Geo Disaster AI Net: An Explainable Deep Ensemble Framework for Real-Time Urban and Rural Disaster Classification and Resilience

Abstract

Naturaldisastersposeescalating threatstobothurban andruralcommunities,necessitating intelligent,real-time classification systems that can support rapid response and resilience planning. This paper presents GeoDisasterAINet, an explainable deep ensemble framework integrating a three-stage hybrid machine learning pipeline—Random Forest (RF) feature learning, probability-based feature transformation, and Support Vector Machine (SVM) classification—achieving robust disaster detection with 82.46% accuracy, 82.1% precision, 82.4% recall, and an AUC-ROC of 0.881. The framework processes 6,857 multi-hazard records spanning floods, earthquakes, cyclones, and landslides across India (2010–2023), achieving approximately 100ms inference time suitable for real-time deployment. Our hybrid architecture demonstrates a 10.87% improvement in error recovery compared to standalone Random Forest baselines, successfully recovering 46 additional disaster cases. We integrate explainability through feature importance analysis, revealing that severity (0.225), fatalities (0.198), and a?ected area (0.165) are the most critical predictors. Furthermore, we propose a quantitative resilience framework incorporating nighttime light (NTL) data, network functionality curves, and composite capital-based metrics to di?erentiate urban and rural disaster impacts. This work bridges the gap between high-accuracy disaster classification and actionable resilience insights, providing a scalable, interpretable solution for heterogeneous geographic contexts.

Introduction

Natural disasters such as floods, earthquakes, cyclones, and landslides cause severe human, economic, and infrastructure damage. With climate change increasing the frequency of extreme events, there is a need for intelligent systems capable of real-time disaster detection and classification. Traditional disaster management methods rely on manual reporting, satellite analysis, and physical simulation models, which are slow, difficult to scale, and unable to efficiently process large multi-source data streams.

Existing machine learning approaches for disaster classification face several issues: limited performance on noisy and imbalanced datasets, lack of explainability in “black-box” models, and limited integration of urban–rural resilience metrics needed for targeted disaster response. Additionally, emergency management requires very fast predictions, creating a challenge in balancing accuracy and low inference time.

To address these problems, the study proposes GeoDisasterAINet, a hybrid machine-learning framework that combines Random Forest (RF) and Support Vector Machine (SVM) in a three-stage ensemble pipeline for multi-hazard disaster classification. The system processes heterogeneous data sources such as IoT sensors, satellite feeds, and reports while providing explainable predictions and resilience insights. The model also incorporates resilience metrics such as nighttime light (NTL) data, network functionality curves, and composite capital-based indices to differentiate disaster impacts in urban and rural areas.

The framework was trained and tested on 6,857 disaster records from India (2010–2023) covering floods, earthquakes, cyclones, and landslides. After preprocessing and feature engineering, the hybrid RF–SVM pipeline achieved 82.46% classification accuracy with about 100 ms inference time, outperforming standalone RF and SVM models. The system also improved error recovery by 10.87%, correctly identifying additional disaster cases that single models missed.

Explainable AI techniques were used to analyze feature importance, revealing that severity, fatalities, and affected area were the most influential factors in disaster prediction. The study also examined urban and rural resilience patterns, showing that urban areas usually recover faster (7–21 days) due to better infrastructure, while rural areas often require 30–90 days for recovery because of accessibility challenges.

Conclusion

Overall, the research demonstrates that hybrid ensemble learning combined with explainable AI and resilience metrics can provide accurate, fast, and interpretable disaster detection systems. The proposed framework supports real-time emergency decision-making, disaster risk assessment, and targeted response planning, making it suitable for large-scale disaster management applications.

References

[1] Raju et al., “GeoDisasterAINet: An explainable deep en-semble framework,” IEEE Access, 2025. DOI: 10.1109/ac- cess.2025.3574451 [2] Humaira et al., “DX-FloodLine: End-To-End Deep Explainable Pipeline for Real Time Flood Scene Object Detection,” IEEE Access, 2023. DOI: 10.1109/access.2023.3321312 [3] Tanvir et al., “Forest Fire Detection Using Ensemble Deep Learning Model With XAI,” 2024. DOI: 10.1109/com- pas60761.2024.10796582 [4] Hanetal.,“ResearchonthePathwaysofInformationTechnology Empowering Cities to Enhance Disaster Resilience,” 2024. [5] Kumar et al., “Ensemble Machine Learning And Deep Learning Framework For Flood Susceptibility Mapping In The Trans- boundary Rapti River Basin,” 2025. DOI: 10.21203/rs.3.rs- 7393798/v1 [6] Cassottanaetal.,“PredictingResilienceofInterdependentUrban Infrastructure Systems,” IEEE Access, 2022. DOI: 10.1109/AC-CESS.2022.3217903 [7] Liu et al., “EQResNet: Real-Time Simulation and Resilience AssessmentofPost-EarthquakeEmergencyHighwayTransporta- tion Networks,” Computation, vol. 13, no. 8, p. 188, 2025. DOI: 10.3390/computation13080188 [8] Abdel-Mooty et al., “Data-Driven Community Flood Resilience Prediction,” Water, vol. 14, no. 13, p. 2120, 2022. DOI: 10.3390/w14132120 [9] Duetal.,“Cross-DomainCriticalityandRobustnessAnalysis of Interdependent Urban Networks,” 2025. DOI: 10.31219/osf.io/ya8z5_v1 [10] Ma, “Urban-Rural Disparity in Disaster Resilience: Harnessing Nighttime Light Data and Interpretable Machine Learning,” 2024. DOI: 10.5194/egusphere-egu24-21837 Garg,“Real-TimeDisasterResponsewithAIOps,”Int. [11] J. Multidisciplinary Res. Growth Eval., 2024. DOI: 10.54660/.ijmrge.2024.5.5.1101-1107 [12] Jankovic et al., “UAV-Assisted Real-Time Disaster Detection Using Optimized Transformer Model,” 2025. [13] Martí et al., “A multitask deep learning model for real-time deployment in embedded systems,” 2017. [14] Sirietal.,“FloodBot-RealTimeInnovativeFlood Resilience Through Machine Learning,” 2025. DOI: 10.1109/iccsp64183.2025.11088698 [15] Cheng et al., “Uncertainty-aware convolutional neural network forexplainableAI-assisteddisasterdamageassessment,”Struct. ControlHealthMonit.,2022.DOI:10.1002/stc.3019 [16] Phanetal.,“Real-timewildfiredetectionwith semantic explanations,” Expert Syst. Appl., 2022. DOI: 10.1016/j.eswa.2022.117007 [17] Shi et al., “Urbanization Impact on Regional Sustainable De- velopment:ThroughtheLensofUrban-RuralResilience,”Int. [18] J.Environ.Res.PublicHealth,vol.19,no.22,p.15407,2022.DOI:10.3390/ijerph192215407 [19] Zhangetal.,“UrbanFloodRiskAssessmentthroughtheIntegra- tion of Natural and Human Resilience Based on Machine Learn- ing Models,” Remote Sensing, 2023. DOI: 10.3390/rs15143678 [20] Lankipalle, “SwarmFusion: Revolutionizing Disaster Response withSwarmIntelligenceandDeepLearning,”arXiv,2025.DOI: 10.48550/arxiv.2507.00005 [21] Sharmaet al., “Artificial Intelligence and Machine Learning for ResilientTransportationInfrastructure,”CureusJ.Comput.Sci., 2025. DOI: 10.7759/s44389-025-09490-2 [22] Linardos et al., “Machine Learning in Disaster Management: RecentDevelopmentsinMethodsandApplications,”Mach. [23] Learn.Knowl.Extr.,2022.DOI:10.3390/make4020020 [24] Suiet al., “Measurement of Coastal Marine Disaster Resilience and Key Factors with a Random Forest Model,” Water, vol. 14, no. 20, p. 3265, 2022. DOI: 10.3390/w14203265 [25] Mishra et al., “Artificial Intelligence and Machine Learning for PredictingNaturalDisastersinSmartCityInfrastructure,”2024. [26] Basit et al., “SMART AI-ML Frameworks for Early Detection andRiskMitigationofFloods,Droughts,andWildfires,”2024. [27] Pasupuleti,“Real-TimeDisasterRiskAssessment Using Predictive Geospatial Analytics,” 2025. DOI: 10.62311/nesx/rphcrdsbdapa3

Copyright

Copyright © 2026 Thanish Kumar T., Y. Vishnu Vardhan, A. Sai Rohitha, P. Ruquitha, D. Sharmila . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.