability of specific, localized areas should then be assessed through more detailed studies, where differentiated weights for the criteria may also be applied according to specific needs and priorities.

From the perspective of sedimentation risk, it was observed that the set of criteria typically considered in SHRBS analyses does not offer much protection, potentially identifying as suitable or highly suitable areas favourable to high sediment production and, therefore, to a rapid loss of usable reservoir volumes.

Besides using specific models for estimating sediment yield, a practical approach could involve extrapolating suitability rules from existing reservoirs that have maintained a significant portion of their original capacity over time.

Practitioners may consider adopting or adapting Model C as a complementary tool, especially during early screening phases. Since Model C is based on an inverse relationship between suitability and estimated sediment yield, it can help flag areas at higher risk of sedimentation. In practical terms, good preliminary results can already be achieved by integrating basic erosion risk maps—such as those derived from empirical models like USLE or RUSLE—into standard suitability analyses. These maps, often readily available or easy to generate for many regions [48], offer a quick and effective way to account for erosion potential at early planning stages.

The full development and integration of a sedimentation-based model like Model C, which is more demanding in terms of modelling effort and data requirements, could then be focused on priority areas pre-identified using such simplified tools. It may also be integrated into existing Weighted Overlay Process (WOP) frameworks either as an independent layer or through composite indicators designed to reflect long-term reservoir sustainability.

Furthermore, the approach presented here could be transferable to other regions with comparable physiographic, climatic, or land use characteristics, particularly in Mediterranean or semi-arid environments where sedimentation poses similar challenges to water storage infrastructure.

This primarily exploratory study can be refined in future research through the use of observed datasets, particularly those containing measured sediment accumulation levels in reservoirs over time. Future work should also consider dynamic sedimentation models that account for temporal land use and land cover changes, as well as field-based validation efforts to calibrate and enhance the reliability of the proposed methods.

In addition, the WOP methodology itself should be further refined—for example, by applying appropriate techniques for sensitivity analysis and for the assignment of weights to the different criteria and by integrating relevant socio-economic factors into the decision-making framework.