Machine learning assisted multi-criteria decision-making approaches for site selection: A systematic review Authors: Disha Upadhyay, Anuj Kumar, Sangeeta Pant, Ketan Kotecha, Ajith Abraham Published: 2026 ABSTRACT Highlights•A systematic review of 189 articles (2015–2025) using PRISMA and the CIMO framework.•Seven domains covered, such as Energy generation, public services, Logistics and infrastructure, Retail, commercial& leisure, Water management, Waste management, Emergency, safety, and disaster.•Proposing a hybrid site selection (MCDM+ML) framework that combines Benchmarking, Explainability, and Validation to enhance methodological transparency and decision support across domains.Keywords: Site selection, Multi-criteria decision-making (mcdm), Analytic hierarchy process (ahp), Technique for order preference by similarity to ideal solution (topsis), Machine learning (ML), Artificial neural network (ANN) Background Location of any manufacturing plant is a competitive advantage for the organization from a point of view of cost, logistics, supply chain, or proximity to customers [1,2,4]. Choosing a site involves looking at multiple factors at a time. For selecting a location that should be economically feasible, environmentally friendly, socially acceptable, in proximity to infrastructure, and also fulfill government regulations [5,6,9]. In growing years, decision makers are in complex situations and dynamic environments where expert-driven approaches are no longer sufficient, but the advantage is having sufficient data [7,8]. (Table 1). Table 1Specifications table.Table 1 dummy alt textSubject areaMathematics and Statistics, Decision support systemMore specific subject areaMulti-criteria decision making (MCDM) methods, Machine Learning (ML) methodsName of the reviewed methodologySystematic review using PRISMA guidelines and the CIMO framework.KeywordsMulti-criteria decision making (MCDM), Machine Learning (ML), Hybrid (MCDM+ML) methods, site selection, decision support modelResource availabilityNot ApplicableReview questionRQ1. In which domains have site selection using MCDM and/or ML techniques been applied between 2015 and 2025? (Context)RQ2. Which specific MCDM and ML methods have been used for site selection? (Intervention)RQ3. What factors influence the selection of these methods for different application areas? (Mechanism)RQ4. What are the commonly used criteria across different domains, and how do these methods enhance decision-making? (Outcome) Specifications table. Initially in site selection context Multi-Criteria Decision Making (MCDM) methods such as AHP, TOPSIS, Analytic Network Process (ANP), and Preference Ranking Organization Method for Enrichment Evaluation (PROMETHEE) have been used for evaluating alternatives [13,16,10,22] with Geographic Information System (GIS). These methods are valuable in stakeholder-driven contexts but are not considered for subjective weighting [21,60]. They have limited capacity to process high-dimensional, heterogeneous datasets for site selection [177]. Machine Learning (ML) methods have powerful predictive and pattern recognition capabilities. In recent studies, it is shown that from offshore carbon storage [155] to urban park planning [158], machine learning is used in large-scale socio-economic and environmental datasets. Algorithms such as Random Forest, Deep Neural Network (DNN), Extreme Gradient Boosting (XGBoost), and Support Vector Machine are effective in capturing complex, non-linear relationships [156,165]. In situations where decisions carry high stakes or require emergency planning, decision makers need not only transparent and explainable outputs but also outputs that help reduce potential bias [159,162]. Recent studies highlight that Hybrid MCDM-ML frameworks have emerged as a promising direction for site selection [179,184]. By combining the interpretability and systematic reasoning of MCDM with the predictive accuracy and adaptability of ML [181,182], these hybrid approaches (MCDM+ML) enhance the processes of weighting and ranking [180,183]. These approaches show insights from large datasets, remain reliable when conditions are unpredictable, and keep transparency for stakeholder involvement, while also making use of computational efficiency [185,187]. This integration shows a movement toward accountable, explainable, evidence-informed decision support [186]. Apparently, site selection can be understood as a multi-objective optimization problem under uncertainty, where decisions are guided by rationality and evidence-based reasoning [30,31]. In this review, the terms site selection, location selection, and siting are used interchangeably to describe the process of identifying and evaluating suitable locations for a specific facility, project, or activity. The terminology varies across studies and application domains, but the underlying concept remains the same. This research directly supports sustainability transition theory and the United Nations’ Sustainable Development Goal 7 (Affordable and Clean Energy) by enabling transparent, data-driven siting of renewable energy and other sustainable infrastructure [35,36,47]. It also contributes to optimization and energy systems planning by offering frameworks that integrate technical performance, cost efficiency, and environmental sustainability [91,94]. This review is among the first to systematically integrate MCDM, ML, and hybrid approaches across multiple domains, including energy generation, public services, logistics and infrastructure, retail and commercial facilities, water management, emergency and disaster management, and waste management over the period 2015–2025. Existing reviews are limited to specific application areas like energy systems or GIS-based site selection [48,49]. Its contributions are fourfold: •Developing a systematic categorization of methods, domains, and criteria. Developing a systematic categorization of methods, domains, and criteria. •Mapping evolution over time from conventional MCDM to advanced ML and hybrid (MCDM+ML) frameworks. Mapping evolution over time from conventional MCDM to advanced ML and hybrid (MCDM+ML) frameworks. •Providing cross-domain synthesis to explain why certain methods dominate in specific contexts. Providing cross-domain synthesis to explain why certain methods dominate in specific contexts. •Proposing a research gap solution framework linking methodological limitations to actionable future directions. Proposing a research gap solution framework linking methodological limitations to actionable future directions. Previous versions of reviews have examined site selection methods that were mostly limited in scope, focusing either on single domains such as renewable energy (e.g., solar, wind, or hydropower) [132,138], on specific infrastructures such as dams or landfills [109], or on individual methodological families such as MCDM [134]. These reviews, though valuable, provide only a limited outlook of the methodological landscape. In contrast, the present study offers the first systematic synthesis that integrates MCDM, machine learning, and hybrid frameworks across multiple domains, including energy, healthcare, waste management, disaster resilience, and commercial infrastructure. Using PRISMA and CIMO frameworks, this review not only maps methodological diversity but also highlights cross-domain trends, interpretability challenges, and emerging opportunities for integration. This placement differentiates our work from earlier reviews and focuses on its findings to both methodological innovation and practical decision support. The rest of the paper follows the PRISMA and CIMO frameworks. It begins by presenting the main insights on MCDM, ML, and hybrid (MCDM+ML) methods used in site selection. The findings are then organized by domain and method category, after which the proposed conceptual framework is introduced. This is followed by the discussion and the final section on future research directions. Method detail This systematic review shows the CIMO logic framework (Context, Intervention, Mechanism, Outcome) to structure and analyse the literature on site selection. The framework ensures clarity and consistency by joining the research domain, the methods applied and their use, and the results achieved. The methodology follows established systematic review practices and incorporates elements of the PRISMA approach to ensure transparency and reliability. Research questions The review was guided by the following research questions (RQs): RQ1. In which domains have site selection using MCDM and/or ML techniques been applied between 2015 and 2025? (Context) RQ2. Which specific MCDM and ML methods have been used for site selection? (Intervention) RQ3. What factors influence the selection of these methods for different application areas? (Mechanism) RQ4. What are the commonly used criteria across different domains, and how do these methods enhance decision-making? (Outcome) Search strategy A systematic literature search was conducted using major peer-reviewed publisher-hosted scholarly platforms, including ScienceDirect (Elsevier), IEEE Xplore, SpringerLink, Wiley Online Library, Taylor & Francis Online, MDPI, and Hindawi. These platforms provide indexed access to peer-reviewed journal articles and conference proceedings. The search focused on publications in English between 2015 and 2025. The initial database searches were conducted on the following dates: ScienceDirect (24 March 2025), IEEE Xplore (25 March 2025), SpringerLink (25 March 2025), Wiley Online Library (26 March 2025), Taylor & Francis Online (26 March 2025), MDPI (24 March 2025), and Hindawi (26 March 2025). As the review progressed, additional searches, screening activities, and reference checks were carried out across the selected databases before the final study selection was completed. Searches were performed primarily in the title and abstract fields across all databases to improve relevance and consistency. In some databases, author keywords or indexed terms were also included where these fields were part of the default platform search behaviour. No additional subject-specific filters or category restrictions were applied. This approach helped maintain consistency and comparability across databases while acknowledging minor differences in platform search behaviour. To identify relevant studies, combinations of keywords related to site selection, MCDM/MCDA techniques, and machine learning were used. A common Boolean search structure was developed and applied across all databases: (“site selection” OR “location selection” OR siting) AND (“MCDM” OR “MCDA” OR “multi-criteria decision making” OR AHP OR TOPSIS OR BWM) AND (“machine learning” OR “decision support system” OR SVM OR “random forest”) This structure was applied across all selected databases. In practice, small adjustments were required to match platform-specific rules such as query length limits, handling of Boolean operators, quotation marks, and field restrictions. Even with these adjustments, the overall keyword groups and search logic remained consistent. Platform-specific query formats and search fields are summarized in Table 2. The IEEE Xplore query was slightly simplified to better align with the platform's search functionality. However, the main concepts related to site selection, MCDM, and machine learning were still represented in the search. No important keyword groups were intentionally omitted, and the overall search strategy remained consistent across all databases. Table 2Platform-specific boolean search queries and search fields.Table 2: dummy alt textPlatformQuery format usedField SearchedScienceDirect(“site selection” OR “location selection” OR siting) AND (“MCDM” OR “MCDA” OR “multi-criteria decision making” OR AHP OR TOPSIS OR BWM) AND (“machine learning” OR “decision support system” OR SVM OR “random forest”)Title and AbstractIEEE Xplore(“site selection” OR “location selection”) AND (“MCDM” OR AHP OR TOPSIS) AND (“machine learning” OR SVM OR “random forest”)Title and AbstractSpringerLink(“site selection” OR “location selection” OR siting) AND (“MCDM” OR “MCDA” OR “multi-criteria decision making” OR AHP OR TOPSIS OR BWM) AND (“machine learning” OR “decision support system” OR SVM OR “random forest”)Title and AbstractWiley Online Library(“site selection” OR “location selection” OR siting) AND (“MCDM” OR “MCDA” OR “multi-criteria decision making” OR AHP OR TOPSIS OR BWM) AND (“machine learning” OR “decision support system” OR SVM OR “random forest”)Title and AbstractTaylor & Francis(“site selection” OR “location selection” OR siting) AND (“MCDM” OR “MCDA” OR “multi-criteria decision making” OR AHP OR TOPSIS OR BWM) AND (“machine learning” OR “decision support system” OR SVM OR “random forest”)Title and AbstractMDPI(“site selection” OR “location selection” OR siting) AND (“MCDM” OR “MCDA” OR “multi-criteria decision making” OR AHP OR TOPSIS OR BWM) AND (“machine learning” OR “decision support system” OR SVM OR “random forest”)Title and AbstractHindawi(“site selection” OR “location selection” OR siting) AND (“MCDM” OR “MCDA” OR “multi-criteria decision making” OR AHP OR TOPSIS OR BWM) AND (“machine learning” OR “decision support system” OR SVM OR “random forest”)Title and Abstract Platform-specific boolean search queries and search fields. ResearchGate was not used as a primary search source. It was used only in a supporting role to help identify additional potentially relevant studies. This was done in two ways: first, by checking the reference lists of already selected papers; and second, by reviewing author-uploaded publication lists when they pointed to formally published work. A study found in this way was considered only if it could be verified as a peer-reviewed journal article or conference paper on an established publisher platform. Once verified, these records were checked against the main database results, and any duplicates were removed before screening. After that, they were treated in the same way as all other records and were screened using the same inclusion and exclusion criteria. Inclusion and exclusion criteria and screening procedure To uphold the validity and applicability of the review, studies were selected based on predefined inclusion and exclusion criteria. Eligible studies met the following conditions: •Peer-reviewed journal or conference papers between 2015 and 2025. Peer-reviewed journal or conference papers between 2015 and 2025. •Studies that applied MCDM, ML, or an integrated MCDM-ML hybrid approach for site selection. In this review, a hybrid model refers to a framework where both ML and MCDM methods are combined within the same workflow. This includes cases where ML supports tasks such as prediction, feature selection, or weight estimation, and MCDM is used for ranking or final decision-making, or vice-versa. Studies that only mentioned both approaches without clear integration were not classified as hybrid. Studies in which machine learning was used only for preprocessing tasks, such as clustering, feature extraction, dimensionality reduction, or data preparation, were not automatically classified as hybrid. A study was considered hybrid only when both the MCDM and ML components contributed directly to the site selection process, such as criteria weighting, prediction, ranking, classification, optimization, or final decision support. When machine learning was used solely as a preprocessing tool and did not influence the final evaluation or ranking of alternatives, the study was classified according to its primary decision-making method. Studies that applied MCDM, ML, or an integrated MCDM-ML hybrid approach for site selection. In this review, a hybrid model refers to a framework where both ML and MCDM methods are combined within the same workflow. This includes cases where ML supports tasks such as prediction, feature selection, or weight estimation, and MCDM is used for ranking or final decision-making, or vice-versa. Studies that only mentioned both approaches without clear integration were not classified as hybrid. Studies in which machine learning was used only for preprocessing tasks, such as clustering, feature extraction, dimensionality reduction, or data preparation, were not automatically classified as hybrid. A study was considered hybrid only when both the MCDM and ML components contributed directly to the site selection process, such as criteria weighting, prediction, ranking, classification, optimization, or final decision support. When machine learning was used solely as a preprocessing tool and did not influence the final evaluation or ranking of alternatives, the study was classified according to its primary decision-making method. •Studies providing a clear methodological framework, defined as including all four elements simultaneously: (i) identification of decision criteria or variables, (ii) description of the model or technique used, (iii) dataset or data source information, and (iv) evaluation or validation procedures where applicable. Studies missing any of these components were excluded to ensure methodological transparency, reproducibility, and comparability. The aim of this review was not only to identify studies using MCDM, ML, or hybrid approaches for site selection but also to compare their implementation, datasets, and validation practices across domains. When these elements were not clearly reported, meaningful comparison was not possible. Although such studies could have been grouped separately, doing so would have increased variation and reduced the overall consistency of the analysis. Studies providing a clear methodological framework, defined as including all four elements simultaneously: (i) identification of decision criteria or variables, (ii) description of the model or technique used, (iii) dataset or data source information, and (iv) evaluation or validation procedures where applicable. Studies missing any of these components were excluded to ensure methodological transparency, reproducibility, and comparability. The aim of this review was not only to identify studies using MCDM, ML, or hybrid approaches for site selection but also to compare their implementation, datasets, and validation practices across domains. When these elements were not clearly reported, meaningful comparison was not possible. Although such studies could have been grouped separately, doing so would have increased variation and reduced the overall consistency of the analysis. •The screening process was carried out in three stages. Duplicate records identified across the databases were removed first. After that, the remaining studies were screened through their titles and abstracts to exclude papers that were not relevant to the scope of the review. The final stage involved full-text assessment of the shortlisted articles against the predefined eligibility criteria. The screening process was carried out in three stages. Duplicate records identified across the databases were removed first. After that, the remaining studies were screened through their titles and abstracts to exclude papers that were not relevant to the scope of the review. The final stage involved full-text assessment of the shortlisted articles against the predefined eligibility criteria. •To reduce selection bias, the screening was performed independently by two authors. Once both authors completed the process, their decisions were compared. Any differences were discussed and resolved jointly until agreement was reached. To further check consistency, a random sample of records was screened again by both authors independently. The inter-rater reliability (IRR) results are presented in Supplementary Table S1 and indicate a high level of consistency in the study selection process. The consistency of screening decisions was assessed using percentage agreement, which reached 95% for the independently screened subset of records. A formal inter-rater reliability measure, such as Cohen's kappa, was not calculated. To reduce selection bias, the screening was performed independently by two authors. Once both authors completed the process, their decisions were compared. Any differences were discussed and resolved jointly until agreement was reached. To further check consistency, a random sample of records was screened again by both authors independently. The inter-rater reliability (IRR) results are presented in Supplementary Table S1 and indicate a high level of consistency in the study selection process. The consistency of screening decisions was assessed using percentage agreement, which reached 95% for the independently screened subset of records. A formal inter-rater reliability measure, such as Cohen's kappa, was not calculated. Excluded papers •Studies that were not directly related to site selection Studies that were not directly related to site selection •Grey literature, such as theses and reports Grey literature, such as theses and reports •Non-English publications Non-English publications •Papers with incomplete or unclear methodological descriptions Papers with incomplete or unclear methodological descriptions During the full-text review stage, 6 studies were excluded based on these criteria. For transparency, Supplementary Table S3 presents the studies excluded during the full-text review stage, along with the main reason for exclusion for each paper. Quality appraisal and risk considerations To ensure the quality of the review, we conducted a structured appraisal during the full-text eligibility stage. A checklist was made to evaluate both the technical and reporting standards of studies, which includes MCDM, ML, or hybrid MCDM+ML approaches to site selection. Each study was independently assessed by two corresponding authors using the predefined criteria, and any differences in judgment were resolved through discussion and consensus. To further assess consistency between authors, a random subset of records was independently reviewed by both authors, and the inter-rater reliability results are presented in Supplementary Table S1, which showed 95% agreement. Each study was evaluated using a checklist based on the following criteria: (i) methodological clarity- It includes a clear description of the modelling framework and criteria selection. (ii) Dataset transparency- It was examined by checking whether data sources and key characteristics, such as size or resolution, were reported. (iii) Validation Procedures-It covers such as Accuracy, Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Area Under the Curve (AUC), Kappa, or Coefficient of Determination (R²), which were reviewed to determine how model performance was evaluated and whether relevant metrics were provided. (iv) Reproducibility-It was considered based on the level of detail given for algorithms and model parameters. Finally, (v) Contextual Relevance-It was used to ensure the study directly addressed a site selection problem. Each criterion was assessed qualitatively rather than scored numerically. Studies were expected to demonstrate adequate coverage across these appraisal domains to be considered methodologically sound. Papers lacking essential detail in any domain, such as unclear dataset description or missing validation information, were excluded during this stage. The quality appraisal was conducted using a qualitative checklist based on these five appraisal domains. No numerical scoring system was applied. The complete appraisal checklist is provided in Supplementary Table S4. Studies with substantial missing methodological information in one or more appraisal domains were excluded following discussion and consensus between the authors involved in the appraisal process. During the quality appraisal stage, 6 studies were excluded after full-text review due to irrelevant scope (n = 1), methodological limitations (n = 2), and insufficient information (n = 3), as shown in the PRISMA flow diagram. The remaining 189 studies met the defined quality standards and were included for the final analysis. All the included studies provide clear descriptions of methodological frameworks and defined decision criteria. Variations in reporting dataset characteristics and validation procedures were observed. These differences reflect variations between MCDM, ML, and hybrid approaches rather than systematic bias. Because the studies varied in their domains, datasets, and validation protocols, we did not apply a quantitative scoring system, a minimum threshold count, or a comparative risk-of-bias index. Instead, performance outcomes were interpreted within the context of each study to avoid overgeneralization. Overall confidence in the evidence was assessed qualitatively, based on the clarity of methods, validation procedures, reproducibility, and consistency of reporting across the included studies. A statistical test for reporting bias, such as publication bias analysis, was not conducted since this review did not include a quantitative meta-analysis. To minimize potential bias, the search covered multiple peer-reviewed databases using predefined strategies, and only peer-reviewed journal and conference publications were included. A formal certainty-of-evidence framework was not applied, as the review focused on methodological mapping rather than effect comparison. Screening and selection process The screening process began with 215 papers on multi-criteria decision making (MCDM), Machine Learning (ML), and hybrid (MCDM+ML) methods in site selection. After removing 18 duplicate records, 197 studies remained for title and abstract screening. At the screening stage, 2 records were excluded because of insufficient research scope, and the remaining 195 studies were utilized in full-text eligibility based on inclusion and quality criteria. 6 studies were excluded in the full-text stage, where 1 was for irrelevant scope, 2 were for methodological limitations, and 3 were for insufficient information. A full-text review was then conducted on the remaining 189 studies to ensure they fulfilled the inclusion criteria. In total, 189 papers were selected for this systematic review. These counts are consistent with the PRISMA flow diagram (Fig. 1). The completed PRISMA 2020 checklist is provided in Supplementary Table S2. PRISMA 2020 flow diagram illustrating the study selection process. Data extraction and categorization A standardized data extraction template was used to record information from each included study. Data extraction was performed by the first author using a predefined data extraction format, and the extracted information was subsequently reviewed by the co-author for consistency and completeness. The following details were recorded: •We recorded the Authors, the paper’s Title, and its year of publication We recorded the Authors, the paper’s Title, and its year of publication •Application in seven domains (e.g., Energy generation, public services, Logistics and infrastructure, Retail, commercial& leisure, Water management, Waste management, Emergency, safety and disaster) Application in seven domains (e.g., Energy generation, public services, Logistics and infrastructure, Retail, commercial& leisure, Water management, Waste management, Emergency, safety and disaster) •MCDM and ML method, and the Hybrid (MCDM+ML) method used MCDM and ML method, and the Hybrid (MCDM+ML) method used •Criteria considered for site selection Criteria considered for site selection •Outcome or findings of the method Outcome or findings of the method •Dataset Characteristics, including data source and size, where reported Dataset Characteristics, including data source and size, where reported •Validation or evaluation approach, including reported performance metrics where available Validation or evaluation approach, including reported performance metrics where available The seven domains were predefined before data extraction. Each study was classified to a single domain based on its primary objective and dominant application area. For studies that could fall on multiple domains in that, the main focus of the research problem was identified and used to find the final domain label. Each study was assigned to only one domain to avoid double counting. For the year-wise trend analysis (Fig. 2, Fig. 3, Fig. 4, Fig. 5, Fig. 6), the 189 included studies were categorized according to their publication year. Each paper was counted once and categorized based on its primary domain and method type (MCDM, ML, or hybrid). Studies using only MCDM techniques were classified under the MCDM category. Those studies using only machine learning approaches were classified as ML. Studies integrating both MCDM and ML within the same framework were categorized as hybrid. Each study was assigned to only one methodological category. Overlapping classifications were not used. The extracted variables are domain, methods used, criteria, geographic scope, data source, sample size, performance metrics, validation methods, and key findings. These extracted data provided the base for the statistical summaries and visualizations presented in the findings section. All performance results and validation measures reported in the included studies were documented as presented by the original authors. When studies reported multiple evaluation metrics, each relevant measure was recorded rather than selecting only specific results. No filtering was applied based on different experimental settings or analysis variations, provided the results were directly related to site selection performance. Problem domain bar chart. Problem domain classification stacked bar chart. MCDM methods in site selection studies (2015–2025). Machine learning models in site selection studies (2015–2025). Hybrid (MCDM+ML) methods in site selection studies. If certain details, such as dataset size or validation procedures, were not clearly stated in a study, this was noted as “not specified”. No missing information was assumed or calculated. Due to heterogeneity in domains, datasets, and evaluation metrics, quantitative meta-analysis was not appropriate, and no sensitivity testing was performed. Information regarding funding sources or author affiliations was not collected, since it was not directly relevant to the methodological focus of this review. As this review used a qualitative synthesis rather than meta-analysis, it presented performance metrics (e.g., Accuracy, RMSE, MAE, AUC, R², Kappa) used as the primary outcome measures. No statistical transformations or data conversions were performed. Extracted information was synthesized as reported in the original studies. This review has some limitations. The search was limited to English-language publications and selected publisher-hosted databases between 2015 and 2025. Grey literature, theses, reports, and non-English studies were excluded. As a result, some relevant studies may not have been captured. In addition, excluding grey literature may have introduced some publication bias, as studies with positive or stronger results are generally more likely to appear in published sources than studies reporting negative or inconclusive findings. The performance metrics reported in the reviewed studies varied depending on the type of site selection problem and the modelling approach adopted. Studies focused on classification tasks typically reported metrics such as Accuracy, AUC, F1-score, and Kappa, whereas studies involving prediction or estimation commonly used RMSE, MAE, and R². Since the reviewed studies differed in their application domains, datasets, objectives, and validation strategies, direct comparison of performance results across studies was often not appropriate. Therefore, the reported metrics were presented and interpreted within the context of each study. Synthesis approach Using the CIMO logic, the studies were analysed by matching the context (problem area), the intervention (MCDM/ML methods), the mechanism (why and how the method was used), and the outcome (the effectiveness of the approach). This allowed for a clearer understanding of how MCDM and ML techniques are being applied in various domains and which approaches work best in which situations. Key insights from ML and hybrid methods in site selection Multi-Criteria Decision Making (MCDM) techniques are prominently used for site selection across various domains where expert judgment and structured evaluation frameworks are necessary. AHP, TOPSIS, PROMETHEE, and ELimination Et Choix Traduisant la REalité (ELECTRE) are among the most utilized methods, applied in contexts such as Energy generation [4,5,7,14,[17], [18], [19], [20],28,[34], [35], [36],38,42,43,[47], [48], [49],51,55,[61], [62], [63], [64], [65],69,72,80,82,83,87,89,91,94,95,97,99,100,102,[110], [111], [112], [113],120,122,123,128,132,134,137,138,141,142,149,150,[152], [153], [154]], Water management [[9], [10], [11], [12],[23], [24], [25],41,45,52,54,77,[107], [108], [109]] and public services [1,2,16,26,29,32,68,75,84,88,125,129,140,147], Waste management [22,31,37,40,44,53,57,58,66,70,71,73,76,[78], [79], [80], [81],86,90,98,103,105,106,114,121,124,127,131,135,139,144] Logistics and infrastructure [3,6,21,27,30,39,46,50,56,60,74,93,96,101,104,116,119,133,136,143,148,151] Retail, commercial & leisure [8,92,145], Emergency, safety and disaster [13,15,33,59,67,85,115,117,118,126,130,146]. The site selection process involves multiple interrelated factors to make optimal decisions. The main key factors include economic, environmental impacts, proximity to location, risk assessment and sustainability. Decision-makers need to balance these factors based on project needs and industry and stakeholders’ requirements. An analytical method that incorporates both the quantitative and qualitative aspects will be more objective. Through available data, analysis and other considerations, the best decision can be made for an organization's future stability. In the last decade, multi-criteria decision-making (MCDM) methods have been extensively employed in different fields for selecting suitable sites. Classical techniques like AHP, Fuzzy AHP and ANP continue to find application in various industries such as public services [32,75], energy generation [47,149], logistics and infrastructure [151], retail, commercial & leisure [169], water management [77], emergency, safety and disaster [67] as well as waste management sectors [70]. Ranking-based techniques such as ELECTRE and its variants [12,109], TOPSIS with its fuzzy extensions [127,102,74], and PROMETHEE and its variants [30] have also gained popularity for site selection. Aggregation-based techniques such as Weighted Linear Combination (WLC) [21,69], and Delphi-AHP hybrid methods [119,60] have also been used. Most recently, the Best–Worst Method (BWM) [141,62], Weighted Multi-Objective Optimization on the basis of Ratio Analysis (Weighted MOORA) [133] and Type-2 fuzzy aggregation [90,28] were added for this purpose. A hybrid MCDM framework, as defined in Section 2.3, combining different techniques with GIS and fuzzy logic has been used [117,179,68,66]. These combinations show that MCDM techniques can be integrated to support site selection, especially when multiple interrelated criteria and spatial factors have equal importance in the context of site selection. All 189 included studies were carefully classified by application domain and methodological approach. However, more detailed analysis was devoted to ML and hybrid methods because these approaches are relatively recent and have received less consolidated attention in existing reviews. Site selection using machine learning (ML) methods Machine Learning (ML) has been integrated into site selection processes, mainly in domains having complex pattern recognition from large datasets. Techniques like Random Forest (RF), XGBoost, Artificial Neural Network (ANN) and deep learning models have been applied in energy generation, where several studies reported strong predictive performance under specific datasets and validation conditions. These studies (2015–2025) confirm that MCDM methods are already explored in the existing literature, including various methodological combinations. While this review acknowledges their breadth, the novel contribution of the present work lies in examining the use of machine learning (ML) and ML+MCDM hybrid methods. These methods are comparatively less explored yet relevant for handling complex site selection problems in several domains. Energy generation [155,156,162,163,165,[175], [176], [177]], Waste management [168], Logistics and infrastructure [158,[170], [171], [172], [173],178] Retail, commercial & leisure [157,161,169,174], public services [166,167], Emergency, safety & disaster [159,160,164]. ML methods offer advanced data-processing capabilities for handling large and complex datasets. However, challenges remain such as limited labelled data, feature selection complexities, and the need for explainability. These techniques contribute to decision-making by uncovering hidden patterns and trends within complex datasets, particularly in real-world site selection applications. Machine learning has advanced site selection studies by helping researchers handle large and diverse datasets. In energy generation domains, machine and deep learning models have reported high predictive capability in several individual studies. For example, DNNs have improved offshore CO₂ storage siting [155]. Machine learning models such as XGBoost, combined with SHapley Additive exPlanations (SHAP) analysis, have enhanced wind farm site planning [162] and ensemble learning approaches have supported national-scale suitability map for wind farms [163]. Similarly, Random Forest and gradient boosting models have supported the siting of solar PV in arid areas [165], while Genetic Algorithm (GA)-based optimization has strengthened planning for energy storage and hybrid renewable systems [176]. In these energy-focused studies, ML methods were reported to support technical feasibility, cost efficiency, and sustainability assessment within the specific application contexts, datasets, and evaluation settings considered. ML has also been applied in public services, Emergency, safety, and disaster domains. Support Vector Machine (SVM) and Random Forest were reported to support the siting of hospitals and elderly care facilities by addressing spatial imbalance and demographic pressures in several reviewed studies [159,166,170]. Emergency and evacuation domain problems and graph-based hybrids have been applied to the allocation of emergency shelters and evacuation centres, with reported reductions in travel distances [160,164] using ANNs and Variational Graph Autoencoder (VGAE). Urban Park planning has also used Decision Trees 158and k-means algorithm has supported business facility location analysis [171]. These examples illustrate how ML-based approaches have been applied to support planning and allocation decisions in public infrastructure, emergency site selection, and logistics-related applications. Several reviewed studies reported that ML methods supported accessibility and allocation of critical services. In studies [157,169] Random Forest was reported to perform better than regression-based models for retail demand analysis and consumer clustering. Ensemble learners and multimodal facilities have been applied to ATM deployment and location strategy analysis for businesses. Clustering and recommendation frameworks have also been used for ATM installation and business sitting in crowded urban markets [173]. Recent studies indicate that, in specific retail and commercial contexts, ML methods were associated with improved technical efficiency while also supporting competitiveness and customer-oriented services. Hybrid ML approaches are used in neuro-fuzzy systems applied for hydropower siting [175] and Artificial Neural Network-Genetic Algorithm (ANN-GA) frameworks for cultural infrastructure [178]. Limited labelled data, complex feature selection, and the need for clear, explainable outputs are the challenges to building such a model that can be trusted by stakeholders and policymakers. For integrating such a framework into mainstream planning and policy for site selection, these issues are very important. Machine learning methods are often used for site selection problems involving diverse datasets and nonlinear relationships. Random Forest, XGBoost, and deep neural networks are commonly applied algorithms in fields such as energy generation, water management, waste management, logistics, and Emergency planning. These methods have been reported to provide reliable predictions and scalability in several domain-specific studies. They have also been used to analyse complex spatial and socio-economic factors, when evaluating suitable sites. Table 3 summarizes representative ML studies with their method used, criteria considered, datasets and size, and performance. Table 3Summary of representative machine learning- based site selection studies.Table 3: dummy alt textAuthor & YearDomainML Method(s)CriteriaGeographic ScopeData SourceSample SizePerformanceValidation MethodKey FindingNassabeh et al. [155]Energy generationXGBoost, RF, MLELM, DNNReservoir, geological, petrophysical, and logistical factorsGulf of MexicoWellbore and geospatial datasets36,000 records; 44 featuresDNN: R² = 0.9937, RMSE = 0.9279Holdout validation using 70/15/15 train-validation-test split with cross-validationHigh predictive performance for CCUS siting.Rahimi et al. [156]Energy generationK-Means, DBSCAN, Hierarchical, K-Medoids, GALocation, population, solar radiation, wind speed, land areaAustraliaAustralian Towns DatasetApprox. 7.2 million rural population records grouped into 10 clustersGA K-Means: DB = 0.42; CH = 611.3GA K-Medoids: DB = 0.57; CH = 498.6Simulation-based and cluster validation using HOMER Pro analysis plus clustering metricsImproved renewable site identification.Zhao et al. [157]Retail, commercial & leisureRF, GD Regression, OLSCommercial services, financial services, residential proximityBeijing, ChinaUrban commercial dataset669,580 POIs (2020); 626,944 POIs (2022); 500 m grid samplingRF: R² = 0.929 (with buffer); R² = 0.915 (without buffer)Holdout validation using a 70/30 train-test split with performance metric comparisonRF achieved the best performance for retail siting.Tang et al. [158]Logistics and infrastructureID3 Decision TreeInfrastructure, medical, commercial, and administrative facilitiesGuangzhou, ChinaPOI data and ArcGIS grids663,749 POIs; 29,793 grids (500 × 500 m)Model agreement with existing parks=93.7%Spatial validation using GIS overlay with actual park distribution and service pressure analysisSuitable and priority sites identified.Almansi et al. [159]Emergency, safety, and disasterSVM, MLP, LREnvironmental, topographic, geodemographicGaza Strip, PalestineDEM and hospital inventory data29 hospital sites and 29 non-sites, 15 conditioning factorsMLP: AUC = 0.849; SVM: AUC = 0.756Mixed validation using 70/30 holdout split and 10-fold cross-validationMLP achieved the strongest predictive performance.Amiran et al. [160]Emergency, safety, and disasterANN, GPC, KNN, SVMLand slope, distance to fire/medical/security centres, population densityTehran, IranOpenStreetMap and San Francisco training dataset373 neighbourhoods; 8 criteriaANN: F1 = 0.9603Spatial transfer validation using training on San Francisco data and GIS-based overlay testing in TehranANN showed the best performance for evacuation siting.Wang et al. [161]Retail, commercial & leisureBP-NN, SVR, RFPopulation, housing prices, POIs, traffic centrality, Weibo check-insBeijing, ChinaCensus, housing, Weibo, and POI datasets5823 signage pointsBP-NN: RMSE = 0.264Cross-validation with ROC-based model evaluationImproved location prediction accuracy.Bilgili et al. [162]Energy generationXGBoost, LightGBM, RF, CART, LR, NB, HGB, SHAPWind speed, elevation, transmission proximityBalıkesir Province, TürkiyeGlobal Wind Atlas, DEM, OSM, ESA Worldcover454 turbines, 2418 pixelsXGBoost: Accuracy = 96.1%; Recall = 98.1%; F1 Score = 96.2%.Spatial validation using GIS overlay against existing wind turbine locationsXGBoost achieved the highest wind siting accuracy.Khalaf et al. [163]Energy generationRF, KNN, SVM, NBNatural, socio-economic, environmental factorsTürkiyeGlobal Wind Atlas, USGS DEM, MapCruzin, Protected Planet7.2 million geo-pointsSVM: Accuracy = 95.1%Spatial validation by comparing GIS suitability maps with existing wind farms.Reliable wind suitability maps generated.Wang et al. [164]Emergency, safety, and disasterVGAE + RF, SVM, KNN, MLPDEM, POIs, population grids, road networks, censusBeijing, ChinaGeospatial Data Cloud, Census, OSM, POI datasetsHexagonal grid units of about 500 mVGAE + RF: F1 = 0.902Cross-validation with ROC and confusion matrix assessmentBalanced and accurate shelter site prediction.Ashraf et al. [165]Energy generationRF, XGBoost, MLP, SHAPGeographical, Socio-Economic, Resource Condition FactorsCholistan Desert, Punjab, PakistanSolargis, USGS DEM, Sentinel-2 NDVI, ERA5 wind, Humdata355 field survey pointsRF: AUC = 0.921; Accuracy = 72%Holdout validation using a 70/30 train-test split with expert interpretation via SHAP.Highly suitable solar sites identified.Li & He [166]public servicesRF, GWR, GWRF, SHAPNatural, socio-economic, facilities, transport, ecology, securityShenzhen, ChinaBaidu POIs, Fang housing, OSM, WorldPop, Earth Observation Group data6670 hexagonal gridsGWR: R² = 0.91; AICc = 110,914.81Mixed validation using 70/30 holdout split and spatial validation through geographically weighted modelling.Key suitability factors identified.Gu et al. [167]public servicesID3 decision treeGovernment, medical, financial, transport, and attractions facilitiesXi’an, ChinaBaidu Map41,275 grid cells of 500 m × 500 m; 472 existing facilitiesID3:Accuracy = 85.8%Holdout validation using a 70/30 train-test split with classification performance assessmentSuitable public facility sites identified.Xia et al. [168]Waste managementMLR, BR, ER, SVR, GBR, RFEconomic, Social, Demographic,Remote Sensing IndicatorsBeijing, ChinaOpenStreetMap144 observations from 16 districts over 2012 to 2020GBR: R² = 0.995; MAE = 2.35Cross-validation using 10-fold CV with spatial interpolation support.Improved waste prediction accuracy.Cai et al. [169]Retail, commercial & leisureCNN, LR, SVM, GBR, RFPopulation, POIs, traffic, competitionZhanjiang, Guangdong, ChinaWorldPop population data, Gaode Maps Open Platform API, OSM road network data20,022 retail stores; 127,503 POIsCNN: RMSE = 0.036; MSE = 0.0013Holdout validation using 60/20/20 train-validation-test split with data augmentationCNN achieved best retail siting performance.Douard et al. [170]Logistics and infrastructureLightGBM classifierPopulation, land cover, elevation, roads, facilitiesGhanaOpenStreetMap376 hospitals and surgical centres in Ghana.LightGBM:Sensitivity =0.94; Precision = 0.98; F1 = 0.96Spatial cross-validation using tile-based partitioning with expert manual validationHigh agreement with expert validation.Khetarpaul et al. [171]Logistics and infrastructureK-Means, HAC, DBSCANVenues, cuisine, ratings, pricing, sanitation, COVID contextNew DelhiFoursquare API and Zomato APIRestaurant venue dataset across multiple districts and about 10 sub-districtsK-Means formed 7 well-defined clusters and outperformed HAC and DBSCANCluster validation using silhouette and elbow methods.High-potential restaurant sites identified.Mazhi et al. [172]Logistics and infrastructureGIS, PCA, MLPOIs, transport, socio-economic, consumer distributionSouth Jakarta, IndonesiaGoogle Maps, ArcGIS, and retail revenue data1581 POIs, 18 test sitesCorrelation (r) = 0.939; MAPE = 5.35%Correlation-based and forecast-error validation using Pearson correlation and MAPESuitable new store sites identified.Rastogi et al. [173]Logistics and infrastructure27 classifiers including LogitBoost, RF, BaggingProximity to schools, apartments, petrol pumps, ATMs, tourist sitesNoida, IndiaGoogle Maps API, Here Technologies API; manually curated dataset42 features, 42 ATM locationsLogitBoost: Accuracy = 85.1%; F1 = 0.851Holdout validation using an 80/20 split with ROC, PRC, and MCC comparisonEnsemble models improved ATM siting.Choudhury et al. [174]Retail, commercial & leisureSVR, RF, XGBoost, Ridge, Lasso, ElasticNet, MLP, CNNsHospitals, amusement, food, roads, education, facilities, buildings, routes, and tourism.Dhaka, BangladeshYelp check-ins, OSM, World Pop, Bing Maps Satellite1108 restaurant images; 150k businesses; 131k check-insMLP: RMSE = 0.050; MAE = 0.037Holdout validation using an 80/20 train-test split with error metric comparisonNeural networks captured complex patterns.Shimray et al. [175]Energy generationMLP, fuzzy inference system, ANNEcology, hostility, cost of energy, water quality, air qualityManipur, Sikkim, Arunachal Pradesh, Himachal Pradesh, IndiaEIA reports of hydropower plant sites; CPCB/MoEF guidelines4 hydropower sitesFinal suitability score = 82.7 for the top-ranked siteRule-based comparative validation using neuro-fuzzy evaluation and ranking across candidate sitesThe neuro-fuzzy framework identified the best site.Xu et al. [176]Energy generationImproved GA + PSOPV curve, economics, safety, climatology, land cost, policy, power qualityChinaMATLAB-generated simulation datasetNot specifiedImproved GA: Accuracy = 85.92%Simulation-based validation using comparative performance metricsImproved GA enhanced PV site optimization.Holloway et al. [177]Energy generationK-Means, K-Medoids, simulationSolar irradiation, wind speed, and wind strengthRural Western AustraliaTownsites LGATE-248 WA GeoJSON dataset and Back4App dataset69 rural townsK-Means: Dunn = 0.146; K-Medoids: Dunn = 0.072Internal cluster validation using the Elbow method, Silhouette score, and Dunn index, and Simulation validation via HOMER ProEffective renewable clustering achieved.Zhao et al. [178]Logistics and infrastructureANN, GAUrban Service Facilities, Spatial Buffer Zone, Socio-Economic & Cultural ContextShanghai, ChinaAutoNavi POI data, Baidu Map API shapefiles, Dianping Visiting Vitality Values, Shanghai administrative map344 CMPBsANN: MSE = 0.0081Holdout validation using untrained test data with scenario-based simulation analysisANN-GA supported site vitality optimization.Note: RF: Random Forest; DNN: Deep Neural Network; MLELM: Multi-Layer Extreme Learning Machine; MLP: Multi-Layer Perceptron; CNN: Convolutional Neural Network; GD: Gradient Descent; LightGBM: Light Gradient Boosting Machine; SVR: Support Vector Regression PSO: Particle Swarm Optimization; LR: Logistic Regression; NB: Naïve Bayes; KNN: K-Nearest Neighbors; SVR: Support Vector Regression; GBR: Gradient Boosting Regression; CART: Classification and Regression Tree; BP-NN: Backpropagation Neural Network; SVM: Support Vector Machine; OLS: Ordinary Least Squares GPC: Gaussian Process Classifier; PCA: Principal Component Analysis; HGB: Histogram-Based Gradient Boosting; GWR: Geographically Weighted Regression; GWRF: Geographically Weighted Random Forest; HAC: Hierarchical Agglomerative Clustering RMSE: Root Mean Square Error; MAE: Mean Absolute Error; MSE: Mean Squared Error; AUC: Area Under the Curve; R²: Coefficient of Determination; DB: Davies–Bouldin Index; CH: Calinski–Harabasz Index; AICc: Corrected Akaike Information Criterion; MAPE: Mean Absolute Percentage Error; VGAE:Variational Graph Autoencoder; HOMER: Hybrid Optimization of Multiple Energy Resources. Summary of representative machine learning- based site selection studies. Note: RF: Random Forest; DNN: Deep Neural Network; MLELM: Multi-Layer Extreme Learning Machine; MLP: Multi-Layer Perceptron; CNN: Convolutional Neural Network; GD: Gradient Descent; LightGBM: Light Gradient Boosting Machine; SVR: Support Vector Regression PSO: Particle Swarm Optimization; LR: Logistic Regression; NB: Naïve Bayes; KNN: K-Nearest Neighbors; SVR: Support Vector Regression; GBR: Gradient Boosting Regression; CART: Classification and Regression Tree; BP-NN: Backpropagation Neural Network; SVM: Support Vector Machine; OLS: Ordinary Least Squares GPC: Gaussian Process Classifier; PCA: Principal Component Analysis; HGB: Histogram-Based Gradient Boosting; GWR: Geographically Weighted Regression; GWRF: Geographically Weighted Random Forest; HAC: Hierarchical Agglomerative Clustering RMSE: Root Mean Square Error; MAE: Mean Absolute Error; MSE: Mean Squared Error; AUC: Area Under the Curve; R²: Coefficient of Determination; DB: Davies–Bouldin Index; CH: Calinski–Harabasz Index; AICc: Corrected Akaike Information Criterion; MAPE: Mean Absolute Percentage Error; VGAE:Variational Graph Autoencoder; HOMER: Hybrid Optimization of Multiple Energy Resources. It is important to note that the performance observations presented in this section are drawn from the results reported in the individual studies reviewed. This work does not include a quantitative meta-analysis or a direct statistical comparison between machine learning and classical MCDM approaches. Therefore, any reported performance improvements should be understood within the specific context and conditions of each respective study. The performance metrics shown in Table 3 are presented as reported in the original studies. As these studies vary in terms of datasets, validation approaches, case contexts, and evaluation measures, the results are not directly comparable and are included for descriptive purposes only. The results summarized in Table 3 illustrate the application of ML methods across different site selection domains, datasets, validation methods, and application contexts. For example, Random Forest achieved ∼92% accuracy in urban library siting [156], while XGBoost reached 96% accuracy in wind farm planning [162]. Deep neural networks for CO₂ storage reported R² values above 0.99 [155], and ANN-based models improved emergency coverage from 27% to 59% [160]. In retail, RF and CNN models achieved R² ≈ 0.93 and reduced error rates compared with regression baselines [157,169]. Across the reviewed studies, ML methods were applied to address objectives related to technical feasibility, cost efficiency, and equity in several application areas, particularly energy, logistics, and public service site selection. However, the reported benefits varied depending on the dataset, study objectives, and evaluation approach adopted in each study. Despite these advantages, the need for large labelled datasets and the limited interpretability of some ML models remain important challenges for wider adoption. Site selection using hybrid (MCDM+ML) methods Hybrid approaches represent a growing trend in site selection. These models combine evaluation approaches, such as AHP and TOPSIS, with ML algorithms like Support Vector Machines (SVM) and Gradient Boosting Regression (GBR), and ANN. This integration provides a framework for decision-making, like Energy generation [181,184,186,187] Waste management [179,180,183], public services [182], Emergency, safety & disaster [188], Water management [189], Logistic and infrastructure [185]. Hybrid (MCDM+ML) models have been applied in areas such as waste management, emergency safety, disaster, public services, and energy generation, etc. It involves combining multi-dimensional datasets with stakeholder contributions, allowing both data-driven and expert-based factors to be considered during site selection. Starting with waste management, Fuzzy Analytic Hierarchy Process (FAHP)-ML-Explainable Artificial Intelligence (XAI) models have been used for landfill siting in Saudi Arabia. In this paper, clustering, ensemble learning, and explainable AI tools like SHAP and Local Interpretable Model-Agnostic Explanations (LIME) were reported to improve both predictive robustness and transparency [179]. Hybrid AHP-RF models have been applied in India for sewage treatment plants and solid waste disposal. By combining expert weighting with ML classification, these studies developed suitability maps for urban and metropolitan cities [180,183]. These studies illustrate how the hybridization framework has been used to incorporate environmental considerations and stakeholder inputs within site selection processes. In the energy sector, ML+MCDM hybrids have been used for renewable energy storage and siting. For example, fuzzy TOPSIS, Additive Ratio Assessment (ARAS), Evaluation based on Distance from Average Solution (EDAS), and Multi-Objective Optimization by Ratio Analysis (MOORA) are used for rankings, and the hybrid method was used to address the complexity of problem [181]. AHP, MLP, and XAI frameworks have been applied to solar PV siting, integrating fuzzy evaluation, clustering, and SHAP interpretability [184]. Models such as AHP, Fuzzy Inference System (FIS), and Self-Attention Convolutional Long Short-Term Memory (SA-ConvLSTM) were reported to improve solar planning by producing more stable suitability maps than static GIS-MCDM approaches in that study [186]. SVM combined with MCDM methods was reported to help stakeholders handle complex datasets while considering multiple factors in site selection [187]. Hybrid AHP+MLP models have been applied to hospital site selection in public services and infrastructure [182]. For rural settlement planning in Emergency, disaster domain RF, AHP, and Entropy methods have been used [188]. The integration of AHP + ML + GIS/RS was reported to provide a useful hybrid framework for strategic site selection [185]. These examples illustrate the application of hybrid approaches across different domains where both stakeholder considerations and practical implementation aspects are important. Dam siting reviews [189] divide methods into GIS/RS, MCDM, and ML, and recommend combining them to improve predictive accuracy and decision quality. Overall, the reviewed studies indicate that MCDM+ML hybrid frameworks have been applied as practical and transparent approaches for site selection across different domains. Table 4 lists representative hybrid studies along with the criteria used, the dataset considered, and outcomes. Table 4Summary of representative hybrid methods (MCDM+ML) in site selection.Table 4 dummy alt textAuthor & YearDomainML Method(s)CriteriaGeographic ScopeData SourceSample SizePerformanceValidation MethodKey FindingAwadh & Mallick [179]Waste managementFAHP, SOM, K-means, Bagging, SHAP, LIMEGeophysical, geoecological, socio-economicAbha-Khamis watershed, Saudi ArabiaALOS PALSAR DEM, Sentinel-2, PME rainfall, USGS imagery1773 km² watershedBagging: MAE =0.799; MSE= 1.187; RMSE= 0.999;R² =0.882Expert judgmentFAHP-SOM-Bagging framework improved landfill suitability assessment.Rajalakshmi et al. [180]Waste managementAHP, RF, fuzzy logicSlope, elevation, LULC, groundwater,soil, proximity to rivers, roads, railways, built-up areasPune, IndiaSentinel-2, SRTM DEM, ASF, ESRI LULC, OSM363.57 km² citySpatial suitability distribution: 0.33% very high suitability; 2.21% high suitabilitySpatial overlay with GIS mapsHybrid RF-AHP approach supported wastewater site selection.Yilmaz et al. [181]Energy generationK-Means++, TOPSIS, ARAS, EDAS, MOORA, BORDAGrid access, renewable proximity, cost, regulations, climateBalikesir, TurkeyGeographic coordinates of WPPs28 WPPsK-Means++: 7 optimal clusters identified using the Elbow method; hybrid MCDM ranking appliedSpatial overlay + consensus MCDAML-MCDM integration improved energy storage site evaluation.Almansi et al. [182]public servicesAHP, MLP, CFSEnvironmental, topographic, and geodemographic parametersMalacca, MalaysiaASTER GDEM (30 m, 2015, Tindak Malaysia)8 hospital sitesMLP: AUC = 0.922, Standard Error = 0.066, R² = 0.998Holdout/test data (10-fold cross-validation)MLP-AHP framework enhanced healthcare facility site selection.Mondal et al. [183]Waste managementAHP, RFgroundwater depth, proximity to sensitive areas, landuse, and topographyEnglish Bazar & Old Malda, IndiaSRTM, ESA, OpenStreetMap, Google Earth ProTwo municipal citiesAHP: Accuracy = 83.83%, Kappa = 0.7894; RF: Accuracy = 80.61%, Kappa = 0.7474Expert judgment + spatial overlayHybrid AHP-RF model supported urban waste management planning.Liao et al. [184]Energy generationAHP, WLC, MLP, K-MEANS, FCE, SHAPClimate, Orography, location, economicsChinaMulti-source GIS, NOAA, USGS, Land China1 km × 1 km resolutionImproved K-Means: CH Index = 2818,164Holdout clustering indicesSHAP-based AHP-WLC-MLP framework improved solar PV suitability mapping.Bojer et al. [185]Logistics and infrastructureAHP, RF, WLCPhysical, Environmental, SafetyAdea District, EthiopiaEsri ArcGIS 10.8, Arcpro 2.9, USGS EarthExplorer936 km²RF: Accuracy = 97%; Kappa = 0.96Spatial overlay with GIS reference layersAHP-RF framework produced robust suitability maps.Jalalifar et al. [186]Energy generationWLC, FIS, SA-ConvLSTMland availability, solar radiation, environmental impactTehran metropolitan region, IranSpatio-temporal load data, GIS layers, DEM, land cover datasetsForecasting horizon = 6 months of load dataSA-ConvLSTM: RMSE = 35.13%; MAE = 26.14%Holdout/test forecastingSA-ConvLSTM integrated spatial and temporal data for energy site planning.Cerna et al. [187]Energy generationSVM, AHPWind speed/direction, elevation, slope, highways, railways, built-ups, Forest zone, scenic areaPhilippinesDEM, GIS DatabasesNational coverageSVM: Accuracy = Luzon (88.7%), Visayas (86.9%), Mindanao (85.6%)Holdout/test dataAHP-SVM framework supported wind power site selection.Shu et al. [188]Emergency, safety, and disasterRF, AHP, Entropy, Information quantity methodPrecipitation, Terrain Slope, Lithology, Land use Type, Altitude, Slope Orientation, distance to Roads, and distance to RiversYibin City, Sichuan Province, ChinaGeological hazard investigation and zoning report, Land use survey data, Transportation, hydrogeology, socio-economic data118 towns and 54 townshipsRF: AUC-ROC = 0.8776Holdout/test dataRF-AHP framework supported safer disaster-related site assessment.Wang et al. [189]Water managementSVM, RF, and gradient augmented tree, AHP, ELECTRE, TOPSIShydrological, geological, topographical, environmental, water quality, socioeconomicGlobal reviewDEM, Ministry of Agriculture and Water Resources1254 initial recordsRandom Forest demonstrated the highest classification accuracy among the evaluated models.Literature synthesis (no explicit validation)GIS-MCDM-ML integration showed potential for dam-siting research.Note: FAHP: Fuzzy Analytic Hierarchy Process; SOM: Self-Organizing Map; RF: Random Forest; MLP: Multi-Layer Perceptron; CFS: Correlation-Based Feature Selection; WLC: Weighted Linear Combination; FCE: Fuzzy Comprehensive Evaluation; FIS: Fuzzy Inference System; SA-ConvLSTM: Self-Attention Convolutional Long Short-Term Memory; SVM: Support Vector Machine; SHAP: SHapley Additive exPlanations; LIME: Local Interpretable Model-Agnostic Explanations; ARAS: Additive Ratio Assessment; EDAS: Evaluation based on Distance from Average Solution; MOORA: Multi-Objective Optimization by Ratio Analysis; BORDA: Borda Count Method; AUC: Area Under the Curve; RMSE: Root Mean Square Error; MAE: Mean Absolute Error; MSE: Mean Squared Error; R²: Coefficient of Determination; CH Index: Calinski–Harabasz Index; Kappa: Cohen’s Kappa; AUC-ROC: Area Under the Receiver Operating Characteristic Curve. Summary of representative hybrid methods (MCDM+ML) in site selection. Note: FAHP: Fuzzy Analytic Hierarchy Process; SOM: Self-Organizing Map; RF: Random Forest; MLP: Multi-Layer Perceptron; CFS: Correlation-Based Feature Selection; WLC: Weighted Linear Combination; FCE: Fuzzy Comprehensive Evaluation; FIS: Fuzzy Inference System; SA-ConvLSTM: Self-Attention Convolutional Long Short-Term Memory; SVM: Support Vector Machine; SHAP: SHapley Additive exPlanations; LIME: Local Interpretable Model-Agnostic Explanations; ARAS: Additive Ratio Assessment; EDAS: Evaluation based on Distance from Average Solution; MOORA: Multi-Objective Optimization by Ratio Analysis; BORDA: Borda Count Method; AUC: Area Under the Curve; RMSE: Root Mean Square Error; MAE: Mean Absolute Error; MSE: Mean Squared Error; R²: Coefficient of Determination; CH Index: Calinski–Harabasz Index; Kappa: Cohen’s Kappa; AUC-ROC: Area Under the Receiver Operating Characteristic Curve. Table 4 presents the reported outcomes of hybrid models within the individual study contexts reviewed. For example, in the water management domain, AHP+RF frameworks achieved an ROC where only 0.33% of land was classified as suitable [180]. In the landfill siting study, the integration of FAHP, Self-Organizing Map (SOM), and Explainable AI improved transparency by making the model outputs easier to interpret through SHAP and LIME explanations [179]. In the energy domain, an AHP + MLP + Explainable AI framework was reported to improve photovoltaic suitability mapping and reduce weight estimation errors by 11.8% in that specific case [184]. Likewise, MCDM + SVM hybrid models achieved reported accuracies of 85–89% across different regions in the Philippines [187].In healthcare planning, an AHP + MLP framework showed a reported R² of 0.99 with 80–83% accuracy, suggesting a high level of agreement between expert-based and machine learning outputs in that study [182]. These findings highlight the use of hybridization frameworks in high-stakes siting applications across the reviewed studies. In this review, a study was considered hybrid (MCDM+ML) when both a multi-criteria decision-making method and a machine learning algorithm were present and combined within the same analytical framework to support the site selection process. This typically involved using MCDM techniques for tasks such as criteria weighting or ranking, alongside ML algorithms for prediction, classification, or optimization within the same study. Studies that only compared MCDM and ML approaches without actually combining them in a single framework were not classified as hybrid. Findings This section shows the findings from this review based on the four research questions. This analysis combines the insights from the literature on site selection using MCDM, ML, and hybrid approaches, as defined in Section 2.3, from 2015 to 2025. Tables, charts, and figures are used to present the findings. Among the 189 included studies, energy generation was the most frequently represented domain with 70 studies, followed by waste management with 34 studies, logistics and infrastructure with 29 studies, public services with 17 studies, emergency, safety and disaster with 16 studies, water management with 16 studies, and retail, commercial and leisure with 7 studies. These counts were derived from the domain classification process described in Section 2.7, where each study was assigned to one primary domain only. RQ1. In which domains have site selection using MCDM and/or ML techniques been applied between 2015 and 2025? (Context) Between 2015 and 2025, site selection research using MCDM and ML has covered seven domains, like Energy generation, Logistics and infrastructure, Emergency, safety and disaster, Water management, Waste management, Retail, commercial & leisure, and public services. The energy sector includes applications in solar PV, wind farms, hydropower, hydroelectric, nuclear plants, and energy storage systems. Studies such as [165,162,176] shows how ML and hybrid frameworks have been applied to balance technical, environmental, and economic key criteria in energy planning. Beyond energy domain, applications are extended to public services including hospitals, libraries, and elderly care facilities. Common criteria include accessibility, population density, and equity considerations. In Retail and commercial siting, applications include restaurants, ATMs, and tobacco shops. This domain has become an active research area, where consumer behaviour data and geospatial analytics are used as key criteria. Waste management applications include landfills, sewage treatment plants, and disaster shelters. Several reviewed studies suggest that hybrid methods can help balance environmental risk with urban growth. Water management [45,77] has also explored site selection. Finally, Emergency safety and disaster response, such as emergency operations site selection [67], safety zone site selection [85] and emergency shelter homes [164]. These studies show that the methods used in this domain can also be adapted for rapid response conditions. In summary, we can say that use of MCDM and ML has grown from energy generation to waste management and also covers infrastructure planning to business competitive areas like logistics and infrastructure [74]. It is showing the capability of these hybrid methods in several domains. In this review, seven domains are classified as Energy generation, public services, Logistics and infrastructure, Retail and commercial & leisure, Water management, Emergency, safety and disaster, and lastly Waste management. Fig. 2 shows the overall distribution of problem domains addressed in site selection studies between 2015 and 2025, indicating that Energy generation is the most frequently studied domain. Fig. 3 presents the year-wise distribution of problem domains and highlights increased research activity in energy, logistics and infrastructure, and public services during 2023 and 2024. RQ2. Which specific MCDM and ML methods have been used for site selection? (Intervention) Across the studies reviewed, the Analytic Hierarchy Process (AHP) remains the most commonly used multi-criteria decision-making (MCDM) technique. It is often combined with fuzzy logic (FAHP) to deal with uncertainty. Other MCDM methods include TOPSIS, PROMETHEE, ELECTRE, MOORA, and Weighted Linear Combination (WLC). These methods are used because they are relatively clear to interpret and can add expert judgment in a structured way. As a result, they are mostly applied in domains such as healthcare, waste management, and urban planning. In machine learning (ML), tree-based algorithms such as Random Forest, XGBoost, and LightGBM are used. Several studies reported that they perform well on geospatial datasets and can handle complex relationships in the data. Neural network models, including MLP, CNN, ConvLSTM, and SA-ConvLSTM, are often applied in domains such as energy generation and logistics and infrastructure, where nonlinear patterns and spatiotemporal forecasting are important. Support Vector Machines (SVM) and clustering methods such as K-means, DBSCAN, and SOM are also used in several studies. This is particularly for classification and exploratory analysis. Hybrid frameworks that combine MCDM with ML have also been applied in site selection problems. Examples from the reviewed literature are AHP-RF for water management [180], FAHP-SOM-XAI for landfill siting [179], and WLC-FIS-SA-ConvLSTM for photovoltaic siting [186]. These combinations are useful because they combine structured expert-based methods and data-driven predictive modelling within the same decision workflow. Fig. 4 shows the year-wise use of MCDM methods in site selection studies, where AHP/ANP appears most frequently across the review period, while increased use of TOPSIS is shown in 2024 and 2025. Other methods, such as WLC and ELECTRE, are also widely used. Fig. 5 shows the distribution of machine learning model families applied in site selection studies between 2015 and 2025. Tree-based models, particularly Random Forest, are the most frequently used approaches across the reviewed literature. Neural network models (ANN, MLP, and DNN) represent the second most commonly applied family, followed by clustering methods. Support Vector Machines, optimization algorithms, and regression-based models are also used. Fig. 6 presents the use of hybrid MCDM+ML methods in site selection studies (2015–2025), with AHP-based hybrids being the most frequently used, followed by WLC-based hybrids. RQ3. What factors influence the selection of these methods for different application areas? (Mechanism) The choice of method is mostly affected by data availability, decision context, and specific needs. In data-rich, domains such as renewable energy, where large volume of climate and geospatial data is available, machine learning methods such as RF, XGBoost, and CNNs are used. They are used due to their predictive strength and ability to capture nonlinear relationships. On the other hand, in situations where stakeholder trust and interpretability are crucial, such as in healthcare, waste management, and disaster planning, multi-criteria decision-making methods like AHP and TOPSIS are more common. In landfill siting, FAHP is used to give structured weighting, while machine learning is applied to refine and validate the predictions. Deep learning is used to forecast demand patterns, while fuzzy inference systems translate these outputs into practical suitability maps in urban PV siting. Additional factors also influence method choice. For example, regulatory frameworks want researchers to use transparent MCDM approaches in public policy settings. In contrast, when sufficient data and computational resources are available, researchers tend to adopt machine learning methods. In the studies where geographical criteria play an important role, often rely on MCDM methods due to their interpretability. On the other hand, national-level studies that involve large datasets tend to use machine learning methods due to their scalability and ability to handle complex data. Overall method selection shows a balance between data complexity, stakeholder requirements, and decision-making needs. RQ4. What are the commonly used criteria across different domains, and how do these methods enhance decision-making? (Outcome) Several common criteria are identified across studies despite domain-specific variations. •Environmental: slope, elevation, geology, hydrology, land cover, biodiversity, hazard zones. Environmental: slope, elevation, geology, hydrology, land cover, biodiversity, hazard zones. •Socio-economic: population density, land price, accessibility, demand, social acceptance. Socio-economic: population density, land price, accessibility, demand, social acceptance. •Infrastructure: proximity to roads, railways, airports, substations, transmission lines. Infrastructure: proximity to roads, railways, airports, substations, transmission lines. •Risk & sustainability: environmental regulations, disaster vulnerability, resettlement impacts. Risk & sustainability: environmental regulations, disaster vulnerability, resettlement impacts. MCDM methods help to make choices by systematically structuring a decision problem and to balance competing goals and priorities. These methods also make the process clear and understandable to policymakers and stakeholders. ML methods are used to improve predictive accuracy, managing large datasets, and identifying hidden patterns. Hybrid methods (MCDM+ML) may reduce bias, increase robustness under uncertainty, and provide well-explained data-driven results. For example, SHAP and LIME explanations in landfill siting [179] demonstrate how explainable AI tools can support more informed decision -making. The integration of MCDM and ML has changed site selection from a subjective, expert-based process to a systematic, evidence-based, and reproducible solution. This transformation has definitely improved the effectiveness of spatial planning decisions as site selection across different contexts. Although existing studies have explored various methods but still several gaps are present across domains. To address these gaps, this review proposes a framework based on three connected pillars: benchmarking, explainability, and validation. Proposed conceptual framework: benchmarking-explainability-validation In this review, we introduced a framework based on three components: benchmarking, explainability, and validation. These stages are supposed to be applied in sequential manner in a practical site selection workflow. Benchmarking creates common grounds for geo-spatial repositories and criteria so that results across different domains (e.g. energy generation, logistics and infrastructure, and water management) can be easily compared. This stage involves collecting, cleaning, normalizing, and spatially aligning datasets before modelling begins in practice. Without benchmarking, hybrid models may lack transferability across different contexts. Explainability enhances trust and accountability by making model outputs transparent and understandable to stakeholders. After benchmarking, the hybrid model is applied, and its outputs are interpreted. Decision makers need to understand why a site is selected, which is supported through explainability tools such as SHAP, LIME, along with MCDM-based weighting methods. Among the reviewed studies, SHAP was the most frequently reported explainability technique and was used in five studies [162,165,166,179,184]. LIME appeared less frequently and was reported in only one study, where it was used together with SHAP in landfill siting studies [179]. Validation ensures reliability, scalability and transferability, particularly in high-stakes applications such as hospital siting, energy plants, and business location planning. At this stage, model outputs are evaluated using appropriate performance metrics and robustness checks. Many existing site selection models are tested within a single region or context. In hybrid models, MCDM components should be tested for robustness under uncertainty, while ML components should be validated across regions to reduce overfitting. Benchmarking was included because many studies used different datasets, criteria set, and validation methods, making direct comparison difficult. Explainability was included because only a few studies used tools such as SHAP and LIME, even though transparency is important in site selection decisions. Validation was included because many studies were tested only in a single region or context, with limited cross-regional or real-world validation. These components together provide a pathway for developing a reliable, transparent, and adaptable decision model for site selection. Table 5 summarizes the implementation stages of BEV Framework. Table 5Implementation process of the BEV framework.Table 5 dummy alt textStagesFocusKey activitiesValidation toolStage 1: BenchmarkingData preparationCollect geospatial datasets, clean and normalize, and align spatiallyEnsures consistency and comparabilityStage 2: ExplainabilityInterpretabilityApply hybrid MCDM–ML models; use AHP/FAHP for weighting; SHAP/LIME for ML interpretationImproves transparency and stakeholder trustStage 3: ValidationReliability & scalabilityCalculate RMSE, MAE, R², AUC; perform cross-validation, regional transferability checks, uncertainty analysisConfirms accuracy and generalizability Implementation process of the BEV framework. Fig. 7 presents the BEV framework, identifying methodological gaps, practical solutions, and potential directions for future research in transparent and scalable site selection. Proposed conceptual framework in hybrid (ML+MCDM) site selection. From the existing study included in this paper, the landfill site selection study by Awadh and Mallick [179] is considered: •Benchmarking: Integration of datasets such as ALOS PALSAR DEM, Sentinel-2 imagery, rainfall data, and ground observations. Benchmarking: Integration of datasets such as ALOS PALSAR DEM, Sentinel-2 imagery, rainfall data, and ground observations. •Explainability: FAHP weighting was combined with SOM clustering and ensemble learning, while SHAP and LIME were used to interpret suitability scores. Explainability: FAHP weighting was combined with SOM clustering and ensemble learning, while SHAP and LIME were used to interpret suitability scores. •Validation: Performance was evaluated using MAE, MSE, RMSE, and R². Validation: Performance was evaluated using MAE, MSE, RMSE, and R². The bagging ensemble method shows an R² value of 0.882. This example shows the sequential application of the BEV Framework. First, benchmarking ensures integration of datasets. Second, explainability supports the interpretation of model output, and third, validation confirms model performance and generalizability. This workflow process shows a path for developing robust, transparent, and transferable decision support for site selection. Applied relevance of the framework Several outcomes are highlighted for different stakeholder groups from 189 studies: For Policymakers, it supports the design of evidence-based and reproducible policies. It also enables transparent and accountable decision-making across domains. Benchmarking ensures comparability, explainability supports defensible decisions, and validation reduces bias and uncertainty across regions. For Urban Planners, it helps to balance technical feasibility with community needs. It also used in decision making such as emergency planning and hospital siting. For Industry and Developers, it reduces uncertainty in high-cost projects. It supports profitability with sustainability and ensures that the models are transferable and applied to standardized datasets. For Cross-Sectoral Practice, Benchmarking reduces the duplication of effort, and explainability makes output comparable and ensures scalability. It enables tools to be applied across domains with broader applicability. Overall, this proposed framework actually bridges academic innovation with real-world decision-making. This framework offers a tool to stakeholder groups that is transparent, reproducible, and transferable. Discussion This review synthesis is based on 189 studies to find how site selection research has evolved and also to highlight the existing challenges. Traditional MCDM methods are preferred for their transparency and stakeholder legitimacy, while ML methods have gained popularity for handling large and complex datasets. Hybrid models, discussed earlier in Section 2.3, combine interpretability with predictive strength and are also capable of handling complex decisions. This development shows that method selection depends not only on data availability and governance needs, but also on the decision context and the nature of the criteria involved. The discussion is organized around three themes: Methodological evolution, Domain-specific adaptations, and Criteria convergence. These themes show how site selection research is moving towards more transferable, integrated, and stakeholder-oriented solutions. Methodological evolution Classical MCDM methods such as AHP, TOPSIS, and PROMETHEE remain widely used due to their transparency and acceptance among stakeholders. For large and heterogeneous datasets, ML techniques such as Random Forest, XGBoost, and neural networks are mostly applied for their predictive accuracy. A recent development is the emergence of hybrid frameworks (e.g., AHP+RF, FAHP+SOM+XAI, AHP+FIS+SA-ConvLSTM). These frameworks combine interpretability with predictive strength and support explainable, interdisciplinary decision support in the field of site selection. The continuous use of AHP and TOPSIS reflects their simplicity and familiarity, but may also indicate limited adoption of newer methods in some contexts. In the case of large labelled datasets, ML methods provide better predictive accuracy, which sometimes can be a problem in data-scarce environment. Hybrid methods offer a balanced solution. However, their complexity may limit practical adoption unless supported by clear and repeatable workflows. Fig. 8 shows the methodological growth in site selection research between 2015 and 2025. The earlier studies were dominated by classical MCDM approaches, which reached their peak around 2017. Machine learning methods were less explored until 2021. They were rapidly explored after that, which reflects the focus on using predictive accuracy and data-driven validation for site selection. Methodological evolution. Hybrid (MCDM+ML) approaches appear less frequently in the literature. These methods still represent an emerging and underexplored integrated framework. This evolution shows a shift from conventional decision-support to information-rich models. The hybrid approach offers a balance between transparency, reduced bias, and predictive capability within specific application contexts. Domain-specific adaptations Method selection is strongly influenced by application context. Energy planning often favors ML and hybrid approaches due to the availability of geospatial and climatic data. Healthcare and public services rely on MCDM methods to ensure transparency and policy alignment. Waste management and disaster planning studies also adopt hybrid approaches to balance uncertainty with transparency. In retail and commercial siting, clustering and predictive ML methods are used to capture consumer behavior and demand patterns. Strategic and security-sensitive domains give importance to robustness and integration of multiple criteria. These patterns confirm that method selection depends on data richness or availability, decision stakes, and governance requirements. Fig. 9 shows the distribution of site selection across the seven domains discussed in this review. Energy generation is the most dominant domain and reflects the maturity of MCDM and hybrid approaches in balancing technical and environmental criteria. Public services, logistics and infrastructure follow, where method selection is influenced by transparency and governance requirements. Waste and water management appear moderately and often apply fuzzy or statistical hybrid methods to address uncertainty. Emergency, safety, & disaster related studies are fewer in number but remain highly significant. Retail, commercial & leisure are less explored and indicate an opportunity for ML-enhanced models. Domain distribution tree map. Fig. 10 presents a Sankey diagram showing how the seven domains are connected to three methodologies: MCDM, ML, and a hybrid (MCDM+ML) approach. Energy generation and public services rely heavily on MCDM methods. It indicates dependence on well-structured and transparent evaluation methods in this domain. The logistics and infrastructure domain is distributed across all three methods, reflecting the need for transparency, prediction, and methodological integration. Retail and emergency domains show greater use of ML, which reflects the demand for rapid, data-driven decisions in these domains. Waste and water management show smaller flows and often combine MCDM and hybrid (MCDM+ML) approaches. Overall, the Sankey diagram highlights that methodological adaptation varies across domains. More established sectors tend to rely on MCDM, and emerging contexts adopt ML, and hybridization serves as a bridging approach across applications. Methodological distribution across application domains in site selection studies (2015–2025). Criteria convergence Across the seven domains, the most common key criteria include environmental factors (slope, elevation, geology, hydrology), socio-economic factors (population density, land price, accessibility), infrastructure factors (roads, railways, substations), and sustainability/risk-related factors (hazard zones, biodiversity, regulations). The pattern suggests the possibility of developing more generalizable frameworks, domain-specific factors such as wind speed in energy applications and groundwater depth in waste management still require contextual customization. However, many studies choose criteria based on data availability instead of what is actually required for a specific domain. This practice may reduce consistency across domains and introduce local biases into otherwise generalizable model. Fig. 11 illustrates how different criteria converge across the seven domains. Environmental and infrastructure-related factors dominate energy and logistics applications. Socio-economic factors are more prominent in public services and retail, where community and market needs are central. Sustainability and risk criteria are more balanced, particularly in water and waste management. Overall, the radar chart indicates that each domains emphasize criteria differently, while also suggesting that hybrid approaches can help balance technical, social, and risk related dimensions. Criteria convergence radar chart. Gaps and challenges Explainability in ML remains underutilized, despite the growing use of tools such as SHAP and LIME. Cross-domain validation remains limited, raising concerns about the transferability of many existing models. In addition, socio-political and behavioural factors are often overlooked, even though they strongly influence real-world siting decisions. Without standardized datasets, reproducible pipelines, and inclusive integration strategies, the full potential of hybrid frameworks may not be fully realized. Positioning against prior reviews Earlier existing reviews have provided valuable insights, but their scope has generally been limited. Most have focused on specific application domains, offering only a partial view of methodological developments across domains. This review contributes to the literature in three ways. First, it brings together MCDM, ML, and hybrid approaches across seven domains. This enables the identification of common criteria and methodological patterns across domains that have not been examined in earlier reviews. Second, it proposes a framework based on benchmarking, explainability, and validation. This helps organize the key methodological gaps into a research agenda. Third, this study highlights visual tools such as Sankey diagrams, radar charts, and Domain distribution tree maps. These visual representations improve the clarity and interpretability of the findings. Compared with previous reviews, this study shows that the site selection has progressed beyond isolated case studies. It also demonstrates the value of cross-domain frameworks for advancing site-selection research in both theory and practice. Benchmarking and cross-regional validation One of the main issues identified across the existing studies is the absence of systematic benchmarking and cross-regional validation. Most models rely on locally sourced datasets, making their results difficult to compare across studies and contexts. These problems show the relevance of the benchmarking, explainability, and validation (BEV) framework introduced earlier. The framework organizes current gaps within a clear research roadmap rather than proposing an entirely new method. The review shows how future methodological improvements can support more credible, transferable, and user-friendly decision-support systems using hybrid methods by positioning these limitations within this framework. Benchmarking is preferable when studies rely on comparable datasets, evaluation metrics, and validation approaches. The reviewed literature shows considerable variation in how models are assessed. The validation methods used are simple holdout splits, cross-validation, spatial validation, and expert-based evaluation. Results were also compared with existing sites or historical data in some studies. Performance is reported using different metrics such as accuracy, RMSE, MAE, or AUC. Direct comparison across studies is actually difficult due to these differences. Validation practices also varied considerably across the reviewed ML and hybrid studies. Among the 35 studies summarized in Table 3, Table 4, holdout validation was the most commonly reported approach (10 studies), followed by spatial validation (8 studies). Mixed validation approaches, expert or rule-based validation, and cluster validation were reported in 4, 3, and 3 studies, respectively, while cross-validation was reported in 3 studies. Simulation-based validation appeared in 2 studies, whereas correlation-based validation and literature synthesis without explicit validation were each reported in 1 study. To avoid double counting, studies that used more than one validation approach were grouped under a mixed-validation category. The wide variation in validation practices highlights the lack of a common benchmarking framework and makes direct comparison of reported performance across studies difficult. This highlights the need for more consistent benchmarking and validation practices in future site selection research. This also supports the significance of the proposed framework for improving comparability and reliability across studies. Practical implications Making standardized criteria across domains can reduce repetitive efforts and contribute to the development of reusable frameworks. At the stage of implementation, the hybrid (ML+MCDM) methods can support infrastructure development decisions that are moretransparent, responsible, and effective for stakeholders. Finally, the importance of explainability and validation is highlighted by this review. Even robust models may face limited adoption if their outputs are not clearly interpreted or validated across different regions. Future site-selection tools should incorporate explainable AI and reproducible validation procedures to improve their credibility and wider adoption. Conclusion Summary of contributions This review provides a comprehensive overview of site-selection studies that combine multi-criteria decision-making (MCDM), machine learning (ML), and hybrid methods, discussed earlier in Section 2.3. The 189 selected studies show the evolution of methods, identify common decision criteria, and highlight the increasing importance of hybrid frameworks for site selection. Limitations The review is limited by the databases searched and the different reporting standards used in the included studies. As noted in Section 2.7, restrictions related to language, publication type, and exclusion criteria may have led to the omission of some relevant studies. Final remarks This study shows how merging MCDM and ML as hybrid methods has influenced the site-selection research in recent years. The study highlights the growing need for hybrid, explainable, and data-rich frameworks to support scalable and transferable decision support. The study provides background for developing site-selection tools that are both scientifically sound and practically relevant. The following section of this paper outlines specific research directions for this agenda. For stakeholders and policymakers, this study proposes a framework that can provide sustainable, fair, and optimal site selection planning outcomes. Future research directions This review identifies several gaps and proposes directions for future work to support the development of more robust and applicable site selection frameworks. Future frameworks should aim to balance methodological rigor, technical advancements, and socio-political considerations. Addressing these gaps may strengthen both the scientific value and practical usefulness of future research for policymakers and stakeholders. •Systematic comparison of methods: Future studies should compare MCDM, ML, and hybrid frameworks to examine their relative strengths in terms of accuracy, interpretability, and scalability. Systematic comparison of methods: Future studies should compare MCDM, ML, and hybrid frameworks to examine their relative strengths in terms of accuracy, interpretability, and scalability. •Benchmark datasets: There is a need to develop standardized and openly available geospatial and socio-economic datasets to support cross-regional validation and fair model comparison. Benchmark datasets: There is a need to develop standardized and openly available geospatial and socio-economic datasets to support cross-regional validation and fair model comparison. •Explainable AI integration: Tools such as SHAP and LIME can improve the transparency and trustworthiness of hybrid frameworks for stakeholders. Explainable AI integration: Tools such as SHAP and LIME can improve the transparency and trustworthiness of hybrid frameworks for stakeholders. •Cross-regional validation: Testing models across diverse regions and domains can improve understanding of their robustness, scalability, and transferability under varied conditions. Cross-regional validation: Testing models across diverse regions and domains can improve understanding of their robustness, scalability, and transferability under varied conditions. •Dynamic decision systems: Leveraging IoT, remote sensing, and real-time data can enable site-selection frameworks to adapt more effectively to changing environments. Dynamic decision systems: Leveraging IoT, remote sensing, and real-time data can enable site-selection frameworks to adapt more effectively to changing environments. •Socio-political integration: Community concerns like health, traffic, fairness, policy priorities, and governance rules are included. This makes the outcome more acceptable with the help of hybrid approaches. Socio-political integration: Community concerns like health, traffic, fairness, policy priorities, and governance rules are included. This makes the outcome more acceptable with the help of hybrid approaches. •Reproducible pipelines: Developing complete and reproducible workflows can enhance both academic rigor and practical implementation. Reproducible pipelines: Developing complete and reproducible workflows can enhance both academic rigor and practical implementation. Ethics statements No data was used in this research. Funding No Funding was used in this research. CRediT authorship contribution statement Disha Upadhyay: Writing – original draft, Methodology, Software, Validation, Data curation, Writing – review & editing. Anuj Kumar: Conceptualization, Methodology, Writing – original draft, Supervision, Writing – review & editing. Sangeeta Pant: Conceptualization, Methodology, Writing – original draft, Supervision, Writing – review & editing. Ketan Kotecha: Supervision, Writing – review & editing. Ajith Abraham: Supervision, Writing – review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.mex.2026.104007. Contributor Information Anuj Kumar, Email: anuj.kumar@dypiu.ac.in. Sangeeta Pant, Email: sangeeta.pant@sitpune.edu.in. Supplementary materials References Associated Data Supplementary Materials