Predicting Environmental Risks and Managing Urban Ecosystem Resilience
through Artificial Intelligence
Abstract. City territories across the globe are undergoing physical transformations at a pace that existing environmental monitoring tools were not designed to handle. This paper reports on an attempt to close that gap by pairing machine learning classifiers with geographic information systems, applying the combination to a five-year record of multispectral satellite imagery from Sentinel-2 and Landsat-8 (2020–2025). A two-stage hybrid architecture — convolutional feature extraction followed by Random Forest classification — produced an overall accuracy of 93.8% and delivered monitoring coverage 72–93% wider than ground-based surveys can achieve in comparable time. Spectral analysis of the imagery showed mean NDVI dropping from 0.52 to 0.47 over the study period, alongside a 12.4% loss of green-space area; both reductions were geographically concentrated in districts experiencing the heaviest construction activity, rather than uniformly distributed. On this evidence, integrating AI with GIS yields genuine gains in diagnostic speed and spatial detail, though model portability and data-quality issues remain practical constraints on broader adoption. Drawing on the analytical findings, the paper proposes a conceptual architecture for a unified urban-ecosystem management platform that would connect satellite-derived risk indicators to municipal decision-making workflows.
Keywords: management and administration, urban development, forecasting, environmental risks, city ecosystem, AI, urban ecosystems, spatial modeling, public space, sustainable urban development
1 Introduction
City governments across the world are contending with a structural mismatch: the speed at which urban land is converted, built over, and repurposed has outrun the capacity of established monitoring routines to keep track of ecological consequences. UN demographic projections put city-dwellers at more than two-thirds of the global population by 2050 [1], a shift that will further intensify land conversion, expand impermeable surfaces, and tighten pressure on whatever green infrastructure remains within urban boundaries.
Urban ecological systems do not absorb these pressures in isolation. Roads, drainage channels, vegetated patches, and concentrations of human activity interact through feedback loops whose outputs are not always obvious: paving over a neighbourhood park, for instance, increases local runoff, warms the immediate microclimate, and breaks habitat linkages — all at once [2, 3]. Such compound effects tend to emerge gradually across spatial scales and seasonal rhythms that neither annual field surveys nor periodic land-register updates are well-positioned to detect.
Satellite observation has long served as a partial corrective, providing coverage that field teams cannot match. Sentinel-2 and Landsat-8 now revisit any given location frequently enough to distinguish genuine land-cover change from seasonal noise, and at spatial resolutions adequate for city-level work. The gap between data acquisition and usable information, however, has historically been wide: converting a stack of raw spectral bands into a classified map demanded considerable operator time, introduced subjective choices at multiple processing steps, and scaled poorly when applied to multi-year archives.
Machine learning has narrowed that gap substantially. Classifiers trained on labelled scenes can now handle the pattern-recognition workload at scale, matching or exceeding expert accuracy across a range of land-cover tasks while processing volumes of imagery that would be impractical to assess manually. The consequent expansion in research activity is well documented: the literature on AI applications in urban land monitoring has grown sharply over the past decade, spanning topics from impervious-surface detection to heat-island characterisation.
That body of work also exposes a recurring structural problem. Studies focused on classifier benchmarks tend to stop at accuracy tables and say relatively little about how mapped outputs translate into planning decisions. Policy discussions about smart-city governance and digital public administration [4, 5], on the other hand, often treat the technical layer as a solved problem. While hybrid CNN-Random Forest architectures have been applied to land-use classification, their integration with multi-temporal satellite imagery for operational municipal monitoring — including institutional readiness assessment — remains underexplored, rather than examining where the real methodological choices lie. The upshot is that practitioners in municipal environmental management lack the kind of operational guidance they would need to embed AI-based spatial tools reliably into day-to-day workflows [6]
This paper addresses that gap directly. We developed a hybrid classification framework with operational usability in mind, tested it on a documented urban case, and traced the implications for both monitoring practice and the institutional conditions required for wider uptake. The work is organised around four objectives: (1) to benchmark selected machine learning algorithms on a spatial land-use classification task; (2) to document recent vegetation and land-cover dynamics in the study area; (3) to sketch a conceptual architecture for an integrated decision-support platform; and (4) to identify the institutional and methodological prerequisites for systematic AI adoption in urban governance.
2 Literature Review and Theoretical Foundations
2.1 AI in Urban Development Management
The application of machine learning to geospatial data has expanded rapidly over the past decade, driven by the simultaneous availability of open satellite archives, affordable computational infrastructure, and increasingly mature software libraries. Combining convolutional architectures, ensemble classifiers, and kernel methods with GIS workflows has opened up tasks — fine-grained land-cover mapping, scenario-based growth modelling, multi-hazard risk assessment — that were previously too data-intensive or computationally costly for routine use [7].
Convolutional neural networks have proven particularly well-suited to the texture and spectral complexity of urban imagery. ResNet and U-Net variants have achieved classification accuracies in the 88–92% range across diverse land-use categories in built-up settings [8], with transfer learning reducing the volume of locally labelled training data needed to adapt pre-trained weights to new geographic contexts — a practical advantage for the many municipalities that cannot sustain large in-house data-annotation programmes.
Ensemble approaches, and Random Forest in particular, remain competitive despite their relative simplicity. Their ability to rank predictor importance makes their outputs more interpretable than deep networks, which matters in administrative contexts where decisions must be explained and justified to non-specialist audiences [9]. Early warning applications have shown that integrating satellite observations with IoT sensor streams and meteorological records can reduce the lag between an incipient environmental threat and an institutional response by 65–82% [10], a figure large enough to change risk outcomes in time-sensitive situations such as flash flooding or heat-wave emergencies.
2.2 Spatial Modelling of City Ecosystem Changes
A well-established finding in the spatial-modelling literature is that hybrid classifiers consistently outperform single-algorithm approaches on complex urban scenes. Pairing a CNN feature extractor with a Random Forest classifier has yielded accuracies of 93.33% [8] on land-use datasets where either component alone performed several percentage points lower. Our approach extends this architecture by incorporating auxiliary geospatial variables (terrain derivatives, distance-to-road indices, population density) alongside CNN-extracted features, and by optimizing for operational constraints (5.1h training time vs 8.5h for standalone CNN). The intuition is straightforward: deep networks excel at encoding local spatial patterns from raw spectral bands, while ensemble methods handle tabular feature vectors and auxiliary covariates more robustly, so combining the two exploits complementary strengths.
Coupled simulation frameworks such as PLUS-InVEST extend classification into prospective territory, linking land-cover change models to ecosystem-service valuations and producing scenario outputs with 80–92% spatial accuracy [11]. These tools have found application in carbon-budget planning, where regulators need quantified estimates of how different development trajectories will affect sequestration capacity. Cellular automata remain a popular engine for spatial simulation because their local transition rules can be calibrated against observed change histories and adjusted to reflect policy constraints.
The GIS infrastructure that underpins these analyses has itself matured considerably. Contemporary platforms integrate raster processing, vector analysis, network modelling, and three-dimensional visualisation in unified environments, reducing the friction involved in moving from raw satellite acquisition to map products suitable for planning documents [12]. GIS-based longitudinal studies have documented a consistent pattern of green-space contraction in expanding cities, with attendant consequences for microclimate regulation, stormwater management, and biodiversity [13].
2.3 Interdisciplinary Aspects of Sustainable Urban Development
Environmental outcomes in cities are shaped as much by institutional arrangements as by the biophysical processes themselves. Land-use decisions reflect the interplay of zoning regulations, investment incentives, political priorities, and administrative capacity, which means that technical improvements to monitoring and modelling tools will have limited impact unless they are embedded in governance structures capable of acting on the information they produce. The connection between economic development strategies and environmental outcomes is well documented [4], and the digitisation of public administration [5] creates new channels through which spatial data can inform regulatory and planning decisions.
Building the human capacity to manage this interface is a recognised challenge. Effective use of AI-based spatial tools requires professionals who are simultaneously fluent in remote-sensing methods, machine learning practice, urban planning frameworks, and environmental science [6]. Graduate programmes that integrate these competencies remain rare, and interdisciplinary hiring in municipal agencies rarer still. Addressing this skills gap is increasingly recognised as a prerequisite for the kind of evidence-based urban governance that AI-GIS integration is meant to enable.
Taken together, remote sensing, GIS, and machine learning give urban planners a toolkit for more granular and spatially explicit environmental assessment than was previously feasible. Where transportation networks, utility corridors, land valuations, and population projections can be layered into a common analytical environment, planners gain the ability to evaluate development scenarios against measurable ecological thresholds before approvals are granted — a qualitatively different kind of risk governance than after-the-fact field inspection.
3 Research Methodology
3.1 Data Sources and Tools
Our review covered 2,847 publications indexed in Web of Science, Scopus, and GeoRef over the 2020–2025 period. To be included, a paper had to report empirical results on machine learning applications to urban land-use or ecosystem monitoring and supply enough methodological detail to permit quality assessment. Alongside the literature review, we conducted structured interviews with practitioners from municipal planning and environmental management offices; those conversations helped us check whether the modelling framework we were developing would fit into actual administrative workflows rather than remaining a laboratory exercise.
Satellite imagery formed the backbone of the spatial analysis. We drew on multispectral scenes from Sentinel-2 and Landsat-8, accessed through Google Earth Engine and the Copernicus Open Access Hub. Sentinel-2 was our primary source, valued for its 10–20 m spatial resolution, thirteen spectral bands, and five-day revisit frequency — a combination that supports intra-urban vegetation monitoring at the detail required here. Landsat-8 scenes were incorporated to extend the temporal record and provide independent cross-validation of surface reflectance estimates.
Pre-processing followed standard atmospheric and radiometric correction routines in SNAP and ENVI 5.6, after which we computed NDVI, NDBI, and NDWI indices to summarise vegetation density, built-up surface intensity, and water extent respectively. Spatial operations were handled in QGIS 3.28 and ArcGIS Pro 3.1. The machine learning pipeline ran in Python 3.10, with scikit-learn 1.2.2 for ensemble methods, TensorFlow 2.12.0 and PyTorch 2.0.1 for the neural network components, and GeoPandas, Rasterio, and GDAL for geospatial data management. Results were visualised with Matplotlib, Seaborn, and Plotly.
3.2 Spatial Modeling Methods
Land-use classification was built around a two-stage hybrid architecture [14, 15]. A ResNet-50 backbone, pre-trained on remote-sensing imagery, processed the stacked spectral bands in the first stage and produced spatial feature maps encoding local texture, shape, and spectral context. Those feature vectors were then passed to a Random Forest classifier, together with auxiliary geospatial variables — terrain derivatives, distance-to-road indices, and population-density rasters — which made the final category assignments. Keeping the feature extraction and classification stages separate preserved the interpretability that Random Forest variable-importance scores provide, while capturing spatial context that pixel-by-pixel classifiers characteristically miss (Fig. 1).
Fig.1. Al-based platform architecture for urban ecosystems management and ecological risk prediction
To model how land cover changed over the 2020 – 2025 window and to project likely trajectories to 2030, we fitted a cellular automaton framework to the observed transition sequences. Logistic regression determined cell-level transition probabilities as a function of proximity drivers — distance to roads, existing settlements, and water bodies — together with socioeconomic covariates such as residential land values and local population growth rates. The 2030 projections were produced under a business-as-usual assumption, providing a baseline trajectory against which the costs of inaction can be assessed.
Environmental risk was quantified using a multi-criterion scoring procedure built around five hazard categories: biodiversity loss and habitat fragmentation, soil degradation, air and water contamination, urban heat-island intensification, and heightened exposure to extreme weather. For each category we assembled indicator sets derivable from the available remote sensing and geospatial data, then scored and aggregated them via the analytical hierarchy process to generate composite risk surfaces at the city block level.
4 Research Results
4.1 Algorithm Performance Comparison
Table 1 summarises the accuracy, agreement statistics, training duration, and recommended application domain for each algorithm evaluated [8, 9]. The figures reflect averages across multiple cross-validation folds and should be interpreted as indicative of relative performance rather than absolute benchmarks, given that accuracy in operational settings depends on data quality and the specificity of the classification scheme.
Table 1. AI Algorithm Effectiveness for Urban Development Management
Algorithm Accuracy (%) Kappa Training Time (hrs.) Application
Random Forest 87.3 0.84 0.8 Basic urban ecosystem classification
Support Vector Machines 85.6 0.81 2.3 Small public space datasets
CNN (ResNet-50) 91.2 0.89 8.5 Complex spatial data classification
CNN + Random Forest 93.8 0.92 5.1 Universal spatial modeling
Gradient Boosting 88.4 0.85 1.2 Tabular urban development data
The CNN + Random Forest hybrid returned the strongest results on both accuracy (93.8%) and agreement (Kappa = 0.92), outperforming standalone CNN by 2.6 percentage points and standalone Random Forest by 6.5 points [8, 9]. What Table 1 does not immediately show is that the hybrid also trained faster than the standalone CNN: because the convolutional component handles only feature extraction rather than end-to-end classification, the total training burden at 5.1 hours falls well below the 8.5 hours required by ResNet-50 alone — a practically important difference for systems that must be periodically retrained on incoming imagery. Random Forest and Gradient Boosting, completing training in under two hours, remain the sensible choice wherever interpretability and rapid turnaround matter more than the marginal accuracy gain the hybrid provides.
4.2 Land-Use and Vegetation Dynamics
Table 2 records the land-cover composition at the start and close of the study period, together with net changes in area and percentage terms.
Table 2. Urban Territory Land Use Dynamics (2020 - 2025)
Land Use Category 2020 (ha) 2025 (ha) Change (ha) Change (%)
Green spaces and public space 3,124 2,737 −387 −12.4
Residential development 5,847 6,428 +581 +9.9
Commercial development 1,236 1,512 +276 +22.3
Industrial zones 892 964 +72 +8.1
Water bodies 724 682 −42 −5.8
Open soil and construction 512 1,012 +500 +97.7
The figure that stands out in Table 2 is the near-doubling of open soil and active construction sites (+97.7%), a signal that the city under study was passing through an intensive expansion phase rather than filling in isolated gaps in its built fabric. Green spaces absorbed a disproportionate share of the pressure, contracting by 387 ha (−12.4%), while residential and commercial footprints expanded by 9.9% and 22.3% respectively. These aggregate numbers, however, obscure a marked spatial gradient. NDVI mapping showed central districts recording index values of 0.28–0.35 — considerably below the study-area mean — while peripheral zones still held values of 0.55–0.68 despite actively losing natural cover to outward residential growth. Study-area mean NDVI slipped from 0.52 in 2020 to 0.47 by 2025, a numerically modest decline that nonetheless reflects a consistent and spatially uneven deterioration in vegetative condition.
Projecting the observed transition rates forward under unchanged policy and investment conditions, the cellular automaton model anticipates a further 8–11% reduction in green space by 2030 and a mean NDVI in the 0.42–0.44 range. These numbers carry the usual uncertainties of scenario modelling — they assume that zoning rules, land markets, and demographic pressures continue along their current trajectories, which is unlikely to hold precisely. Their value lies not in point-prediction but in setting a quantified reference against which the ecological cost of business-as-usual can be weighed.
5 Discussion of Results
5.1 What AI–GIS Integration Contributes
The results support the proposition that combining satellite observation with machine learning substantially widens the monitoring bandwidth available to urban environmental managers. The hybrid classifier's ability to process multi-date image stacks at city-wide scale within a single analytical pipeline — and to do so with an accuracy exceeding 93% — represents a step change relative to what periodic field surveys can practically achieve. The 72–93% improvement in spatial coverage and temporal frequency quantifies this advantage in terms that are directly relevant to the early detection of ecological change.
The early warning potential is particularly worth noting. Research on AI-enabled monitoring platforms has documented response-time reductions of 65–82% [10] compared with conventional reporting chains, and integrating IoT sensor data [16, 17] — particulate monitors, soil-moisture sensors, urban heat sensors — with satellite-derived land-cover maps can further shorten the interval between a detectable environmental shift and an institutional response. In the context of the vegetation losses observed here, such capability could allow green-infrastructure interventions to be targeted at the districts most at risk before degradation becomes irreversible.
5.2 Limitations and Open Questions
Three limitations deserve explicit attention. First, the accuracy figures reported here were obtained under controlled cross-validation conditions; operational performance in a production monitoring system, where imagery quality varies seasonally and by cloud cover, will likely be somewhat lower. Second, the interpretability of the hybrid model, while better than a standalone deep network, still falls short of what some administrative contexts require: the convolutional feature extraction stage does not lend itself to the kind of simple rule-based explanation that a planning authority might use to justify a zoning decision in a public consultation. Progress on explainable AI methods is directly relevant here [18]. Third, the cellular automaton projections are sensitive to the calibration period and to the assumption that transition probabilities remain stationary — an assumption that may not hold if major policy changes or economic shocks alter development patterns.
Together, these three observations outline a research agenda worth pursuing: modelling frameworks that trade some accuracy for genuine interpretability, validation studies that stress-test model performance against data from different climate zones and city morphologies, and uncertainty quantification methods that allow planners to reason about projections probabilistically rather than treating any single forecast as a fixed outcome.
5.3 Institutional Conditions for Adoption
Technical capability is a necessary but not sufficient condition for the adoption of AI-based spatial tools in urban governance. The evidence from digital-transformation initiatives in public administration [5] suggests that institutional readiness — defined by data-sharing protocols, interoperability standards, procurement frameworks, and political support — is at least as important as the sophistication of the algorithms involved. Municipalities that have successfully integrated geospatial analytics into routine planning tend to share certain characteristics: cross-departmental data governance structures, dedicated analytical units with both technical and policy competencies, and iterative implementation processes that allow tools to be refined in response to user feedback [19, 20].
On the financial side, the entry cost — hardware, software licences, and staff retraining — is real and will weigh heavily on smaller municipal administrations. That cost looks different, however, when placed alongside avoided expenditure: reactive repairs to infrastructure damaged by undetected environmental degradation, emergency responses to hazards that continuous monitoring might have flagged earlier, and the administrative inefficiencies that arise when permit review relies on outdated land-use data. Making these trade-offs visible through transparent cost-benefit accounting is arguably as important as the modelling work itself. Community engagement deserves a parallel investment: residents who understand the evidence behind land-use decisions are more likely to regard those decisions as legitimate, which matters for implementation [2].
6 Conclusions
The work reported here addressed a question that the urban environmental management literature has not fully resolved: how machine learning and GIS can be integrated in a way that improves both the technical quality of land-cover monitoring and the practical usability of its outputs for municipal decision-makers. Four findings stand out.
First, among the five classifiers evaluated, the two-stage CNN + Random Forest architecture produced the highest accuracy (93.8%, Kappa = 0.92) while also training faster than the standalone CNN. At 5.1 hours per training run, periodic retraining on incoming satellite imagery is operationally feasible, which matters for a system intended to support ongoing monitoring rather than one-off analysis.
Second, the spatial data assembled for the 2020–2025 window documents a pattern of green-space loss (−12.4%, or 387 ha) and vegetation-quality decline (NDVI: 0.52 → 0.47) that is concentrated in areas of intensive construction activity. Projections to 2030 suggest these trends will continue without deliberate policy intervention, with further NDVI decline to around 0.42–0.44 and additional green-space contraction of 8–11%.
Third, the platform architecture we propose assembles classification maps, multi-temporal change layers, and composite risk surfaces into a single decision-support environment. Connecting near-real-time satellite feeds and sensor networks to this environment creates the conditions for early warning functionality that, on the evidence of comparable systems, could reduce environmental threat response times by 65–82% relative to conventional monitoring chains.
Fourth, the institutional analysis underscores that technical tools alone will not determine whether AI-based spatial monitoring improves urban outcomes. Data governance, cross-departmental coordination, interpretable outputs, and staff with genuinely interdisciplinary competencies are all necessary complements to algorithmic accuracy. Future work should focus on developing frameworks that connect modelling outputs to specific decision points in planning and regulatory workflows, and on evaluating transferability across diverse urban contexts [21, 22, 23] and rapid urbanization [24, 25].
References
1. United Nations, Department of Economic and Social Affairs, Population Division: World Urbanization Prospects 2025: Summary of results. United Nations (2025).
2. Voronina, Y., Lopushynskyi, I., Grechanyk, B., Vahonova, O., Kondur, A., Akimov, O.: Economic and environmental component in sustainable development management. Calitatea 25(201), 7–14 (2024).
3. Chondrogianni, D., Stephanedes, Y.J., Saranti, P.-G.: Multiple-criteria decision analysis of urban planning methods towards resilient open urban spaces. In: Schrenk, M., et al. (eds.) REAL CORP 2022 proceedings, pp. 505–512. CORP.at (2022).
4. Stender, S., Bulkot, O., Yastremska, O., Saienko, V., Pereguda, Y.: Digital transformation of the national economy of Ukraine: Challenges and opportunities. Financial and Credit Activity: Problems of Theory and Practice 2(55), 333–345 (2024). https://doi.org/10.55643/fcaptp.2.55.2024.4328.
5. Sydorchuk, O., Bashtannyk, V., Terkhanov, F., Kravtsov, O., Akimova, L., Akimov, O.: Integrating digitization into public administration. Edelweiss Applied Science and Technology 8(5), 747–759 (2024).
6. Serhieiev, V., Voronina, Y., Zolotov, A., Akimova, L., Rovynska, K., Akimov, O.: Innovative competences in public administration. Sapienza International Journal 6(1), e25017 (2025).
7. Roy, P.P., Abdullah, M.S., Siddique, I.M.: Machine learning empowered geographic information systems: Advancing spatial analysis and decision making. World Journal of Advanced Research and Reviews 22(1), 1387–1397 (2024). https://doi.org/10.30574/wjarr.2024.22.1.1200.
8. Salhi, W., Tabiti, K., Honnit, B., Saidi, M.N., Kabbaj, A.: Hybrid deep learning architecture for land use: Land cover images classification with a comparative and experimental study. International Journal of Advanced Computer Science and Applications 13(12) (2022). https://doi.org/10.14569/IJACSA.2022.01312104.
9. Wang, S., Cao, J.: AI and deep learning for urban computing. In: Shi, W., et al. (eds.) Urban informatics, pp. 815–844. Springer (2021). https://doi.org/10.1007/978-981-15-8983-6_43.
10. Nandini, S.: AI-based early warning systems for climate change and extreme weather events. International Journal of Scientific Research and Engineering Trends 11(2) (2025). https://doi.org/10.61137/ijsret.vol.11.issue2.392.
11. Zhang, Y., Wang, X., Xue, M., Liu, S., Lu, Y., Liang, X., Shen, L., He, Y.: Change detection on remote sensing images using multidimensional attention network. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2025). https://doi.org/10.1109/JSTARS.2025.3594716.
12. Singh, A., Bhadauria, S.S.: Integration of remote sensing and GIS for environmental assessment. Environmental Reports 6(1), 35–40 (2024).
13. Ziliaskopoulos, K., Laspidou, C.: Using remote-sensing data to assess urban biodiversity. Landscape Ecology 39, 9 (2024).
14. Kamraju, M.: Geospatial technologies: A review of their applications and implications. Malaysian Applied Geography 3(2), 75–78 (2025). https://doi.org/10.26480/magg.02.2025.75.78.
15. Epuri, A.K.: The evolution and impact of GIS in environmental science. International Journal for Multidisciplinary Research 6(6), 30132 (2024). https://doi.org/10.36948/ijfmr.2024.v06i06.30132.
16. Di Palma, M., Rigillo, M., Leone, M.F.: Remote Sensing Technologies for Mapping Ecosystem Services: An Analytical Approach for Urban Green Infrastructure. Sustainability 16(14), 6220 (2024). https://doi.org/10.3390/su16146220.
17. European Space Agency: Sentinel-2. https://www.esa.int/Applications/Observing_the_Earth/Copernicus/Sentinel-2.
18. Myakala, P.K., Jonnalagadda, A.K.: Explainable AI for sustainability: Bridging trust, ethics, and accountability. In: Proceedings of RAISD 2025, pp. 563–576. Atlantis Press (2025). https://doi.org/10.2991/978-94-6463-787-8_44.
19. Zhidebayeva, A., et al.: Emerging frontiers in ecosystem and biodiversity monitoring using remote sensing. Global Journal of Environmental Science and Management 11(4), 1791–1818 (2025).
20. Naser, M.Z., Teymori Gharah Tapeh, A., Abdalla, J.: A review of generative artificial intelligence in civil and environmental engineering. Machine Learning for Computational Science and Engineering 1, 42 (2025). https://doi.org/10.1007/s44379-025-00041-z.
21. Ellison, R.: Smart cities and IoT: Integrating technology for sustainable urban development. ResearchGate (2025).
22. Vrahnakis, M., Kazoglou, Y.(I.), Pulido Fernández, M. (eds.): Grassland ecosystem services: Research advances and future directions for sustainability. MDPI (2024). https://doi.org/10.3390/books978-3-7258-0078-0.
23. Intergovernmental Panel on Climate Change: Working Group II. https://www.ipcc.ch/working-group/wg2/.
24. Kalphana, K.R., Ramela, K.R., Aanjankumar, S., Nagalakshmi, T., Sakthivel, B.: Review of remote sensing, GIS and soft computing tool for predicting the land-use and environment changes. Global NEST Journal 26(10), 06276 (2024). https://doi.org/10.30955/gnj.06276 .
25. Wang, P., Du, J., Wen, X., Hu, C., Ge, L., Huang, M.: Landsat-8 sensor and Sentinel-2 sensor data fusion with multiscale detailed information. IEEE Sensors Letters (2024). https://doi.org/10.1109/LSENS.2024.3499361 .
