Introduction



Sustainable ecotourism development requires the careful integration of ecological conservation, environmental capacity, spatial accessibility, and local development priorities. In environmentally sensitive and infrastructure-limited regions, tourism planning cannot be based only on the presence of natural and cultural attractions. It must also determine whether the landscape has sufficient ecological capability to support tourism activities without causing long-term environmental degradation. Poorly planned tourism development may increase pressure on natural resources, intensify habitat fragmentation, reduce landscape quality, and create conflicts between conservation and economic development. Therefore, ecological suitability assessment has become an essential tool for identifying areas where tourism activities can be developed in a controlled, sustainable, and environmentally responsible manner.



Khodabandeh County, located in Zanjan Province in northwestern Iran, has considerable potential for ecotourism because of its ecological diversity, topographic variation, natural attractions, and cultural heritage. Important attractions such as Katale Khor Cave, the Sojas River, Qeydar Nabi Tomb, and local cultural assets including Afshar carpet weaving, pottery, and metalwork provide a strong basis for tourism development. However, despite this potential, tourism development in the county has remained limited due to insufficient spatial planning, lack of comprehensive ecological suitability evaluation, and weak integration of environmental constraints into development strategies. The spatial distribution of elevation, slope, climate, land use, water resources, conservation zones, and tourist attractions creates a heterogeneous landscape in which ecotourism suitability is unevenly distributed.



This study evaluates the ecological suitability of Khodabandeh County for sustainable ecotourism development using an integrated spatial decision-support framework. The framework combines the Delphi technique, Analytic Network Process (ANP), fuzzy logic, Geographic Information Systems (GIS), and Random Forest (RF) modeling. The main objective is to identify suitable and restricted zones for ecotourism development and to compare expert-based and data-driven suitability outputs. By integrating expert judgment, fuzzy standardization, spatial multi-criteria analysis, and machine-learning-based comparison, the study provides a scientific basis for regional planning and sustainable tourism policy.



Materials and Methods



This research was designed as an applied and descriptive–analytical study. The methodological framework consisted of six main stages: spatial data collection and preprocessing, expert consultation using the Delphi method, weighting of ecological criteria through ANP, fuzzy standardization of environmental variables, GIS-based spatial integration using weighted linear combination, and Random Forest modeling for comparative spatial-consistency assessment. Six major ecological and infrastructural criteria were selected based on literature review, expert consultation, and available spatial datasets: water availability, climate conditions, landform characteristics, land use, conservation status, and accessibility to tourist attractions. These criteria were represented through subcriteria including distance to rivers, temperature, sunshine duration, elevation, slope, slope direction, land use/land cover, conservation areas, and proximity to tourist attractions.



The Delphi method was applied to refine and validate the selected criteria and subcriteria. A panel of experts with backgrounds in environmental planning, ecotourism development, GIS, and spatial analysis participated in the consultation process. Their judgments were used to structure the decision-making model and support the subsequent ANP weighting procedure. ANP was selected because ecological suitability assessment involves interdependent relationships among criteria rather than completely independent variables. Unlike hierarchical models, ANP allows feedback and interaction among criteria and subcriteria, making it more appropriate for complex socio-ecological systems. Pairwise comparison matrices were constructed using Saaty’s scale, and final weights were calculated in Super Decisions software. The consistency ratio was below the acceptable threshold, confirming the reliability of expert judgments.



Fuzzy logic was used to standardize heterogeneous environmental and infrastructural layers into a common suitability scale ranging from 0 to 1. This step was necessary because ecological variables usually affect suitability gradually rather than through fixed and abrupt thresholds. For example, slope, elevation, temperature, distance to rivers, and distance to tourist attractions influence tourism suitability through continuous spatial transitions. Fuzzy membership functions allowed these gradual changes to be represented more realistically. Areas closer to rivers and tourist attractions were assigned higher suitability values because of their greater recreational potential, accessibility, landscape attractiveness, and visitor experience value. Temperature and sunshine duration were standardized according to their suitability for outdoor recreational activities and visitor comfort. Land use and conservation zones were standardized based on ecological compatibility, tourism attractiveness, environmental sensitivity, and regulatory restrictions.



The fuzzy-standardized layers were integrated using a weighted linear combination model in GIS. In this process, ANP-derived weights were multiplied by their corresponding fuzzy membership layers and aggregated to produce the primary ecotourism suitability map. In parallel, a Random Forest machine-learning model was developed to generate a complementary data-driven suitability map and compare its spatial pattern with the ANP–fuzzy–GIS output. The RF model incorporated the same environmental and tourism-related predictor variables used in the ANP–fuzzy analysis. A total of 410 spatial reference points were prepared, including randomly distributed samples across the study area and location-based samples associated with known tourist attractions and relevant spatial features. Each point was assigned to one of five suitability classes: completely suitable, suitable, relatively suitable, unsuitable, and completely unsuitable. The dataset was divided into training and testing subsets, with 70% used for training and 30% used for testing. The RF model was configured with 290 decision trees, and the Gini impurity index was used as the splitting criterion.



Model performance was assessed using internal classification accuracy and variable importance values. In addition, the RF-derived map was compared with the ANP–fuzzy–GIS map through visual interpretation and pixel-based overlay analysis. Before comparison, both maps were harmonized in terms of projection, spatial resolution, processing extent, snap raster, and NoData mask. Since the RF sample labels were derived from suitability classes and spatial data may exhibit autocorrelation, RF accuracy was interpreted as an internal sample-level performance indicator rather than independent ground-truth validation. Therefore, the RF output was used primarily for comparative spatial assessment and consistency analysis.



Results



The ANP weighting results showed that proximity to tourist attractions received the highest final global weight, followed by temperature, sunshine duration, river proximity, elevation, slope, conservation areas, slope direction, and land use. These results indicate that ecotourism suitability in Khodabandeh County is strongly influenced by accessibility to tourism attractions, climatic comfort, and water-related landscape attractiveness. Land use and conservation areas received lower weights, not because they are unimportant, but because they function more as limiting or regulatory factors than as direct drivers of tourism attractiveness. This distinction between enabling and constraining factors is important for sustainable ecotourism planning, because areas with high recreational appeal may still require strict environmental management if they overlap with ecologically sensitive or restricted zones.



The final ANP–fuzzy–GIS suitability map showed clear spatial differentiation across Khodabandeh County. Completely suitable areas accounted for approximately 8% of the county and were mainly concentrated in the northern parts of the study area. Suitable areas covered about 26%, while relatively suitable areas represented the largest share, approximately 31% of the county. In contrast, unsuitable and completely unsuitable areas covered approximately 22% and 13%, respectively. Overall, around 65% of the county fell within the suitable and relatively suitable categories, indicating considerable potential for ecotourism development under appropriate planning and environmental management conditions. Higher-suitability zones were generally associated with favorable topography, better access to attractions and water resources, and relatively moderate climatic conditions. Conversely, the southern and southwestern parts of the county were more frequently classified as unsuitable or completely unsuitable due to climatic limitations, conservation restrictions, environmental sensitivity, and reduced accessibility.



The Random Forest model produced a broadly similar spatial pattern. Higher-suitability zones were mainly concentrated in the northern and northeastern parts of Khodabandeh County, while unsuitable and completely unsuitable zones were mostly located in the southern and southwestern areas. The RF model achieved an overall classification accuracy of 0.9538 on the testing subset, indicating strong internal sample-level predictive performance. Variable importance results showed that tourism attractions, river proximity, and temperature were among the most influential predictors in the RF model. This pattern was broadly consistent with the ANP-derived priorities, suggesting convergence between expert-based weighting and data-driven modeling in identifying accessibility, water availability, and climatic comfort as key determinants of ecotourism suitability.



The overlay-based comparison between the ANP–fuzzy–GIS and RF maps provided a more detailed interpretation of model agreement. Exact overall spatial agreement between the two maps was 39.83%, and Cohen’s kappa coefficient was 0.247, indicating fair exact agreement beyond chance. However, because the suitability classes are ordinal, exact pixel-to-pixel agreement alone does not fully represent the degree of consistency between the two outputs. When agreement within one adjacent suitability class was considered, the agreement increased to 90.57%, and the quadratic weighted kappa reached 0.715. These results show that most disagreements between the two models occurred between neighboring classes, such as suitable versus relatively suitable or relatively suitable versus unsuitable, rather than between completely contrasting classes. Therefore, the RF output supports the broad spatial pattern of the ANP–fuzzy–GIS model, especially at the regional zoning level.



Discussion



The findings demonstrate that ecotourism suitability in Khodabandeh County is shaped by the interaction of climatic, topographic, hydrological, accessibility-related, and conservation factors. The concentration of higher-suitability areas in the northern and central parts of the county indicates that these zones provide better conditions for controlled ecotourism development. These areas combine more favorable terrain, better proximity to tourism attractions, access to water resources, and relatively suitable climatic conditions. In contrast, the southern and southwestern parts of the county require more restrictive and conservation-oriented management because they are affected by environmental limitations, reduced accessibility, and higher ecological sensitivity.



The comparison between ANP–fuzzy–GIS and RF results strengthens the interpretation of the spatial suitability pattern. Although exact agreement between the two maps was moderate, the high agreement within one adjacent class and the substantial quadratic weighted kappa indicate broad ordinal consistency. This means that the two models generally agree on the regional pattern of suitability, even where they differ in the exact allocation of class boundaries. This is important because expert-based and data-driven models rely on different forms of reasoning. The ANP–fuzzy–GIS model reflects expert judgment, ecological thresholds, and uncertainty-aware spatial integration, while RF identifies patterns based on predictor variables and reference samples.



Conclusion



This study developed an integrated ANP–Fuzzy–GIS–RF framework for ecotourism suitability assessment in Khodabandeh County. The results showed that approximately 65% of the county falls within suitable and relatively suitable categories, while about 35% is classified as unsuitable or completely unsuitable. Completely suitable areas are limited and mainly concentrated in northern parts of the county, indicating that ecotourism development should be spatially targeted rather than uniformly promoted.



From a planning perspective, northern and central zones can be considered priority areas for controlled ecotourism investment, infrastructure improvement, visitor management, and community-based tourism initiatives. Development in these areas should follow low-impact design principles and include environmental monitoring, carrying-capacity considerations, and coordination among local authorities, conservation agencies, tourism planners, and local communities. In contrast, unsuitable and completely unsuitable areas should be protected from intensive tourism development and managed through conservation-oriented strategies.



Overall, the proposed framework provides a transparent and replicable model for sustainable ecotourism planning. The integration of ANP and fuzzy logic enables the incorporation of expert judgment and ecological uncertainty, while RF modeling offers complementary data-driven support for assessing spatial consistency. However, RF results should be interpreted as comparative evidence rather than complete independent validation, because spatial samples and environmental predictor layers may contain spatial autocorrelation. Future studies should incorporate field observations, visitor-use data, stakeholder interviews, higher-resolution environmental datasets, and socio-economic indicators to strengthen empirical validation and improve the policy relevance of ecotourism suitability assessments.