Investigating potential supply of ecosystem services in cultural landscapes through efficiency analysis

To substantiate the proposed framework, the empirical part of this study was conducted in southwestern Slovenia, a region with a sub-Mediterranean climate that covers approximately 72,288 hectares and is home to about 30,000 inhabitants. It spans the Karst, Brkini, and Sub-Dinaric Littoral regions and includes five municipalities, namely Sežana, Komen, Divača, Hrpelje-Kozina, and Miren-Kostanjevica (Fig. 2). The economic development of the region was based on agriculture until the second half of the 19th century, followed by viticulture and stonemasonry after the railroad was built (Panjek 2015). After World War II, the level of education increased, and at the same time, the industrial sector developed, leaving traditional landscape features such as dry-stone walls, tiny shepherds’ houses, stone wells, and old fields, and succession began. People abandoned agriculture in search of better and easier incomes in the emerging automotive, timber, and logistics industries or other industries in larger cities such as Koper and Sežana. This transition also led to migration to larger cities, changing people’s lives and reorienting their needs and resources to other areas of activity.

The Natura 2000 network covers 70% of the study area, which is also considered one of the European biodiversity hotspots (Živi Kras-Carso 2014). The study area is in a hilly topography at an altitude between 200 and 800 m above sea level (Marušič et al. 1998). Forests cover 65% of the area, followed by meadows and pastures (20%) and old field succession (4%). Agricultural land occupies less than 3% of the area, urban and artificial land about 4% in traditionally scattered settlements. There is only less than 0.1% surface water, mainly due to the predominantly calcareous bedrock. Water is found mainly in underground caves, passages, and natural pools.

Traditionally, the study area was mainly used by the local residents for agricultural purposes, especially as pastureland. In the past, the forests were sparse and consisted mainly of deciduous species such as oaks, hornbeams, and ash trees. In fact, the grasslands that dominated the Karst (most of the northerly part of the study area) at the end of the 18th century were transformed mainly into forests over the next 250 years (Kaligarič and Ivajnšič 2014). This transition was made possible by the artificial introduction of a non-native Austrian pine (Pinus nigra Arnold) in the 19th century. This introduction, which was considered positive and beneficial by foresters and policymakers, led to prolonged resentment among local residents and landowners, who were deprived of their livelihoods as a result (Remec 2021). Eventually, local residents accepted the new species, which became an integral part of their identity. Nowadays, the Austrian pine is losing vitality, and more native species (e.g., Quercus petraea (Matt.) Liebl., Ostrya carpinifolia Scop.) are shaping the structure of forest undergrowth. The study area thus represents a traditional cultural landscape that is now returning to its natural state, making it an interesting subject for research. And since similar cultural landscapes can be found all over the world, we believe it is a suitable study region.

Based on the preliminary research conducted in the study area and after reviewing the existing literature (e.g., Panjek 2015; Marušič et al. 1998; Živi Kras-Carso 2014), seventeen ESs were selected from three main ES groups, namely regulating, cultural, and provisioning. The preliminary research encompasses face-to-face interviews with stakeholders in the study area prior to the ES mapping and assessment. The nineteen stakeholders interviewed were local residents, farmers, foresters, hunters, agricultural advisory officers, a municipal officer, a tourist guide, a cultural worker, and a student. The interviewees were asked, among other things, which ESs are considered important for most people in the study area. After analyzing the results, a list of identified ESs was created. The naming of ESs follows the CICES v5.1 classification on a group-, class- or class-type level. Table 1 shows the selected ESs, ES groups, short names, indicators, data sources, and method of mapping.

Based on the conceptual framework described above, square LUs (i.e., rectangular cells) with a side length of 1 km were chosen. The grid surrounding the study area was built using the Create Fishnet tool in ArcGIS PRO 3.1 (ESRI) software. LUs that were completely outside of the study area were excluded, as well as LUs with an area less than 0.999 km². These LUs were excluded to ensure comparability between LUs and avoid issues related to their size. The grid’s final composition consists of 627 cells. We used detailed LULC maps of the Slovenian Ministry of Agriculture, Forestry, and Food as a basis for assessing the potential supply of ESs and for visualizing the results. We employed a deterministic approach to map the potential supply of ESs through direct or indirect measurements, mainly using proxy indicators. The usage of LULC-based proxy indicators is considered as not preferred alternative, especially for small-scale analyses (Eigenbrod et al. 2010). On the other hand, when information on ecosystem conditions and direct supply of ESs are inexistent or insufficient, proxies are necessary to develop (Burkhard and Maes 2017).

Nevertheless, two ESs (i.e., landscape esthetics and tourism) were assessed by modeling with a modeling tool, and five ESs (i.e., climate regulation, flora observation, crop production, wild plant food, wood production) were assessed directly with the LULC layer and parameters provided by the authorities. The list of ESs, indicators, data sources, and methods is shown in Table 1. It is of utter importance to note that we opted for indicators that were spatially explicit and publicly available. For this reason, some indicators do not entirely match with the selected ESs name (e.g., flora observation, fauna observation). The reported mapped ESs were rescaled using min-max normalization between 0 and 100 to account for the different scales. This involved subtracting the minimum values from the dataset values and dividing the intermediate result by the difference between the maximum and minimum values of the dataset (Rolo et al. 2021). The final step was to multiply the result by 100 to obtain a range between 0 and 100. As noted by Paracchini and Capitani (2011), building composite indicators requires non-dimensional variables. In our study, we compare the weighted ES values with their corresponding costs using a complex procedure rather than simply adding up values. For technical reasons related to the limitation of the DEA software used (i.e., zero values could lead to infinite cycling of the software, yielding no results), we replaced zero values of inputs and outputs with the value 0.001 (Coelli et al. 2005).

Erosion control was mapped using the presence of erosion control measures published by the Slovenian Meteorological Agency. We assign values to LUs that are proportional to the area covered by these measures, assuming that ecosystems present in LUs with erosion control measures provide lower erosion control potential (Crossman et al. 2013). ES climate regulation was mapped as a proxy value of carbon stored in soils (Burkhard and Maes 2017; Wei et al. 2017). Each LULC was assigned a corresponding carbon storage value, and then a weighted arithmetic mean was calculated for each LU. Carbon values for LULCs with sources and calculation methods are listed in Table A.1 (see Appendices).

Educational ES has been mapped by the presence of protected natural areas (e.g., Natura 2000 sites, ecologically important areas). The larger the share of protected areas in a LU, the higher the potential supply of ESs on that LU, as these areas are important to promote environmental awareness through education (De Dominicis et al. 2017). Touristic ES has been modeled using the InVEST model Visitation Tourism (Baró et al. 2016, Outdoor Recreation Activities—Supply 2018, Sharp et al. 2021). The InVEST Visitation tourism model displays average visitation rates in different landscapes for the period based on Flickr photos with geotags that were taken in a specific area. Since we wanted to obtain a long-term average, we chose to use the maximum period, i.e., 2005–2017. ES flora observation was mapped as the diversity of tree species—the higher the number of different tree species in the forest, the higher the potential to provide this ES – aggregated for LUs. A similar indicator has been used for assessing biodiversity appreciation (e.g., Queiroz et al. 2015). Fauna observation ES was mapped by the presence of protected large carnivores (i.e., brown bear, lynx, and wolf). The larger the share of wildlife territories present in a LU, the higher the potential supply of ESs on that LU, as the animals can be seen in that area. Species richness (n ha⁻¹) is often used as an indicator (e.g., Crouzat et al. 2015), but due to a lack of data, we used the abundance of the territory per square kilometer. Inspirational ESs have been mapped based on the density of immobile cultural heritage sites. The idea is that a higher potential for providing ES lies in areas with a higher density of cultural heritage sites. Immobile cultural heritage sites (e.g., buildings and monuments) serve as inspiration for culture, art, and design (Berglihn and Gómez-Baggethun 2021). Spatial explicit maps with sites’ locations were appended to each LU to obtain the number of locations. The number of different wild animals that can be hunted found in the LU was used as a proxy for ES hunting. This indicates the distribution of wild animals and the diversity of hunting experience, as more species offer more opportunities for various hunters. Importantly, in this study, we consider ES hunting as a cultural ES and ES wild animals used for nutrition as a provisioning ES (Turner et al. 2014; Schulp et al. 2014). Spiritual ESs have been mapped based on the density of immaterial cultural heritage sites. Higher potential for providing ES lies in areas with a higher density of intangible cultural heritage, e.g., performing arts, traditional knowledge, practices, and rituals, which provide a foundation for spiritual experiences in nature (Berglihn and Gómez-Baggethun 2021). Recreational ES was evaluated using the recreational potential index, which has been described in detail by Paracchini et al. (2014) and Outdoor Recreation Activities—Supply (2018). The recreation potential index is calculated as the average normalized value of six indicators, namely recreation value of protected areas, degree of naturalness, distance to water bodies, number of LULC types, terrain roughness, and density of mountain peaks. Due to the low density of significant mountain peaks in the area, this indicator was replaced by the distance from the centroid of each LU to the nearest of the 48 mountain peaks. The only criterion for selecting the peaks was an elevation of more than 250 m above sea level. Esthetics ESs have been modeled using the InVEST model Scenic Quality (Baró et al. 2016, Outdoor Recreation Activities—Supply 2018, Sharp et al. 2021). For the InVEST Scenic Quality model, a digital elevation model and data from the Ministry of the Environment and Spatial Planning were used to identify features that negatively impact scenic quality (e.g., surface mines and industrial areas). The areas with unaffected views were considered as the areas that potentially supply the most ES landscape esthetic.

The soil index calculated by the Ministry of Agriculture, Forestry, and Food was used as a proxy for ES crops. The soil index can be interpreted as a “productivity index” as it indicates the productive capacity of the soil (Egoh Benis et al. 2012). The higher the values, the more productive the soil is and thus more suitable for agricultural production. The number of households collecting medicinal plants was used as a proxy for ES medicinal plants. We used Lovrić et al. (2020) estimate that 37.9% of Slovenian households collect wild medicinal and aromatic plants and calculated the number of households per settlement—the higher the number, the higher the potential supply of ES. Importantly, this indicator does not reflect the actual availability of certain plants in the ecosystems near the household. However, it is assumed that individuals are familiar with the flora in their immediate vicinity and therefore gather plants in close proximity to their residences. Water was mapped by calculating the nearest distance to an above-ground water body or registered water source. Proximity to water bodies is a proxy indicator of water yield; the closer the water body, the higher the potential supply of water (cf. Burkhard and Maes 2017; Vigerstol and Aukema 2011). The volume of chestnut in the forest stand was used as a proxy for assessing ES wild plants (Burkhard et al. 2014). Chestnut is the only species used as a proxy for plants in general, as no other flora (e.g., mushrooms) has been mapped or spatially assessed previously. The number of wild animals hunted was used as a proxy for ES wild animals. We used the Slovenia Forest Service data on a number of shot wild animals that are often used for nutrition per hunting district (Burkhard et al. 2014; Raudsepp-Hearne et al. 2010). Finally, forest growing stock in the forest stand was used as a proxy for assessing ES timber production (Burkhard et al. 2014). This indicator is related to forest areas only and shows the capacity of a forest stand to produce wood.

DEA was used to evaluate the efficiency of landscape units in supplying various ESs (i.e., outputs), given management costs (i.e., inputs). DEA is a popular nonparametric method for measuring decision-making unit (DMU) efficiency developed by Charnes et al. (1978). In this study, the concept of DMUs is equivalent to the concept of LUs. DEA allows the use of multiple inputs and multiple outputs with different units. This means that one can assess multiple economic, environmental, and social aspects without having to deal with units and data transformations (Kapfer et al. 2012). The more outputs produced with given inputs, the more efficient the DMU (note: in our study, we use LU as equivalent to DMU). DEA uses linear programming to calculate efficiency as the ratio between the weighted sum of outputs and the weighted sum of inputs and can be represented as described below (Coelli et al. 2005, Huguenin 2013). Suppose we have j different LU_j, j = 1,…, n and each LU_j requires m inputs I, i = 1,…, m, in order to produce s different outputs r, r = 1,…, s. A LU_j requires x_ij units of input i to produce y_rj units of output r. We assume that each LU_j has at least one positive input and one positive output, where x_ij and y_rj are bigger or equal to 0. We want the technical efficiency LU_j to be maximized (Eq. 1) as to other LU_j, j = 1, …, n. The difference between weighted outputs and weighted inputs is greater than or equal to zero for all LU_j (Eq. 2). The maximum technical efficiency value of each LU_j is 1 (Eq. 3), and the values of weights u_r and v_i must be greater than 0 (Eq. 4).subject to

In our study, we are interested in identifying the LUs with the highest potential supply of multiple ESs, given the estimated cost of management per LULC, so we opted for output orientation. Since the LUs have the same acreage and we assume they are operating at optimal scale, constant returns to scale technology were applied. The 17 ESs evaluated were used as outputs in the efficiency analysis. We calculated one input, i.e., the cost of managing the various LULCs within an LU, which is a proxy for obtaining the set of ESs in that specific LU (Table A.2 in the Appendices). Specifically, the associated management costs from the economic model calculations of the Agricultural Institute of Slovenia and the Slovenian Forestry Institute were assigned to the respective LULCs and averaged within each LU. Urban and artificial areas were excluded from the average cost calculations and were not included in the total area of LUs because they do not represent a natural or semi-natural ecosystem and, thus, do not provide ES for the purposes of this study.

The idea behind assigning the average management cost of producing products associated with a particular LULC composition to LUs as a proxy lies in the fact that the supply of various ESs is theoretically the result of independent LULC management for a particular use. Take pasture, for example: if a landowner manages it for sheep herding, that pasture is also a (potential) storage area for carbon in the soils, an area for observing and studying flora and fauna, a landscape of inspiration, and other benefits. The cost of “co-producing” all these ESs is therefore proportional to the management cost of producing the main products or services (e.g., grazing in the example from which dairy or meat products are produced). This is only true for a static temporal analysis with ESs that are strongly connected with particular LULCs. In a dynamic configuration and with other indicators, this assumption may not hold. In other words, the use of pastureland also reflects the potential contribution of that pastureland to the supply of various other ESs, including those indirectly associated with a particular ecosystem. Win4Deap v2.1 (^© 2015 Michel Deslierres) was used for the calculations of DEA efficiency. The result of our efforts is an LUs efficiency map that can help regional planners improve the potential supply of ESs in inefficient LUs by looking at the LULC composition of the most efficient LUs, corresponding ESs, and the associated costs.

For the sake of argument and to verify the spatial distribution of ESs on different groups of LULCs, the results were summarized and contrasted within landscape types, as suggested by Hladnik (2005). The typology chosen presents the results from a different perspective that takes into account terrain features (e.g., forests are primarily located on higher and steeper terrains, which are less suitable for agriculture; Hladnik 2005). He distinguished three landscape types: (a) forest landscapes are landscapes with a forest percentage of more than 85%, (b) forested landscapes are landscapes with a forest percentage between 40 and 85%, and (c) agricultural and urban landscapes are landscapes with a forest percentage of less than 40%. To test whether there are statistically significant differences between two or more groups (e.g., ES groups, LULCs, landscape types), we used the Kruskal-Wallis H test; to show the correlation between ESs, landscape types, and LULCs, we used Spearman’s rank correlation coefficient.

Finally, Tobit regression was used to assess the underlying drivers of change that affect the efficiency of LUs. Tobit regression is a special case of censored regression, in which the dependent variables are bounded from below, above, or both (Hoff 2007). We opted for a two-limit Tobit regression analysis since the efficiency scores were bounded between zero and one. As noted by Ramalho et al. (2010), Tobit and linear regression (ordinary least squares) models are traditionally used to regress efficiency analysis results. The Tobit model assumes the existence of a latent variable, y^*_i, which is unobservable and is hypothesized to underlie the observed dependent variable scores. As pointed out by Hoff (2007), the Tobit model is a misspecification for second-stage efficiency analysis with DEA, yet it is sufficient for modeling DEA efficiency values against independent variables. The DEA efficiency value is defined by several environmental variables (x_i) and a latent random variable (y^*_i), as shown in Eq. 5.where x_i is a 1 × k vector of observations on the constant and k – 1 efficiency factor explanatory variables, β is a k × 1 vector of unknown coefficients, and ε_i is the random error term that is normally distributed with mean 0 and variance σ² (McDonald 2009). If y^*_i ≤ 0, the efficiency score for the ith production unit, y_i = 0, if y^*_i ≥ 1, y_i = 1, and if 0 < y^*_i < 1, y_i = y^*_i

Using this latent variable y^*_i, the efficiency values y_i are defined as censored above 1 and below 0. This means that if y^*_i is greater than 0, we observe the actual value of y_i, and if y^*_i is less than or equal to 0, we observe a censored value of 0. Likewise, if y^*_i is greater or equal to 1, we observe the censored value of 1, otherwise, the actual value of y_i. Tobit regression models censored dependent variables by explicitly accounting for an unobservable latent variable that influences the observed outcomes, thereby providing insights into the factors contributing to the observed distribution. Tobit regression analysis was performed using Gretl 2022a software (^© Allin Cottrell and Riccardo “Jack” Lucchetti).

Based on the literature reviewed and survey results, we chose to analyze the impact of seven drivers of change that may influence the efficiency of LUs. The main criterion for the selection of driving factors was that the factor has already been recognized and described by the research community. The second important criterion for including variables was the ability to obtain spatially explicit data with some variation. The second criterion is a severe restriction that disables the inclusion of several interesting variables such as average income, education, life satisfaction, and community-level economic development indicators. We followed the classification of factors affecting landscape change proposed by Bürgi et al. (2004) and divided them into five groups: socioeconomic (population density, number of different LULCs), natural (average slope, average elevation), political (proportion of Natura 2000 area), technological (road density), and cultural (distance to cultural heritage sites). We hypothesized that lower population density would increase the efficiency of LUs by positively affecting the heterogeneity of LULC use and expressing synergies among ESs (Feng et al. 2021). Furthermore, we hypothesized that a higher proportion of Natura 2000 network, higher road density, higher number of different LULCs per square kilometer, greater distance from cultural heritage sites, greater slopes, and higher average elevation would also increase LUs efficiency (Schirpke et al. 2013; Paracchini et al. 2014; Allen 2015; Rolo et al. 2021; Sasaki et al. 2021; González-García et al. 2022). The spatially explicit slope and elevation data were extracted from a digital elevation model with a resolution of 10 m. The number of different LULCs was calculated from the spatially explicit LULC map provided by the Ministry of Agriculture, Forestry, and Food. Other maps were retrieved from national authorities: the Ministry of Environment and Spatial Planning for Natura 2000 maps, the Survey and Mapping Authority for road network maps, and the Ministry of Culture for cultural heritage sites. All data refer to the state in 2020.