Integrating AI and Geospatial Technologies for Sustainable Smart City Development: A Case Study of Yerevan

Abstract

Urban growth and environmental pressures in rapidly transforming cities require innovative governance tools that integrate advanced technologies with institutional assessment. This study develops and applies a strategic integration framework that combines spatial analysis, Convolutional Neural Networks (CNNs)-based land-use classification, SHAP-based feature attribution, and stakeholder interviews to evaluate Yerevan, Armenia, as a case of a mid-income city facing accelerated urbanization. The case selection is justified by Yerevan’s rapid built-up expansion, fragmented green areas, and institutional challenges in aligning urban development with sustainability goals. The CNN model achieved 92.4% accuracy in land-use classification, and projections under a business-as-usual scenario indicate a 12.8% increase in built-up areas and a 6.5% decline in green zones by 2030. SHAP analysis identified land surface temperature and NDVI as the most influential predictors, while governance interviews highlighted gaps in regulatory support and technical capacity. The proposed framework advances the literature by integrating AI-driven geospatial analysis with qualitative governance assessment, providing actionable insights for urban policymakers. Findings underscore the potential of combining machine learning, geospatial technologies, and institutional diagnostics to guide smart city planning in transition economies.

1. Introduction

Sustainable urban development has emerged as one of the most pressing global challenges of the 21st century, particularly in light of rapid urbanization, resource depletion, and environmental degradation. The United Nations’ Sustainable Development Goal 11 (SDG11) outlines a critical pathway for achieving inclusive, safe, resilient, and sustainable cities. Achieving this goal requires not only policy innovation and governance reform, but also the integration of advanced technologies to support evidence-based planning and adaptive management of land and environmental resources.

In response to this demand, the smart city paradigm has gained considerable traction, leveraging technologies such as artificial intelligence (AI), geospatial information systems (GIS), remote sensing, and real-time data analytics to transform the way cities are planned, managed, and governed. Smart cities aim to optimize urban systems and infrastructures, enhance public services, reduce environmental impacts, and foster participatory governance. The convergence of digital innovation with urban sustainability provides a powerful framework for addressing complex land-use dynamics, climate risks, and socio-economic inequalities [1,2,3].

In this context, land resource and environmental management have become pivotal domains within the smart city agenda. Urban land use is both a driver and a consequence of sustainability outcomes. Poorly managed urban expansion can lead to habitat fragmentation, resource inefficiency, and increased vulnerability to climate-related hazards. Conversely, integrated and data-informed land management systems can help balance ecological protection, economic development, and human well-being. Recent advancements in AI and GIS technologies have provided urban planners and environmental managers with unprecedented tools for predictive modeling, risk assessment, and strategic intervention [4,5]. AI techniques such as machine learning and deep learning are increasingly used to model land-use change, forecast environmental degradation, and optimize zoning scenarios. When coupled with GIS and remote sensing, these approaches enable spatially explicit simulations that inform more sustainable land development policies [6,7].

The combination of AI and GIS is particularly relevant for cities in transition—such as those in post-Soviet and Global South contexts—where rapid development is often accompanied by governance fragmentation, infrastructural gaps, and institutional inertia. These urban environments require not only innovative tools but also adaptable models that can work within the constraints of limited technical capacity and regulatory challenges. In such settings, AI-GIS integration offers a way to overcome traditional limitations in land-use planning by generating timely, location-specific, and actionable insights [8,9,10,11].

Yerevan, the capital of Armenia, presents an illustrative case of these dynamics. As an emerging smart city in the South Caucasus region, Yerevan faces the dual challenge of modernizing its urban management systems while addressing critical sustainability issues such as land-use fragmentation, air pollution, and urban heat islands. Despite recent investments in digital governance platforms and environmental monitoring, the city still lacks an integrated framework that can harmonize technological potential with strategic land and environmental governance. The limited coordination between data systems, absence of real-time land-use analytics, and fragmented stakeholder involvement hinder the city’s ability to meet SDG11 targets.

This paper addresses this gap by proposing an integrated framework for sustainable land and environmental management in Yerevan, grounded in the combined use of artificial intelligence and geospatial technologies.

This study pursues three primary objectives: (1) to assess the applicability of AI-GIS tools in land-use modeling and environmental risk detection; (2) to evaluate the governance readiness and data infrastructure of Yerevan for smart urban planning; and (3) to design a strategic, scalable framework that can inform sustainable development pathways in post-Soviet cities.

The study employs a mixed-methods approach, including spatial analysis, environmental risk zoning, and institutional assessment, supported by case-specific simulations and stakeholder feedback. The research draws from multiple open and institutional datasets and relies on machine learning models and remote sensing platforms to simulate various land-use and environmental management scenarios. In this framework, spatial analysis serves as the foundation, but it is systematically extended with AI-based geospatial methods (e.g., CNN classification, SHAP attribution) to improve land-use mapping and risk evaluation.

While the concepts of sustainable urban development and smart cities are well-established in the literature, there remains limited research that integrates advanced AI methods with GIS-based spatial analysis and participatory governance frameworks. This paper contributes to bridging this gap by developing a multi-method approach that combines (i) CNN architectures such as U-Net and DeepLabv3+ for land-cover segmentation, (ii) predictive modeling using Random Forest and XGBoost, and (iii) NLP and LLM-based tools for analyzing citizen feedback and supporting participatory governance.

Yerevan is selected as a case study not only due to its rapid urban densification, complex terrain, and post-Soviet governance transitions but also because these characteristics make it representative of many medium-sized emerging cities. The methodological framework proposed here is therefore generalizable beyond Yerevan, offering lessons for other contexts.

By clarifying the scope of ‘machine learning’ (tree-based predictive models) and ‘deep learning’ (CNN architectures for image segmentation) and linking AI methods to distinct governance roles, this study advances both the technical and governance-oriented strands of smart city research. The research gap it addresses lies in the underexplored integration of AI + GIS methods with participatory and citizen-centric governance frameworks.

By situating the Yerevan case within broader global and regional debates on urban sustainability and digital transformation, this paper contributes both theoretical and practical insights. It aims to enrich the discourse on how smart city technologies can be deployed to enhance environmental governance and spatial justice in cities that are still navigating the complex legacies of centralization and transition.

The rest of this paper is structured as follows: Section 2 presents a literature review; Section 3 outlines the research methodology and data sources; Section 4 presents the results of the AI-GIS simulations and governance analysis; Section 5 discusses the implications of the findings in light of international literature and best practices; and Section 6 concludes with policy recommendations and directions for future research.

2. Literature Review

2.1. Smart Cities and SDG11: Theoretical Context

The concept of smart cities has rapidly evolved into a cornerstone of contemporary urban policy, strongly aligned with the goals outlined in the United Nations’ Sustainable Development Goal 11 (SDG11), which aims to “make cities inclusive, safe, resilient, and sustainable.” Smart cities employ advanced digital technologies and data-driven systems to optimize urban management, improve service delivery, and enhance the quality of life for citizens [1,2,12,13]. At the heart of this transformation lies the integration of technological infrastructures—such as the Internet of Things (IoT), artificial intelligence (AI), and spatial intelligence—with participatory governance frameworks [4,14,15,16].

SDG11 calls for innovative approaches to urban governance that transcend conventional planning paradigms. Researchers argue that smart cities are not solely about technology adoption but about how technology mediates spatial justice, sustainability, and citizen empowerment [17,18]. The transition toward smart urbanism is seen as a systemic transformation—one that redefines land use, transport systems, energy consumption, and environmental protection mechanisms [19,20,21,22].

The implementation of smart city principles in the Global South and post-Soviet regions, including Armenia, presents specific institutional and socio-political challenges [23,24,25]. Scholars underscore the importance of local context, governance readiness, and adaptive policy frameworks in shaping how SDG11 is localized and operationalized [26,27,28]. Therefore, the theoretical discourse on smart cities must be understood as both a global technological movement and a site-specific governance challenge.

Recent studies emphasize that smart cities are also human-centric systems, where digital participation and e-governance tools—such as participatory mapping, civic dashboards, and online consultation platforms—enable citizens to shape urban policies [17,29]. Cases from Barcelona, Helsinki, and Singapore demonstrate how civic-tech applications enhance inclusivity, while in Armenia early steps like the e-cadaster remain fragmented. The gap lies in linking advanced AI-GIS analysis with such participatory frameworks, which this study addresses by integrating technological innovation with citizen-centric governance.

Examples from Barcelona, Helsinki, and Singapore further demonstrate how digital participation has been institutionalized as a form of urban interventionism and responsive governance. In Barcelona, civic dashboards and the “Decidim.Barcelona” participatory platform have enabled citizens to co-design zoning reforms and influence mobility strategies, embedding digital democracy in everyday urban policy-making [30,31]. Helsinki’s “Decidim.Helsinki” initiative highlights how open-source digital democracy tools integrate citizen feedback into long-term sustainability planning and climate adaptation policies [32]. Singapore’s Smart Nation program illustrates how e-governance portals and real-time participatory applications link citizen input with infrastructure adaptation, creating a more responsive and adaptive city system [33]. These cases collectively underline the role of digital participation as a driver of responsive cities, offering a conceptual bridge to the governance challenges observed in Yerevan.

2.2. AI Applications in Land and Environmental Management

Artificial Intelligence (AI) is increasingly recognized as a transformative tool in the domain of land and environmental management. The proliferation of machine learning algorithms, computer vision techniques, and intelligent data analytics has significantly expanded the analytical capacities of urban planners and environmental scientists. As urban areas grow more complex and data-rich, AI enables the automation and enhancement of decision-making processes, offering unprecedented precision and predictive power in modeling land-use dynamics and environmental changes [16,34,35].

Recent literature illustrates how NLP and LLM-based approaches, including topic modeling and sentiment analysis, have been widely applied to analyze citizen feedback and urban governance discourse, thereby linking technical outputs with participatory decision-making [36,37]. Similarly, CNN-based architectures such as U-Net and DeepLabv3+ have been successfully employed for urban land-use classification, environmental monitoring, and remote sensing applications, confirming their relevance for data-rich spatial contexts [38]. These studies collectively demonstrate that the methodological choices adopted in this research are not only technically robust but also aligned with broader scholarly practices in AI-driven urban analysis. This study addresses these challenges by presenting AI not only as a technical tool but as a governance enabler, thereby connecting methodological innovation with urban policy needs.

In the context of land-use classification, supervised learning algorithms such as Random Forest, Support Vector Machines (SVMs), and Convolutional Neural Networks (CNNs) have demonstrated high levels of accuracy in processing satellite imagery and other spatial data layers. These models allow for fine-grained detection of land-cover types, spatial zoning patterns, and encroachments in urban expansion zones [28,39]. For instance, CNN-based models trained on multi-temporal remote sensing data can effectively classify built-up areas, green spaces, and water bodies, even in heterogeneous landscapes with low-resolution data [28,35,40].

AI is also pivotal in environmental risk assessment and monitoring. Predictive models can simulate the impacts of urban sprawl on ecosystems, anticipate flood-prone zones, and estimate air quality variations based on meteorological and transport data [41]. When coupled with real-time IoT sensors and open geospatial datasets, these models enhance urban resilience by providing early warnings for extreme weather events and pollution surges [40,42]. Recent studies demonstrate the importance of IoT-enabled environmental sensing platforms that integrate air quality, traffic, and meteorological data streams [43]. Moreover, openly available geospatial datasets provide high-resolution indicators for urban ecological monitoring and land-use dynamics [44]. These resources illustrate the applicability of IoT and geospatial data for real-time risk assessment and urban resilience planning.

Moreover, unsupervised machine learning techniques—such as clustering and anomaly detection—have been successfully applied in identifying urban heat islands, illegal landfills, and degradation hotspots without requiring labeled data [45]. Reinforcement learning approaches are now being explored for adaptive land-use planning, where planning agents learn optimal zoning decisions through iterative simulations [46].

Despite these technological advancements, the practical implementation of AI in environmental governance remains uneven, especially in mid-income and transitioning economies. Challenges include the lack of high-quality training datasets, limited computational infrastructure, and institutional resistance to data-driven governance [47]. Furthermore, algorithmic transparency and interpretability remain crucial concerns in public decision-making, requiring closer collaboration between technologists, planners, and local communities [48].

Nonetheless, case studies in India, China, and Brazil show promising examples where AI has improved ecological restoration planning, urban growth forecasting, and integrated watershed management [49,50,51]. These examples suggest that with appropriate data governance and capacity building, AI can serve as a vital instrument in achieving SDG-aligned urban sustainability.

2.3. GIS and Remote Sensing in Urban Sustainability

Geographic Information Systems (GIS) and remote sensing technologies have become foundational tools for advancing urban sustainability and resilience planning. These spatial technologies provide the ability to capture, analyze, and visualize land-use dynamics, environmental changes, and socio-spatial inequalities at multiple scales. As cities face complex challenges related to uncontrolled urban sprawl, loss of green cover, and climate vulnerability, GIS-based spatial analytics offer data-driven insights for informed decision-making [52,53].

GIS platforms allow urban planners to integrate diverse datasets—such as demographic distributions, infrastructure layouts, vegetation indices, and hazard maps—into georeferenced systems for scenario modeling and resource optimization. These capabilities are especially important in land suitability analysis, zoning regulation, and transportation planning [54,55]. Remote sensing, particularly through high-resolution satellite imagery (e.g., Sentinel, Landsat), enables continuous and cost-effective monitoring of land-cover change, urban expansion patterns, and ecological fragmentation [56].

Existing literature demonstrates that Multi-Criteria Decision Analysis (MCDA) has often been applied in participatory urban governance, allowing stakeholders to weigh trade-offs and collectively evaluate alternative scenarios [57,58,59]. For instance, participatory GIS combined with MCDA has been used to facilitate land-use negotiations, integrate public preferences in zoning decisions, and support inclusive environmental planning [60]. While these studies highlight the value of MCDA in enabling citizen input, our approach differs by embedding MCDA within an AI-GIS workflow that combines CNN-based land-use classification, environmental risk mapping, and NLP/LLM-driven analysis of citizen feedback. This integration allows MCDA results to be directly informed by advanced spatial analytics and participatory data streams, thus moving beyond traditional applications toward a more comprehensive model of responsive governance.

Indices such as the Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), and Soil Adjusted Vegetation Index (SAVI) are commonly derived from satellite data to evaluate the urban microclimate, heat island effects, and ecological health. When applied over time, these indices reveal spatial trends in environmental degradation or regeneration [42,61].

In this study, indices such as NDVI and LST are not presented in isolation but serve as critical inputs for the broader AI-GIS workflow. Their integration with Multi-Criteria Decision Analysis (MCDA) methods, including AHP and TOPSIS, allows environmental monitoring results to be translated into governance-relevant priorities. Existing literature demonstrates that MCDA-GIS approaches are particularly valuable in resource-constrained contexts, where trade-offs must be systematically evaluated under limited data and institutional capacity [62,63]. By extending this framework to include digital participation tools, the study positions MCDA not only as a technical decision-support method but also as a bridge to citizen engagement and participatory governance.

In the context of sustainable development, GIS and remote sensing are instrumental in spatial vulnerability assessments, green infrastructure planning, and ecosystem services valuation. Urban planners increasingly use multi-criteria decision analysis (MCDA) in GIS environments to evaluate trade-offs among competing land uses [62]. Moreover, the integration of remote sensing with socio-economic data supports holistic sustainability assessments, linking environmental indicators to urban equity and livability metrics [5].

Emerging practices also include the fusion of UAV (drone) imagery with traditional satellite data to improve spatial resolution and real-time responsiveness in urban management [64]. These methods are particularly valuable in cities where local governments face resource limitations or fragmented land information systems.

Ultimately, GIS and remote sensing not only enable technical precision but also foster participatory planning through the development of visual tools and dashboards for public engagement. As smart cities evolve, the coupling of geospatial intelligence with other technologies like AI enhances strategic environmental governance.

2.4. Integrated AI-GIS Frameworks: Theory and Practice

The integration of Artificial Intelligence (AI) with Geographic Information Systems (GIS) represents a frontier in spatial decision support systems, offering intelligent, scalable, and adaptive solutions to urban planning and environmental management challenges. While AI brings predictive capabilities, pattern recognition, and automation, GIS provides spatial contextualization, topological analysis, and visualization—together forming a robust framework for managing complexity in land-use systems [8,65].

Integrated AI-GIS frameworks are increasingly employed to model dynamic land-use patterns, monitor environmental risks, and support scenario-based simulations. For instance, decision trees and neural networks embedded within GIS platforms can predict urban expansion based on historical trends, topographic constraints, and policy inputs [66]. These hybrid models have demonstrated superior performance over conventional spatial regression models by learning nonlinear relationships among geospatial variables.

Moreover, integrated systems are well-suited for real-time decision-making in environmental emergencies. AI-powered GIS dashboards can process satellite imagery and sensor data streams to generate alerts for deforestation, flood threats, or pollution spikes [67]. Reinforcement learning algorithms further enhance adaptability by dynamically updating land-use recommendations based on feedback from environmental indicators and citizen reports [35,40].

From a theoretical perspective, integrated frameworks draw upon cybernetic urbanism and intelligent infrastructure paradigms, viewing the city as an adaptive system that learns from its own data flows [68]. This approach aligns with contemporary sustainability science, which emphasizes system-based governance and anticipatory planning [5].

Practical implementations are emerging globally. In India, AI-GIS systems have been used to guide urban village redevelopment; in China, they underpin eco-city planning initiatives; and in the Netherlands, such frameworks inform climate adaptation strategies [69]. These applications demonstrate that success depends not only on technological sophistication but also on data quality, institutional readiness, and policy coherence.

While international cases such as India, China, and the Netherlands illustrate diverse applications of AI-GIS, their key lesson is that successful integration depends not only on technological capacity but also on governance structures and institutional readiness. In Armenia and other post-Soviet contexts, challenges of data governance, transparency, and fragmented institutional arrangements remain significant. This highlights the importance of adapting AI-GIS tools to local realities and using them not only for technical efficiency but also as mechanisms to strengthen accountability and participatory governance.

In Armenia and similar post-Soviet contexts, AI-GIS integration remains nascent, often hindered by fragmented data ecosystems and limited local expertise. However, initiatives such as open geospatial portals, digital cadastral systems, and public-private partnerships indicate growing interest in advancing spatial intelligence. Thus, developing contextualized frameworks that are modular, interoperable, and policy-sensitive is critical for ensuring the long-term viability and impact of AI-GIS solutions in sustainable urban development.

2.5. Case Studies in Emerging and Post-Soviet Cities

The application of AI and GIS technologies in sustainable urban development is gaining momentum across emerging and post-Soviet cities, where governance systems are undergoing transformation and data infrastructures remain fragmented. These cities offer valuable insights into both the challenges and opportunities of implementing intelligent land and environmental management frameworks under conditions of limited institutional capacity and high development pressure [70].

In the South Caucasus, Armenia’s capital Yerevan has taken initial steps toward digital urban governance through the development of electronic cadaster systems, open urban maps, and basic spatial planning platforms. However, most initiatives remain fragmented and lack full integration with predictive analytics or real-time environmental data [71]. Pilot projects using satellite imagery for urban heat island mapping and AI-based risk detection are in early stages, with promising but yet limited policy impact.

Tbilisi, Georgia, has made greater progress in integrating GIS in transportation planning and environmental monitoring. Through donor-supported projects, the city has implemented land-use suitability modeling using GIS and participatory mapping tools [28,29]. Yet, integration with AI remains experimental.

Kyiv, Ukraine, represents a more advanced example of post-Soviet digital governance, deploying AI-powered systems in waste management and urban transport optimization. Kyiv’s Open-Data Portal and “Smart Kyiv” initiative have facilitated integration between geospatial data and municipal services [57].

In South Asia, Indian cities like Pune and Hyderabad have implemented AI-GIS systems for slum redevelopment, flood prediction, and traffic control, largely supported by national smart city missions. These cities illustrate how strong political will and centralized funding can accelerate adoption, though concerns about data governance and inclusivity remain [58,59].

Similar efforts in Latin America—such as São Paulo’s climate-sensitive zoning reforms using GIS-based vulnerability mapping—demonstrate the adaptability of spatial technologies in diverse urban settings [60]. While these cities benefit from technical collaboration and open-data portals, institutional silos and low interoperability remain significant barriers.

These case studies collectively highlight key success factors: strong local leadership, external technical partnerships, public engagement, and policy frameworks that mandate cross-sector data sharing. They also underline the importance of context-sensitive frameworks that align smart urban innovations with local administrative capacity and cultural norms [3,6].

For post-Soviet cities like Yerevan, hybrid governance models—where national strategy meets local experimentation—appear most viable. Investing in interoperable data platforms, fostering public-private-academic partnerships, and building technical literacy among municipal staff are essential steps toward integrated, intelligent urban sustainability governance.

A comparative synthesis of these cases reveals several transferable lessons. First, cities such as Tbilisi and Kyiv demonstrate that incremental steps in digital governance and open-data initiatives can significantly strengthen institutional readiness, even under resource constraints. Second, Indian experiences highlight how AI-GIS tools can be effectively applied to urban challenges when combined with community engagement and participatory platforms. These patterns suggest that for Armenia, the adoption of AI-GIS must be coupled with governance reforms that improve transparency, citizen trust, and institutional coordination. Such insights justify the relevance of using Yerevan as a case study, since its challenges mirror those of other transitional urban contexts.

2.6. Gaps and Future Research Directions

Despite the growing body of literature on AI and geospatial technologies for sustainable urban management, several critical gaps persist that hinder broader application, especially in emerging and post-Soviet urban contexts. These gaps are evident in technological integration, institutional capacity, socio-political adaptation, and the theoretical framing of smart city development [26,27,29].

First, a major gap lies in the limited integration of AI and GIS into coherent, interoperable platforms. Many studies emphasize the separate usage of these technologies without addressing how they can be systemically combined within a unified decision-support architecture. This fragmentation leads to inefficiencies in data processing, limits scalability, and inhibits policy uptake [29].

Second, there is a pronounced lack of longitudinal and real-time datasets, particularly in cities with underdeveloped digital infrastructures. Without high-resolution temporal data, predictive models remain speculative, and land-use simulations become less accurate in representing rapid urban change. Moreover, in post-Soviet regions, access to public data is often constrained due to institutional inertia and lack of open-data policies.

Third, institutional readiness remains a barrier to implementation. While technological capacity may exist through partnerships or donor projects, the institutional mechanisms for absorbing and utilizing AI-GIS tools are often underdeveloped. Research is needed on governance models that can bridge the gap between technological innovation and administrative practice, particularly in decentralized or hybrid governance systems [17,19].

Fourth, ethical and equity concerns regarding algorithmic decision-making are insufficiently explored. There is limited engagement with how AI-based planning might reinforce existing inequalities or exclude marginalized populations from participatory processes. The future of smart governance must be anchored in transparency, accountability, and social inclusion.

Finally, there is a need for context-sensitive theoretical frameworks that move beyond Western-centric models of smart urbanism. Urban development in post-Soviet, African, or South Asian contexts follows distinct trajectories, shaped by legacy systems, informal practices, and geopolitical dynamics [26,27]. Thus, future research should prioritize localized models, co-created with stakeholders that balance technological sophistication with cultural and institutional fit.

These gaps align directly with the three objectives of this study. The limited integration of AI and GIS highlights technical barriers for accurate land-use modeling (Objective 1). Weak institutional readiness and constrained data governance emphasize the need to evaluate governance capacity and participatory mechanisms (Objective 2). Finally, the lack of scalable and context-sensitive frameworks underscores the importance of developing a transferable methodological model for emerging smart cities (Objective 3). Moreover, issues of equity and participation call for greater engagement with literature on citizen inclusion, digital participation, and socio-technical integration to ensure that AI-GIS tools enhance transparency and inclusivity rather than reinforce existing inequalities.

To address these challenges, scholars recommend advancing modular AI-GIS toolkits, fostering transdisciplinary research collaborations, and investing in municipal capacity-building. Moreover, the development of comparative case studies and open-source urban simulation environments can democratize access to smart planning tools and support adaptive sustainability strategies in diverse cities.

This review highlights that while extensive research exists on smart cities, AI, and GIS individually, far fewer studies integrate these domains into a unified methodological framework. In particular, the connection between AI-driven geospatial analysis and digital participation/governance mechanisms remains underexplored. Building on this gap, the present study contributes by combining CNN-based land-use classification, Random Forest and XGBoost predictive modeling, spatial autocorrelation techniques (Moran’s I, LISA), and NLP/LLM-driven analysis of citizen feedback. This integration provides both technical innovation and governance relevance, offering a reproducible framework for emerging smart cities.

In summary, the literature highlights significant progress in smart city research but also reveals unresolved challenges that are highly relevant for emerging and post-Soviet contexts. This study responds by focusing on three key dimensions: (i) the technical integration of AI-GIS for land-use and risk modeling, (ii) the institutional and governance capacity needed for participatory smart city planning, and (iii) the development of a scalable methodological framework adaptable to diverse urban settings. By synthesizing insights rather than repeating descriptive accounts, the literature review directly supports the objectives and sets the stage for the methodological approach that follows.

3. Materials and Methods

This study employs a mixed-methods research design that integrates artificial intelligence (AI) algorithms and geospatial technologies to assess and enhance sustainable land and environmental management in the city of Yerevan, Armenia. The methodology combines spatial data processing, predictive modeling, and governance analysis within a structured four-stage workflow: (1) data acquisition and preprocessing, (2) AI-GIS model development, (3) simulation and validation, and (4) output synthesis and policy translation.

3.1. Data Sources and Preprocessing

A variety of spatial and non-spatial datasets were collected from official and open-access sources. Satellite imagery was acquired from the European Space Agency’s Sentinel-2 database (https://dataspace.copernicus.eu/explore-data/data-collections/sentinel-data/sentinel-2) (accessed on 19 September 2025) with 10 m spatial resolution and 5-day revisit frequency. Land-use and cadastral maps were obtained from the National Cadaster Committee of Armenia (https://www.e-cadastre.am/) (accessed on 19 September 2025). Climate data (temperature, precipitation, wind) were retrieved from the Armenian Hydrometeorological Service (http://env.am/en/environment/environmental-monitoring) (accessed on 19 September 2025), while demographic data were sourced from the National Statistical Committee https://www.armstat.am/en/ (accessed on 19 September 2025). Additional topographic and infrastructure data were integrated from ASTER GDEM and Yerevan’s municipal urban development archives.

All datasets were standardized to a common spatial reference system (WGS84, UTM Zone 38 N) and formatted into interoperable layers (GeoTIFF, Shapefile, CSV). Data preprocessing included noise reduction, image rectification, and normalization of attribute tables for model training.

The datasets used in this study cover the period 2015–2023, ensuring both historical and recent observations for temporal analysis. Demographic data are available at the district level, while cadastral datasets provide parcel-level resolution. To ensure data consistency, missing values were handled through interpolation for continuous variables and cross-verification with secondary official statistics for categorical data. The analysis was conducted for the entire municipality of Yerevan, although selected districts were examined in greater detail to illustrate spatial heterogeneity.

The Normalized Difference Vegetation Index (NDVI) was calculated from Sentinel-2 imagery using the red band (Band 4, 665 nm) and the near-infrared band (Band 8, 842 nm). Land Surface Temperature (LST) was derived from the Thermal Infrared Sensor (TIRS) of Landsat 8, specifically Band 10, by applying the mono-window algorithm. Atmospheric correction parameters required for the LST computation were obtained from the official USGS metadata files associated with each scene. This ensured consistent and scientifically robust estimation of vegetation health and surface temperature indicators for subsequent analyses.

To ensure consistency across datasets, Sentinel-2 (10 m) and Landsat 8 (30 m) imagery were co-registered to a common spatial reference system (WGS84, UTM Zone 38 N). The thermal bands of Landsat 8 were resampled to 10 m resolution using bilinear interpolation to match Sentinel-2 layers. Temporal alignment was achieved by selecting cloud-free imagery acquired within a ±5-day window and applying atmospheric correction and cloud-masking procedures. This preprocessing guaranteed comparability of LST, NDVI, and land-cover inputs in the subsequent AI-GIS analyses.

The temporal coverage of the dataset spans the period 2018–2023, allowing for multi-year monitoring of land-use and environmental dynamics. To ensure data quality, only scenes with less than 10% cloud cover were selected. Cloud and shadow masking was applied using the QA bands provided by both Sentinel-2 and Landsat 8 products, supplemented by additional cloud-detection algorithms to minimize atmospheric noise. These procedures guaranteed the consistency and reliability of the inputs used for subsequent NDVI, LST, and land-cover analyses.

3.2. AI and GIS Integration

To enhance the transparency of model outcomes, we complemented the accuracy metrics with post hoc interpretability analyses. For the Convolutional Neural Network (CNN), we applied Gradient-weighted Class Activation Mapping (Grad-CAM) and Guided Grad-CAM to visualize class-discriminative regions on Sentinel-2 imagery. These saliency maps consistently emphasized roof textures, impervious surface boundaries, and low-NDVI patches for built-up classes, while green areas were associated with canopy structure and high-NDVI zones. Water bodies were primarily identified through low-NIR and high-blue spectral responses.

In this study, machine learning refers specifically to tree-based predictive models (Random Forest, XGBoost) used for environmental risk forecasting, while deep learning refers to Convolutional Neural Networks (CNNs) such as U-Net and DeepLabv3+ applied to land-use classification. NLP and LLM-based techniques were additionally employed to analyze stakeholder interviews and citizen feedback, linking technical outputs with participatory governance. This clarified scope ensures methodological precision and avoids generic use of AI terminology.

For the tree-based models (Random Forest, XGBoost), we employed SHapley Additive exPlanations (SHAP) to compute both global and class-specific feature attributions. The most influential predictors across land-use categories were Land Surface Temperature (LST) and NDVI, followed by proximity to industrial and traffic corridors, urban density, and elevation. These results were further supported by partial dependence plots, which illustrated the marginal effects of each predictor.

For the Support Vector Machine (SVM), where native interpretability is limited, we applied permutation feature importance and partial dependence analysis to ensure comparability.

Overall, these interpretability procedures enabled us to assess whether the models relied on ecologically and spatially meaningful features. The findings provided additional confidence in the validity of the predictive outputs and informed the subsequent environmental risk analysis.

In this study, machine learning is defined as tree-based predictive models (Random Forest, XGBoost) used for environmental risk forecasting and feature attribution, while deep learning refers specifically to Convolutional Neural Networks (CNNs) such as U-Net and DeepLabv3+ applied to land-use and land-cover classification. Additionally, NLP and LLM-based tools were employed to analyze textual data from stakeholder interviews and public participation platforms, enabling the integration of citizen-centric insights into the technical framework. This clarified scope ensures methodological transparency and prevents ambiguity in the application of AI concepts.

CNN Architecture and Training. The land-use classifier employs a CNN with five convolutional blocks (3 × 3 kernels, ReLU activations), each followed by batch normalization and max-pooling. A global average pooling layer feeds a dense layer (128 units, ReLU) and a softmax output over the target classes. Training used the Adam optimizer (learning rate 1 × 10⁻³) with categorical cross-entropy loss, for up to 50 epochs and batch size 64, with early stopping (patience = 7) on validation loss. Regularization included dropout (0.3) and L2 weight decay (1 × 10⁻⁴). Data augmentation comprised random rotations (±10°), horizontal/vertical flips, random crops, and brightness/contrast jitter. Class imbalance was addressed via class-weighted loss.

It should be noted that the use of a conventional 70/30 train-test split may introduce potential spatial leakage in remote-sensing applications. While this approach is widely adopted, alternative strategies such as spatial cross-validation could further minimize this risk. We therefore acknowledge this as a methodological limitation of the current study.

3.3. Simulation and Risk Zoning

Using the best-performing AI model (CNN), a scenario-based simulation was conducted to project land-use patterns under a “business-as-usual” trajectory for 2030. Environmental risk zones—such as flood-prone areas, urban heat islands, and air pollution corridors—were delineated using a composite index of NDVI, land surface temperature (LST), and traffic density. These zones were then overlaid on current land-use maps to identify critical areas of intervention.

The weighting scheme of the Environmental Risk Index (ERI) was informed by both prior literature and expert consultation. Land Surface Temperature (LST) was assigned the highest weight (40%), reflecting its strong documented correlation with urban heat island intensity and ecological stress [42,45]. NDVI was weighted at 30% as a direct proxy of vegetation health and ecosystem resilience, consistent with common practice in urban ecological assessments [62]. Proximity to industrial and traffic corridors was also weighted at 30%, acknowledging its contribution to air pollution and localized heat accumulation [41]. Expert input from local environmental planners confirmed that this distribution best reflects Yerevan’s urban conditions. This weighting structure ensured that the ERI captured both biophysical and anthropogenic drivers of environmental risk. The simulation was applied across the full extent of Yerevan municipality, while selected districts with the highest composite risk scores were mapped in greater detail to illustrate local vulnerabilities. This approach ensured that the Environmental Risk Index was both citywide in scope and contextually sensitive.

Comparative model evaluation showed that CNN outperformed Random Forest, SVM, and XGBoost, achieving the highest classification accuracy (92%) and F1-score (0.89), thereby justifying its use for scenario simulation. For the Environmental Risk Index, all input layers (NDVI, LST, traffic density) were normalized to a common scale prior to weighting. The weighting scheme (40% LST, 30% NDVI, 30% traffic density) combined evidence from prior literature with empirical validation through consultation with local environmental experts. Threshold values for high-risk classification were thus derived using both established ecological indicators and context-specific expertise, ensuring methodological robustness.

3.4. Governance and Stakeholder Assessment

A governance readiness matrix was developed to assess Yerevan’s institutional capacity to adopt smart urban technologies. The matrix evaluated six dimensions: institutional capacity, data infrastructure, policy alignment, public engagement, technical expertise, and regulatory support. Input was gathered through expert interviews and a review of municipal strategic documents.

The governance readiness matrix was conceptually informed by established frameworks, including the UN E-Government Development Index, OECD digital government principles, and World Bank governance diagnostics, while adapted to Armenia’s institutional context. The six dimensions selected—data governance, institutional capacity, transparency, participation, regulatory alignment, and technological infrastructure—thus combine international best practice with local relevance. To validate these dimensions, 12 experts were interviewed, representing government agencies, academia, and civil society. Semi-structured interviews were conducted between March and May 2023, with participants balanced in terms of gender (7 male, 5 female) and professional background. This design ensured both methodological rigor and contextual grounding in assessing governance readiness.

The scoring procedure of the Governance Readiness Matrix employed a 1–5 Likert scale across all governance dimensions. Two independent raters scored each interview transcript, and discrepancies were reconciled through discussion. Inter-rater reliability was measured using Cohen’s kappa (κ = 0.82), reflecting strong agreement and calibration across evaluators.

4. Results

4.1. Model Performance Evaluation

To evaluate the predictive accuracy and operational efficiency of various artificial intelligence algorithms applied in land-use classification and environmental risk assessment, four models were trained and tested: Random Forest (RF), Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Gradient Boosting (XGBoost). The performance of each model was assessed using standard classification metrics, including accuracy, precision, recall, and F1-score. Additionally, interpretability and computational time were used as auxiliary criteria in model selection.

As shown in

Table 1. AI model performance metrics. Source: Generated by authors based on model output, 2025 *.

Model	Accuracy, %	Precision	Recall	F1-Score	Computation Time	Interpretability
CNN	92.4	0.91	0.93	0.92	Moderate	Moderate
XGBoost	88.7	0.87	0.86	0.86	Fast	Moderate
Random Forest	84.1	0.82	0.83	0.82	Fast	High
SVM	82.5	0.8	0.78	0.79	Slow	Low
CNN	92.4	0.91	0.93	0.92	Moderate	Moderate

* Note: Interpretability is labeled as “Moderate” to indicate the availability of reliable post-hoc explanations (e.g., Grad-CAM for CNN; SHAP for tree-based models), although the models are not inherently transparent in structure.

, CNN outperformed other models across all performance metrics, achieving an overall classification accuracy of 92.4%, compared to 88.7% for XGBoost, 84.1% for RF, and 82.5% for SVM. The CNN model also demonstrated the highest precision and recall scores, particularly in detecting mixed-use and green zones, which are typically challenging due to spectral similarity with adjacent urban classes.

In this table, ‘Fast’ indicates average training and inference times under 2 min per run, while ‘Slow’ indicates runs typically exceeding 5 min. Interpretability was labeled as ‘Moderate’ when reliable post-hoc tools (e.g., Grad-CAM, SHAP) were available, and as ‘High’ when the model structure was inherently transparent (e.g., Random Forest).

The final selection of the CNN model for subsequent simulation tasks was justified not only by its superior performance but also its ability to learn spatial features across layers of remote sensing data, which are essential for mapping urban complexity.

Hyperparameter tuning was performed for all models to ensure fair comparison. For Random Forest and XGBoost, the number of trees, depth, and learning rates were optimized via grid search; for SVM, kernel type and regularization parameters were adjusted; for CNN, learning rate, batch size, and epochs were tuned through iterative testing. In addition to accuracy and F1-scores, interpretability and computational time were evaluated: ensemble methods (RF, XGBoost) offered higher interpretability but required longer training, while CNN achieved superior accuracy with moderate training time. This balance of predictive performance and computational feasibility provided the basis for selecting CNN as the best-performing model.

In addition to overall accuracy, class-wise performance metrics were calculated for all eight land-use categories (

Table 2. Class-wise performance metrics. Source: Generated by authors based on model output (CNN classification results, 2025).

Class	Precision	Recall	F1-Score
Residential	0.91	0.89	0.90
Industrial	0.87	0.85	0.86
Commercial	0.84	0.83	0.83
Recreational	0.81	0.79	0.80
Green areas	0.92	0.90	0.91
Water bodies	0.95	0.93	0.94
Transport	0.86	0.84	0.85
Undeveloped	0.83	0.82	0.82

Note: The class-wise performance metrics demonstrate high reliability for residential, industrial, and green areas, while mixed-use categories (commercial and undeveloped) show moderate values due to overlaps with adjacent urban classes. These results provide a more transparent and detailed evaluation of classification outcomes beyond overall accuracy.

).

Precision, recall, and F1-scores indicated high reliability for residential, industrial, and green areas, while mixed-use zones showed moderate accuracy due to overlaps with adjacent classes. To further clarify classification performance, a confusion matrix (Figure 1) has been added, highlighting the distribution of misclassifications across categories. These additions provide a more detailed and transparent evaluation of the classification outcomes.

In addition to accuracy metrics, we examined model explanations. For the CNN, Grad-CAM saliency maps consistently emphasized edges and textures of impervious surfaces and low-NDVI patches for built-up classes, while high-NDVI canopy structure dominated attributions for green areas; water predictions followed low-NIR/high-blue responses. For RF/XGBoost, SHAP values ranked LST and NDVI as the most influential features, followed by proximity to industrial/traffic corridors, urban density, and elevation, mirroring the drivers observed in the environmental-risk analysis. Mixed-use boundaries remained the main source of ambiguity across models. We therefore label interpretability as ‘Moderate’ (Table 1, footnote), indicating robust post hoc explanations despite non-transparent internal representations.

4.2. Land-Use Classification and Simulation (2023–2030)

The land-use classification for the city of Yerevan was carried out using the CNN model, trained on Sentinel-2 imagery and validated against official cadastral datasets. The classification identified eight land-use categories: residential, industrial, commercial, recreational, green areas, water bodies, transportation infrastructure, and undeveloped land. The spatial resolution of 10 m enabled fine-grained detection of intra-urban heterogeneity, including fragmented green spaces and informal settlements.

Figure 2 presents the classified land-use map for the year 2023. Urban areas are primarily concentrated in the central and northern districts (Kentron, Arabkir, and Ajapnyak), while peripheral zones such as Erebuni and Shengavit exhibit mixed-use and transitional land types.

To simulate future land-use dynamics, a scenario-based projection was performed for the year 2030 under a “business-as-usual” (BAU) assumption. The simulation incorporated population growth rates (projected at 0.4% per annum), urban infrastructure expansion plans (Yerevan Master Plan 2021–2030), and spatial development constraints such as slope and protected green zones. Land transformation probabilities were computed using the trained CNN model, and spatial cellular automata were applied to simulate urban expansion patterns.

The “business-as-usual” (BAU) scenario was developed using official planning and demographic sources. In particular, the Yerevan Master Plan 2021–2030 provided projections of infrastructure expansion and zoning priorities, while population growth forecasts from the National Statistical Committee (0.4% annual increase) informed the expected demand for residential areas. These drivers were translated into transition probabilities by linking demographic growth to residential land expansion, planned infrastructure projects to industrial and commercial development, and zoning regulations to restrictions on green area conversion. This ensured that the simulation reflected both demographic pressures and institutional planning frameworks.

The land-use simulation was implemented through a cellular automata framework. A 5 × 5 Moore neighborhood structure was applied to capture spatial interactions among adjacent cells. Transition probabilities were derived from observed land-cover changes during 2018–2021, reflecting empirically observed conversion trends between built-up, green, and mixed-use classes. Model calibration followed a split-sample validation design, with 2018–2020 data used for training and 2021–2022 data used for testing. Constraints were imposed to restrict conversions in protected natural areas and water bodies, thereby ensuring ecological plausibility. These specifications enhanced the transparency, robustness, and reproducibility of the simulation outcomes projected for 2023–2030.

Model validation for the 2023 classification was quantified through a confusion matrix, with overall accuracy of 91% and per-class accuracies ranging from 86% (grassland) to 94% (built-up areas). These results substantiate the reliability of the classification. The assumptions of the business-as-usual scenario were based on official demographic projections [71], municipal infrastructure plans, and spatial growth constraints from cadastral datasets, ensuring that the scenario reflects realistic trends. Transformation probabilities in the cellular automata model were derived from historical land-use transitions (2015–2020) and parameterized following established approaches [28], with calibration against observed 2021–2022 changes. The projected figures—12.8% increase in built-up areas and 6.5% decline in green areas by 2030—are consistent with recent historical trajectories in Yerevan and comparable patterns observed in Tbilisi and Baku, confirming their plausibility.

The resulting 2030 projection revealed a 12.8% increase in built-up areas, primarily through the conversion of undeveloped and green spaces in the southern and eastern peripheries. Industrial land expanded modestly (by 3.1%), while green areas faced a projected decline of 6.5%, emphasizing the urgency of implementing protective zoning regulations. These values represent central estimates under the BAU scenario and should be interpreted as indicative trends rather than precise predictions, as a full uncertainty quantification (e.g., Monte Carlo simulations) was beyond the scope of this study.

In addition, spatial fragmentation of green areas became more pronounced, leading to potential degradation of urban ecological corridors. This result underscores the need for integrated urban planning approaches that account for ecological connectivity, not just surface area preservation.

Figure 3 visualizes the projected land-use composition for 2030, demonstrating the spatial extent of densification and ecological stress zones.

The base map of Yerevan’s administrative districts provides geographic context, while overlay layers indicate simulated built-up expansion (red) and identified environmental risk zones (yellow). This visualization integrates the CNN-based classification results with scenario projections, ensuring that readers can clearly distinguish background geography from modeled outputs.

Overall, the results emphasize the dual pressure of population growth and weak land-use enforcement in shaping Yerevan’s urban form. The AI-GIS simulation provides actionable spatial intelligence for policymakers to identify areas of potential conflict between development and ecological sustainability.

4.3. Environmental Risk Mapping and Hotspot Identification

Environmental risk assessment was conducted using a composite spatial index that integrates vegetation cover, land surface temperature (LST), air quality proxies, and topographic constraints. The aim was to identify urban hotspots prone to ecological stress, enabling proactive land and environmental management.

Three core indicators were used:

• Normalized Difference Vegetation Index (NDVI)—derived from Sentinel-2 imagery to assess vegetation health.
• Land Surface Temperature (LST)—retrieved from thermal bands of Landsat 8 using radiative transfer algorithms.
• Proximity to Traffic and Industry—used as a proxy for air pollution and urban heat islands.

Each layer was standardized on a 0–1 scale and combined using a weighted sum model. Weights were assigned as follows: NDVI (30%), LST (40%), Proximity Index (30%). The resulting Environmental Risk Index (ERI) highlighted areas with high ecological degradation risk.

High-risk zones were predominantly found in the Malatia-Sebastia, Erebuni, and Shengavit districts, which are characterized by dense housing, low green coverage, and proximity to major industrial corridors. These areas also exhibit elevated land surface temperatures and lower NDVI values.

Conversely, districts such as Kanaker-Zeytun and Nork-Marash, with higher elevation and vegetation density, presented moderate to low-risk values.

The spatial distribution of risk indicates a concentric pattern of environmental stress, with increasing intensity from the urban core toward expanding industrial peripheries. This confirms that unregulated urban sprawl, when combined with weak enforcement of environmental zoning, significantly intensifies ecological vulnerability.

A further overlay of risk zones with the projected 2030 land-use expansion (Figure 2) revealed a disturbing overlap between future development zones and current high-risk ecological areas. This calls for immediate regulatory interventions and the integration of environmental constraints into land-use planning frameworks.

To support this analysis,

Table 3. Key factors influencing environmental risk in Yerevan. Source: Compiled by authors based on spatial analysis results, 2025.

Risk Driver	Correlation with ERI	Influence Level
Land Surface Temperature (LST)	+0.81	Very High
NDVI (Vegetation Index)	−0.78	Very High
Proximity to Industrial Sites	+0.66	High
Urban Density	+0.59	Moderate
Elevation	−0.42	Moderate

presents key risk drivers ranked by their spatial correlation with the Environmental Risk Index (ERI).

These findings demonstrate the potential of AI-GIS tools in pinpointing vulnerable zones, enabling data-driven environmental planning, especially in cities undergoing rapid transformation like Yerevan.

To test the robustness of the Environmental Risk Index (ERI), a sensitivity analysis was performed by varying the weights of NDVI, LST, and Proximity by ±10%. The resulting hotspot maps showed more than 85% spatial overlap with the baseline scenario, indicating that the identification of high-risk zones is not strongly dependent on the exact weighting scheme. This confirms that the ERI provides stable and reliable results under reasonable variations in parameterization.

The weighting scheme of 40% (LST), 30% (NDVI), and 30% (traffic/industrial proximity) was based on prior ecological assessments [62] and validated through sensitivity analysis. Similar composite indices have been widely used in urban ecological studies to capture the combined effects of vegetation cover, land surface temperature, and anthropogenic pressures on environmental vulnerability [42,52,62]. Moreover, the correlation between LST and urban heat island intensity has been consistently documented in cities undergoing rapid densification [42,52]. Studies also emphasize the importance of NDVI as a proxy for ecological resilience and its inverse relationship with urban ecological risk [62]. LST was derived from Landsat 8 thermal bands for the period 2015–2023, with surface temperature estimates validated against meteorological station records in Yerevan. The observed overlap between projected development zones and existing high-risk areas highlights the need for integrating environmental risk layers into municipal planning processes, particularly in relation to zoning regulations, infrastructure expansion, and citizen safety.

4.4. Governance Readiness

Effective implementation of AI- and GIS-based tools for sustainable land and environmental management requires not only technological capacity but also robust governance mechanisms. To evaluate Yerevan’s institutional preparedness for such transformation, a six-dimension Governance Readiness Matrix (GRM) was developed and applied. The GRM assesses key enablers across institutional, technical, policy, and participatory domains.

The six dimensions include the following:

• Institutional Capacity—availability of specialized departments and human resources in digital planning.
• Data Infrastructure—existence and quality of open spatial databases, interoperability standards, and update frequency.
• Policy and Legal Alignment—consistency of urban planning regulations with sustainability and digital governance frameworks.
• Public Participation—opportunities for citizen engagement in planning, reporting, and feedback mechanisms.
• Technical Expertise—local proficiency in AI, geospatial analysis, and data interpretation.
• Regulatory Support—legal mandates for digital transformation in land management.

Using expert scoring and document analysis, each dimension was rated on a 5-point scale, with 1 = Very Weak and 5 = Very Strong. The results are summarized in

Table 4. Governance readiness matrix for AI-GIS-based land management in Yerevan. Source: Developed by authors based on document review and expert interviews, 2025.

Governance Dimension	Score (1–5)	Readiness Level
Institutional Capacity	3	Moderate
Data Infrastructure	2	Weak
Policy and Legal Alignment	3	Moderate
Public Participation	2	Weak
Technical Expertise	4	Strong
Regulatory Support	3	Moderate

.

These findings suggest that Yerevan demonstrates moderate readiness in institutional and policy terms but faces significant limitations in data infrastructure and participatory governance. While the presence of GIS professionals in urban departments is encouraging, the absence of a unified urban data platform and low public engagement present key barriers.

Notably, technical expertise scored highest due to Armenia’s strong STEM education base and growing tech ecosystem, which positions Yerevan as a potential hub for smart city experimentation. However, these capabilities must be better institutionalized and linked with legal and planning frameworks.

The GRM results provide a clear direction for policy intervention: strengthening data governance systems, formalizing public participation processes, and enacting legal mandates to mainstream AI-GIS tools in municipal planning.

The scoring of the six dimensions was informed by 12 semi-structured expert interviews conducted with representatives of government agencies, academia, and civil society. Scores were first assigned individually and then reconciled through a moderated workshop discussion to ensure consensus and reduce individual bias. This process enhanced the robustness of the matrix. The framework draws on established references such as the UN E-Government Development Index and OECD digital government principles, while being adapted to post-Soviet contexts where institutional fragmentation and data governance challenges remain significant. In addition, Yerevan’s strong STEM capacity is documented in recent higher-education and innovation reports [72], which highlight the city’s growing base of skilled professionals as a potential enabler for digital transformation.

4.5. Strategic Impact Assessment of the AI-GIS Framework

The integration of artificial intelligence and geospatial technologies into urban land and environmental management in Yerevan has revealed significant strategic implications. This section synthesizes the framework’s impact across five critical dimensions: spatial accuracy, policy relevance, decision support, sustainability alignment, and institutional integration.

Enhanced Spatial Accuracy. Improvements in this area (92.4%) enabled finer detection of fragmented green spaces, informal settlements, and transitional zones, providing urban planners with reliable micro-level spatial intelligence.

Policy Relevance and Forecasting Capability. A simulation module projected land-use changes under a business-as-usual (BAU) scenario up to the year 2030. This predictive capability enables policymakers to formulate proactive policies, anticipate densification patterns, identify emerging urban sprawl, and inform zoning regulations, environmental preservation strategies, and infrastructure prioritization.

Risk-Sensitive Planning. Environmental risk mapping, leveraging land surface temperature (LST), normalized difference vegetation index (NDVI), and proximity to pollution sources, identified high-risk zones, which empowers municipalities to target ecological interventions, such as urban greening initiatives or regulatory overlays, on an evidence-based rather than an assumption-driven basis.

Alignment with the SDGs and Smart City Agenda. The framework directly supports SDG11 (Sustainable Cities and Communities) by addressing the following key targets:

• 11.3: Inclusive and sustainable urbanization.
• 11.6: Reducing adverse environmental impact of cities.
• 11.7: Provision of green and public spaces.

Moreover, the digital nature of the AI-GIS system aligns with Armenia’s broader Digital Transformation Strategy, contributing to e-governance and data-driven public service delivery.

Institutional Learning and Governance Capacity. The application of the Governance Readiness Matrix provided actionable insights into institutional gaps. Specifically, the low scores in data infrastructure and public participation suggest the need for strategic investment in (1) urban data portals and open-access frameworks; (2) civic-tech platforms for participatory planning; and (3) legal reforms embedding AI tools in regulatory mandates. Figure 4 below summarizes the framework’s strategic impact across five dimensions.

Figure 4 has been revised to consistently present four dimensions; each explained in detail. Beyond reporting spatial accuracy, we also assessed sources of uncertainty and potential misclassification errors, noting that vegetation and built-up classes showed the highest confusion in certain districts. These limitations inform a more cautious interpretation of results. The alignment of the framework with SDGs has been clarified, linking outputs specifically to SDG 11.3 (inclusive urbanization) and SDG 13.1 (climate resilience), without overstating broader contributions. Finally, institutional recommendations are grounded in the Armenian context, referencing ongoing reforms in open-data portals, e-governance platforms, and urban planning regulations, thereby ensuring policy relevance and feasibility.

In conclusion, the AI-GIS framework not only enhances urban spatial planning but also enables transformative shifts in governance, sustainability, and citizen engagement. For cities like Yerevan, this approach marks a strategic pivot from reactive regulation to predictive, participatory, and environmentally intelligent planning.

4.6. Implementation Roadmap

In order to operationalize the AI-GIS framework for sustainable land and environmental management in Yerevan, a phased implementation roadmap is proposed. This roadmap is designed to transition the framework from a proof-of-concept to a fully institutionalized system, supporting long-term urban resilience and smart governance.

The roadmap includes six sequential phases, each linked with expected outputs, responsible stakeholders, and capacity requirements (

Table 5. AI-GIS implementation roadmap. Source: Generated by authors.

Phase	Timeline	Main Objective	Key Actions	Stakeholders	Output
Phase 1 Onboarding	Q1–Q2 2025	Build consensus among key governmental and municipal agencies	Establish an inter-agency AI-GIS task force	Yerevan Municipality	Official adoption plan and institutional commitment
			Align goals with national smart city and digital transformation strategies	Ministry of Environment Ministry of High-Tech Industry
Phase 2 Data Consolidation and Infrastructure Setup	Q3–Q4 2025	Ensure the availability and interoperability of core spatial and administrative datasets	Integrate cadastral, topographic, and environmental data	Cadaster Committee	Centralized, regularly updated geo-database
			Develop urban data portal with access protocols and update cycles	ArmStat GIS contractors
Phase 3 AI Model Calibration and Testing	Q1–Q2 2026	Localize and optimize machine learning models for Yerevan’s spatial profile	Fine-tune CNN models using local imagery	Universities AI research centers	Locally adapted, high-performance AI models
			Validate risk indices and land-use classification accuracy
				International technical partners
Phase 4 Pilot Deployment in High-Risk Districts	Q3–Q4 2026	Test the system in priority areas facing ecological or planning stress	Deploy dashboard and alert systems for selected districts Engage local authorities in feedback and refinement	Yerevan District Offices Environmental NGOs Civil tech groups	Operational early warning and planning tool
Phase 5 Legal and Policy Integration	2027	Embed AI-GIS tools into legal urban planning procedures	Amend planning codes to incorporate spatial intelligence tools Define responsibilities, liabilities, and public disclosure rules	National Assembly, Ministry of Justice, legal experts	AI-enabled planning legislation and compliance protocols
Phase 6 Full Rollout and Capacity Building	2028 onwards	Institutionalize the system city-wide and train personnel	Scale system to all districts Provide continuous training for public planners and data managers Launch civic education campaigns to involve residents	Training institutes, media, schools, professional associations	Fully operational, inclusive, and transparent AI-GIS governance system

).

While the phased roadmap provides a structured pathway, its feasibility in Armenia depends on addressing several enabling conditions and risks. Coordination among government agencies, municipalities, and civil society actors requires clearly defined accountability mechanisms and conflict resolution procedures. Potential barriers—such as restrictions on inter-agency data sharing, institutional inertia, and political resistance—may slow implementation and must be explicitly planned for. The assumption of seamless legal integration has therefore been revised to acknowledge these challenges. Finally, the phase on capacity building and citizen engagement has been expanded by drawing on international best practices in participatory governance (e.g., OECD, UN-Habitat), ensuring that the roadmap incorporates not only technical but also institutional and social dimensions.

5. Discussion

While the results demonstrate the potential of AI-GIS integration, several methodological limitations must be acknowledged. Data availability and quality remain uneven, particularly for cadastral and demographic inputs, which constrains model precision. Uncertainty also arises from classification errors, as noted in the confusion matrix, and from assumptions in the business-as-usual scenario. These limitations do not undermine the core findings but require cautious interpretation. The framework’s alignment with SDGs and the New Urban Agenda has been refined by linking outputs directly to measurable targets—SDG 11.3 on inclusive urbanization and SDG 13.1 on climate resilience—thereby grounding claims in evidence. Beyond technical integration, this study highlights concrete pathways for AI-GIS tools to support inclusive governance. Participatory mapping platforms, NLP-based analysis of citizen feedback, and open-access urban dashboards can enable residents to contribute to land-use planning and environmental monitoring. International examples from Helsinki and Barcelona demonstrate how civic-tech tools improve transparency and accountability, offering lessons for Armenia. However, feasibility is constrained by political and institutional barriers: inter-agency data silos, limited legal frameworks for data sharing, and potential resistance from entrenched interests. The implementation roadmap must therefore be viewed not only as a stepwise plan but also as a negotiation process that requires accountability mechanisms, conflict resolution tools, and incremental trust-building with citizens. Comparative reflection indicates that while Yerevan shares governance barriers with other post-Soviet cities (fragmented institutions, weak data governance), its growing STEM base and civic-tech potential differentiate it from many Global South contexts. This highlights both opportunities for transferability and the need for context-sensitive adaptation.

The Implementation Roadmap, originally presented in the Results section, is here reframed as part of the discussion to emphasize its interpretive nature. The phased structure continues to provide clarity, but its feasibility depends on addressing institutional barriers, political constraints, and coordination mechanisms. Integrating the roadmap in this section also highlights its function as a forward-looking framework rather than a descriptive output, linking the technical findings to governance reforms, capacity-building, and citizen participation. This repositioning avoids repetition while expanding the discussion on practical implementation and future directions.

The findings of this study underscore the transformative potential of integrating artificial intelligence (AI) and geospatial technologies (GIS) for urban sustainability planning in emerging smart cities like Yerevan. The high-performance metrics achieved by CNN-based land-use classification (92.4% accuracy) reflect the growing maturity of machine learning methods in urban data environments and validate the increasing reliance on deep learning for complex spatial categorization tasks. These results echo the broader literature, which recognizes convolutional neural networks as effective tools for parsing heterogeneous urban landscapes [28,35].

One of the critical insights from this research is the spatial mismatch between current development patterns and areas of high ecological vulnerability. The environmental risk zones identified through composite indices strongly correlate with projected urban expansion under a “business-as-usual” scenario, raising immediate concerns about unregulated sprawl, ecological fragmentation, and heat island intensification. This finding is consistent with existing empirical studies from the Global South and post-Soviet regions, where land-use regulations remain under-enforced and disconnected from environmental indicators [6,28,35,40].

The Governance Readiness Matrix (GRM) further exposes systemic constraints that limit the operationalization of smart urban tools. In particular, Yerevan’s weak performance in data infrastructure and public participation points to institutional inertia and fragmented stakeholder engagement, challenges that are widely noted in transitional governance systems. Nonetheless, the city’s strong technical expertise—bolstered by its growing tech sector and academic talent—offers a foundation upon which to build capacity for long-term digital transformation.

Importantly, the AI-GIS framework proposed in this study not only offers technical improvements in land-use monitoring and risk assessment but also serves as a strategic tool for policy learning and adaptive governance. When embedded in municipal planning workflows, this integrated approach facilitates anticipatory decision-making and enhances alignment with global sustainability agendas, such as SDG11 and the New Urban Agenda. Moreover, the participatory potential of spatial dashboards and open-data portals, if realized, can democratize urban knowledge and empower local communities—goals that remain underdeveloped in Yerevan’s current planning ecosystem.

Comparative analysis with other post-Soviet and emerging cities [25,26,27] suggests that successful AI-GIS adoption is contingent not merely on technological availability but on multi-level governance, inter-agency coordination, and legal mandates. In Yerevan’s case, the proposed implementation roadmap offers a practical blueprint for bridging this gap. However, sustained political commitment, civic engagement, and institutional reform are essential for transitioning from pilot projects to systemic change.

6. Conclusions

This study introduced a novel AI-GIS framework designed to enhance sustainable land and environmental management in emerging smart cities, using Yerevan as a case study. By integrating artificial intelligence with geospatial technologies, our framework provides a strategic paradigm shift toward data-informed and resilient urban governance.

This study advances the literature by demonstrating how AI-GIS integration can be operationalized not only as a technical innovation but also as a governance-oriented framework in transitional contexts. The unique contribution lies in linking deep learning, predictive modeling, and spatial statistics with participatory governance mechanisms, thereby bridging methodological innovation and institutional application. Nevertheless, the research has clear limitations: data constraints in cadastral and demographic records, potential misclassification errors, and biases arising from governance structures. These limitations underscore the need for cautious interpretation and for iterative refinement of the framework. The outlook for future research should focus on strengthening longitudinal datasets, refining risk modeling with participatory inputs, and testing the framework’s scalability in other post-Soviet and emerging urban contexts. By doing so, this study contributes both conceptually and practically to ongoing debates on smart urbanism, while offering lessons that are both locally grounded and globally relevant.

Our findings demonstrate the framework’s effectiveness through several key outcomes:

• The Convolutional Neural Network (CNN) model achieved over 92% accuracy in land-use classification, providing unparalleled spatial resolution for monitoring green infrastructure and urban sprawl;
• Simulated land-use changes projected through 2030 offered crucial insights into anticipated urban growth and identified specific ecological stress zones;
• Environmental risk mapping successfully pinpointed vulnerability hotspots driven by factors like temperature extremes, vegetation loss, and industrial proximity;
• The Governance Readiness Matrix (GRM) revealed moderate preparedness in Yerevan, highlighting weaknesses in data infrastructure and participatory mechanisms while confirming the framework’s strong alignment with Sustainable Development Goal 11 and Armenia’s Digital Transformation Strategy.

Based on these results, we propose the following recommendations and avenues for future research:

• We urge the integration of AI-GIS frameworks into urban planning regulations and the establishment of an Urban Data Authority to standardize data collection. Enhancing civic-tech platforms and open-data portals is also crucial for boosting participatory governance.
• We recommend expanding AI model training with local data, deploying early warning systems in high-risk areas for proactive resilience, and developing capacity-building programs to train urban planners and municipal staff in geospatial intelligence and AI ethics.
• The framework should be applied to other Armenian cities, such as Gyumri and Vanadzor, for comparative regional assessments.

Future research could explore the integration of digital twins and metaverse environments to create more immersive and participatory land-use simulations. Moreover, further investigation is needed into the socio-ethical implications of AI-driven decision-making, particularly concerning data privacy and algorithmic transparency.

In sum, this study provides a replicable model for other emerging cities aiming to balance rapid urbanization with the imperatives of sustainable development.