In many traditional conservation oriented spatial analyses, conservationists have focused on mapping and preserving those areas that represent the most significant ecological value while at the same time isolating those areas of human activity (Hersperger et al., 2018; Jalkanen et al., 2020). While such an approach is useful to identify ecological priority areas, it generally provides little practical direction in terms of land-use management for those landscapes that are dominated by a mix of both settlement, agricultural production areas, and infrastructure. The limitations of such a traditional conservation orientation are clearly demonstrated in the Mediterranean Region where long standing and highly intense uses by humans of large areas of land create conditions which make rigid conservation boundary definitions very difficult to apply in land-use planning.

To confront this issue, this research develops a new framework called the dual logic spatial zoning model (DLSZM), conceptualises land management as a multi-zone system that reflects varying levels of planning intervention, rather than framing space through a simple protect-versus-use perspective. Each zone represents a different level of planning intervention, and the zoning structure is based upon four main planning zones: Strict Conservation; Managed Use; Restoration; and Development Guidance, with an additional general management background class where appropriate. The underlying assumption is that sustainable spatial planning is more likely to emerge from the combined interpretation of ecological value, human demand and benefits, and stress and vulnerability, instead of relying on a single dominant planning priority.

The “dual-logic” design enables comparison and joint interpretation of two decision-generation pathways. The first pathway follows a rule-based, expert-driven zoning logic, widely used in spatial decision support and multicriteria evaluation (Malczewski and Jankowski, 2020). In this pathway, environmental variables are reclassified into a common ordinal scale and combined using expert-informed weights, producing zones aligned with explicit planning principles. The second pathway uses the same spatial predictors as continuous raster surfaces and applies a machine-learning model to learn zoning patterns from data. Here, the objective is not to produce an unconstrained “black-box” classification, but to examine whether the expert-defined zoning logic is learnable and to obtain a smoother, potentially more generalizable decision surface that is less dependent on predefined reclassification thresholds (Figure 1).

This study was conducted in Antalya Province, Türkiye, selected as an illustrative case where ecological sensitivities and human-driven spatial demands strongly intersect. Antalya metropolitan city is characterized by a combination of various forest and coastal ecosystem types in the same geographic space where there are both high levels of tourism and agriculture and urban development pressure (Cetin et al., 2022). The region experiences consistent demands for tourism and resultant growth in tourism-related infrastructure that continues to be built out into the coastal zone as well as at the coastal belt and in forest–coast transition zones (Sancar and Güngör, 2020; Sütçü et al., 2020). In this context, Antalya provides a useful setting to test multi-zone planning logic that navigates both conservation priorities and spatial development demands (Figure 2).

The analysis relied on spatial datasets encompassing remote sensing products, LULC layers, infrastructure and socio-economic proxies, and environmental disturbance records. Because inputs were heterogeneous in resolution and data type, all layers were harmonized to a 30 m analysis grid. Categorical datasets were resampled using nearest-neighbor interpolation, while continuous variables were resampled using bilinear interpolation to minimize artificial class mixing and preserve value continuity. All rasters were projected into a common coordinate reference system and aligned to a single reference grid. Analysis was restricted to pixels with valid data across all predictors; pixels with missing values in any layer were excluded. All spatial datasets were accessed between January and March 2025 from their respective official repositories (URLs provided in Table 1). Interpolation-related uncertainty was assessed qualitatively by comparing bilinear and nearest-neighbor outputs for representative variables, confirming that resampling did not introduce systematic spatial artefacts or class mixing beyond acceptable limits for regional-scale zoning analysis.

The predictor set was organized into three thematic components reflecting the coupled socio-ecological structure of Antalya: Ecosystem Services and Biodiversity (ESB), Human Use and Benefits (HUB), and Stress and Vulnerability (SV).

The ESB included indicators representing biodiversity patterns, habitat structure, and ecosystem functioning, such as species and endemic density, NDVI-based vegetation vigor, habitat quality and continuity metrics, hydrological proximity, and temperature-related proxies (e.g., LST) relevant to ecosystem functioning and regulating services.

Importantly, the same thematic structure was retained across both pathways, but variables entered the zoning process differently. In the expert-driven pathway, predictors were reclassified to a 1 to 5 ordinal scale to ensure comparability and interpretability. Most predictors were classified using their empirical distributions (e.g., natural breaks or quantiles), while selected pressure-related distance variables were grouped into literature- and practice-informed distance bands (e.g., proximity to roads, settlements, streams, wildfire, and hunting areas). In contrast, the machine-learning pathway used predictors in their continuous form (without ordinal reclassification) to reduce dependence on threshold choices and to allow the model to learn interactions in continuous feature space.

The expert-driven zoning pathway followed a multicriteria decision-making logic commonly used in spatial planning applications. Criterion importance was elicited from an expert panel (n = 7) spanning biogeography, forestry, urban and regional planning, public administration, and tourism. Experts expressed judgments using triangular fuzzy numbers (TFNs) to represent uncertainty (Zadeh, 1965; Deveci, 2023). Individual judgments were aggregated using the geometric mean to reduce the influence of extreme assessments.

The aggregated fuzzy weights were then defuzzified and transformed into crisp criterion weights using the Logarithmic Methodology of Additive Weights (Log LMAW) (Pamučar et al., 2021), applying a weighted-centre approach commonly used for TFN defuzzification (Opricovic and Tzeng, 2004; Aytekin et al., 2023). Using these weights, ESB, HUB, and SV component indices were calculated and subsequently translated into multi-zone outputs through explicit rule-based thresholds designed to reflect planning-relevant intervention levels rather than mathematically sharp boundaries. In order to assess robustness of expert weighting, a leave-one-expert-out (LOEO) sensitivity analysis was performed by recalculating Log-LMAW weights while excluding each expert in turn. Weight stability was quantified using coefficient of variation and rank consistency (Spearman’s ρ).

Since the zoning classes are grounded in planning logic rather than direct observations, model evaluation emphasised spatial consistency and coherence in decision logic, rather than relying solely on absolute validation. First, agreement between machine-learning zoning and expert-driven zoning was assessed using a zone-by-zone confusion matrix and class-based precision, recall, and F1-scores, emphasizing macro-averaged metrics to mitigate class imbalance effects. Second, zoning outputs were examined against proxy layers for natural areas (e.g., forest and semi-natural covers) and human pressure indicators to test whether zones exhibit expected spatial behavior (e.g., strict zones aligning with natural proxies; development guidance aligning with accessibility and built-up proxies). Finally, stability across spatial partitions was examined by comparing zoning patterns across blocks, providing an additional lens on spatial generalizability and sensitivity. Lower precision but higher recall observed for Development Guidance reflects its intentionally inclusive planning role, prioritizing sensitivity to development pressure over strict exclusion, consistent with its function as a guidance rather than regulatory zone.

A comparison of the large scale spatial zoning outputs produced using both the expert-based framework and the machine learning-based framework shows a number of similarities in terms of spatial extent, but also clear local variations in the zoning classifications produced (Figure 3). In both models, Managed Use is identified as the dominant zoning classification type, which are found to be extensive and relatively continuous spatial patterns across the interior parts of the study area. These zones occupy landscapes characterised by moderate ecological value combined with manageable levels of human pressure.

Strict Conservation zones primarily exist within ecologically intact areas of high habitat continuity and vegetation productivity. Expert-based framework tends to produce compact and spatially coherent core areas, while the ML-based framework tends to fragment these areas into smaller patches, especially along transitional land types. The fragmentation is not random, rather it indicates that the ML-based framework has greater sensitivity than an expert-based framework to finer spatial variations in overlapping environmental gradients.

Development Guidance zones are largely concentrated in coastal belts, on the edges of urban areas, and along key accessibility corridors. Both of the zoning frameworks show that these zones are geographically fragmented and closely aligned with current settlement structures, transport networks, and tourism-oriented land-use patterns. Restoration zones occur in a scattered distribution pattern, often corresponding to areas experiencing cumulative stress signals. In comparison to zoning developed by experts, the zoning derived from the ML model allocates significantly more area to the Restoration zone, which indicates an increased focus on signal of degradation and vulnerability.

Quantitative comparison of the outcomes of zoning shows that the spatial patterns produced from both the expert-derived method and the ML-based method show a high degree of similarity to one another across the area-weighted confusion matrix (Figure 4). The two methods produce similar classifications in managed-use zones in approximately 7,630 km² of area.

These zones are located where there is an intermediate level of ecological value and human utilization, and therefore can be used for ongoing land use but subject to specific managerial controls. These zones typically occur in transitional areas or “transition belts” between strictly protected cores and areas of the landscape designated to provide guidance for future development. Therefore, they act as buffers to protect sensitive ecosystems from excessive pressures of human activities while providing some form of social and economic functionality. A significant amount of agreement is also seen in the designation of Strict Conservation and Restoration zones by both methods of zoning, which demonstrates the degree of concurrence in identification of core priority conservation areas.

Although the expert framework and ML derived framework have a good agreement, there is some clear evidence of major transitions between the different zoning categories. Some areas that were classified as Strict Conservation by the expert-based method are categorized as Development Guidance or Restoration by the ML results. Also, there is evidence that the ML has assigned different classifications to those lands in the Development Guidance zones identified by the expert-based methods. These diagonal transitions suggest that the ML has systematically reassigned many of the same landscape features rather than made arbitrary disagreements about classifications.

Spatially, there are large, contiguous clusters of areas in which both frameworks agree; these are mainly located in the interior parts of the region that have been designated as being for Managed Use. Conversely, areas of disagreement are generally confined to coastal frontages, areas of urban expansion and ecological transition zones that coincide with intersections of biodiversity value, accessibility and vulnerability signal values. The spatial pattern of this variation implies that the main divergence between the two frameworks relates to landscapes that contain multiple conflicting signals and cumulative pressures from different types of planning activity.

The way different types of land cover are distributed through the zoning categories demonstrates how the expert-based framework and the ML-based framework have differing structural approaches to the distribution of natural and artificial surface types when they are combined (Figure 5). In the expert-based framework, most of the natural land-cover types were found in the Strict Conservation zone and the Managed Use zone, which together include the majority of the area of natural land-cover types (Figure 5A). The restoration zone includes a much smaller proportion than this because it was designed to be specifically designated as a target in the rule-based planning process. On the other hand, the artificial land-cover types were primarily located in the Managed Use zone and the Development Guidance zone; these correspond to a plan which emphasizes controlled or regulated use and controlled development in human dominated landscapes (Figure 5B). The separation of natural from artificial land-cover types is based upon the LULC classification scheme described in Section 2.4. Forest, shrubland, semi-natural vegetation, and wetlands are classified as natural cover while urban, industrial, transportation, and agricultural surfaces are classified as artificial cover.

Under the ML-framework derived, there are noticeable shifts to these distributions for each of the two land-cover categories. The largest percentage of natural area still falls into the Managed Use category; however, a significantly higher percentage of the natural land cover is classified as Restoration (Figure 5A) than under the expert-defined zoning. Similarly, when comparing the zoning classifications for artificial surface land covers generated by the ML-methodology to those produced by the expert methodology, the zoning generated by the ML-approach classifies a greater portion of the artificial surface land covers as Restoration (Figure 5B) than did the expert method. These redistributions indicate that the ML framework more commonly recognizes both natural and artificial landscapes as being in need of proactive management due to signals of overlap with regard to stress, vulnerability, and disturbances.

These patterns show collectively that the ML-derived zoning describes the land-cover types based on what is currently ecologically or functionally active, as well as their expose to cumulative pressures. This results in a greater prominence for restoration focused zoning across natural and modified environments, reflecting a more intervention oriented spatial interpretation when comparing to the expert-based framework.

The spatial configuration of zoning outcomes relates directly to the distribution and interaction of the three thematic elements: Ecosystem State (ESB), Human Use and Pressure (HUB), and System Vulnerability (SV). Understanding each of these thematic elements separately provides insight into the spatial reasoning underlying both expert-based and machine learning based zoning decision making processes (Figure 6).

Feature importance and explainability analysis in expert-based and machine learning (ML) frameworks were conducted to elucidate how zoning outcomes arise. The purpose of these analyses was to provide two complementary perspectives on decision drivers; i.e., normative expert prioritization of decision drivers versus data-driven learning outcomes (Figures 8–10).

The weighting of the criteria in the expert-based framework using the TFN-Log LMAW method clearly show the ESB component has a large weight given to ecological integrity. Habitat continuity, habitat quality and vegetation productivity are the variables with the largest weightings and represent an ecosystem functionally operating as it should. On the other hand, the HUB variable have the majority of their weightings applied to the HUB variables. In addition, the indicators for disturbance were the indicators associated with disturbances which had the largest weightings out of all of the indicators in the SV component; these being forest fire recurrence and climate stress.

Similar to global feature importance based on the ML-based CatBoost model, the most important drivers are represented similarly, however relative weights given to each component are redistributed with respect to the non-linear interactions that exist among predictors (Figure 9). The ESB variables will continue to be important in terms of how zoning is developed as an environmental protection vehicle. However, the ML method provides more weight to cumulative/interacting stressors, such as access, frequency of land use changes, and disturbance regimes, and therefore the methods provide a more balanced representation of the stressors related to accessibility and the drivers.

Further detail is provided on how each variable influences zoning decisions when individual variables combine in the context of different management regimes through class-specific SHAP analysis (Figure 10). For the Strict Conservation type, the most important positively contributing factors were the state of the ecosystems, including habitat continuity, habitat quality and vegetation production, whereas the human related factors showed negatively for all the factors considered. On the other hand, the results of the development guidance type were mainly explained by accessibility factors, land use changes, disturbance factors, and the other factors that were characterized as having moderate ecological value. Therefore, these results show how zoning decisions result from different combinations of ecological conditions and human pressure for different management approaches.

The spatial intensity of divergence between expert-based and ML-based zoning frameworks is illustrated through a continuous difference surface and direct map comparison (Figures 11, 12). The difference surface reveals that discrepancies are not randomly distributed, but instead form spatially coherent clusters across the landscape.

Most of the positive differences occurred within transitional environments that have a combination of various stressors (i.e., land use change, frequent fire activity, and gradient of accessibility) that are changing at an overlapping rate. Within these transitional areas, the ML-derived framework has assigned greater support for Restoration or Development Guidance more frequently than it did when the expert-based framework was used. Most of the negative differences were found in relatively intact ecological “interiors” where the expert-based framework most consistently supported either a decision to prioritize a Strict Conservation outcome or a Managed Use outcome.

The majority of study area exhibits minor difference in zoning, showing that overall there is a large scale convergence in how the two zoning methods are being applied. Therefore, most divergence will be found at the boundary areas where ecological values, vulnerabilities, and pressures from humans overlap. The patterns show that the major difference between expert based and machine learning (ML) based zoning occur because of structured decision logic versus inconsistent application of methodologies.

Comparing expert-driven to ML-based zoning outputs shows the variation in spatial prioritisation should be interpreted primarily as an outcome of their differing views of decision-making, rather than as artefacts of methodological inconsistency. Both of the two approaches to zoning use the same understanding of what constitutes ecological value and human pressure but operationalize them in different ways. These differences become particularly apparent in landscapes characterised by gradual transitions and overlapping planning signals, rather than in areas with clearly dominant priorities.

Expert driven zoning is based on an expert based cautionary approach to land use planning that emphasizes ecological integrity, habitat continuities, and limited disturbances by way of pre-defined rules and weights assigned to each rule. This type of approach is consistent with established principles of systematic conservation planning in which expert judgement is used to mitigate uncertainty in order to protect valuable systems from irreparable loss or degradation (Margules and Pressey, 2000; Pressey et al., 2013). In the present case, this logic results in spatially coherent zoning patterns, especially within ecologically intact interior areas where conservation priorities are relatively clear.

While ML-based zoning exhibits a greater ability to respond to gradual transitions, as well as overlapping pressures throughout the landscape, rather than using specific thresholds, it incorporates interactions among several spatial factors, including accessibility, land-use dynamics and disturbance-related signals. As mentioned earlier, similar behavior has also been documented in prior research, indicating that ML models generally exhibit a higher level of sensitivity when assessing heterogeneous or transitional environments (Verburg et al., 2016; Hagenauer and Helbich, 2017). This level of sensitivity, in this study, can be seen in the increased number of Restoration and Development Guidance zones being defined in the areas experiencing active change, and especially in the coastal and peri-urban corridors where a combination of pressures are present; in the case of Antalya, this pattern is clearly linked to tourism driven coastal development, coastal accessibility gradients, and the rapidly changing nature of forest-urban transition zones.

While a large portion of the study area shows zoning consistency across both methods, other areas show some degree of zoning divergence. These areas are typically characterised by well-defined ecological conditions and relatively stable levels of human pressure, suggesting that convergence between expert-based and ML-based zoning may point to robust and reliable planning signals. In contrast, areas which show a larger degree of zoning divergence between the two methods are not randomly distributed. They tend to coincide with landscapes where ecological value, vulnerability and human demand intersect, indicating locations where planning decisions are inherently more contested and context dependent (Knight et al., 2006).

From this point of view, the divergence in zoning approaches should not be seen as a limitations, but rather, it provides a valuable insight into how sensitive zoning results are to the underlying assumptions and the modelling selections made and where further evaluation and/or stakeholders and adaptive management strategies could possibly be required.

Understanding why there are differences in zoning results is as much of the planning challenge as determining where they occur. As such, the study has used both global importance metrics and class-oriented interpretability analysis to determine the factors that influence zoning decisions.

Results of CatBoost modeling at the aggregated scale clearly show that there is a substantial difference between the variables that are most closely associated with conservation-related zoning, versus those which are most closely associated with development and/or restoration. The variables most closely associated with ecosystem condition appear to have greater influence on the distribution of Strict Conservation zones, while indicators related to human use intensity, accessibility and disturbance appear to have greater influence on Development Guidance and Restoration areas. This overall pattern broadly mirrors the weighting logic embedded in the expert-based framework, suggesting conceptual consistency rather than contradiction between expert judgement and data-driven learning (Elith et al., 2008; Breiman, 2001).

Nevertheless, aggregated importance scores do little to elucidate how each of the individual variables operate across the various spatial contexts in which they are applied; to identify this relationship and thus to overcome the limitations described above, interpretability analysis were employed to illustrate how the relative impact of each variable is affected by a variety of zoning decisions. These results show that an identical driver can be associated with different levels of development potential based upon varying environmental context. For instance, accessibility-related indicators tend to increase the likelihood of Development Guidance zoning in areas with moderate ecological value, while simultaneously reducing the probability of Strict Conservation designation in similar contexts.

Such context-dependent behaviour can’t be easily captured in a rule-based systems based on fixed thresholds by design. Data driven models that are applied to continuous feature spaces capture such interactions much more naturally than rule based systems do. By explicitly showing these interactions, analysis of the interpretability reduce the black box nature typically associated with machine learning and increase transparency and planning value of ML based zoning output.

Although there was a general consistency between zoning results generated through the use of expert-based and ML-based methods for land-use/zoning classification, several different areas of uncertainty will require special attention. The primary sources of these uncertainties will stem from the data-related constraints and also from the fact that all definitions of planning priorities have an inherent normative nature.

The use of expert based weighting schemes can also give a higher weighting to some ecological attributes whilst giving an under-weighting to cumulative or emerging threats. On the other hand, the zoning outputs from the application of machine learning (ML) algorithms will heavily depend on the quality of the data, spatial resolution and the adequacy of proxy variables used to represent human activity and disturbance (Pontius and Millones, 2011). Therefore, zoning outputs should be viewed in terms of supporting decision making as opposed to being used for prescriptive planning solutions.

From a practical planning perspective, one can say that areas with relative convergence between both methods are more robust in a relative sense, than areas of divergence, and thus areas would potentially benefit from targeted field surveys, and/or stakeholder engagement or adaptive management interventions. In this sense, disagreement between zoning approaches functions as a diagnostic signal highlighting planning sensitivity, rather than a methodological weakness.

In terms of a regulatory and policy perspective, the proposed zoning framework provides a basis for zoning and policy to support differentiated approaches as per the spatial context. This framework will be used to ensure that the current high levels of conservation remain and to limit further development of this land for infrastructure development. Managed Use zone will be used to justify adaptive zoning regulations which can facilitate the balancing of ecosystem conservation with the continued use of these lands for low-impact tourism, regulated forestry, and controlled agricultures. Development Guidance zones will assist in identifying areas in which there is an expectation that pressure for growth will increase and in which proactive planning tools (zoning regulation, development control and/or environmental impact assessment) will be most useful. Finally, Restoration Zones will be used to identify those areas of the landscape having the greatest need for ecological rehabilitation and risk reduction, particularly in areas that have been subject to recurring disturbances and/or changes in land use.

The transferability of the proposed framework should be treated with care. Although the general structure of the framework is generally applicable outside of the case study location, the zoning outcomes and decision logic will depend upon the local ecological and governance context. Therefore, in order for the framework to provide locally relevant results, both local calibration and expert input will continue to be necessary. Additionally, future research could improve this method through the incorporation of scenario-based analyses that specifically examine long-term environmental and socio-economic change.