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

Abstract

Urban and rural areas are increasingly vulnerable to natural disasters, including floods, cyclones, earthquakes, and wildfires, requiring accurate and timely classification for resilient communities. Traditional disaster detection approaches often rely on manual analysis or singlemodel predictions, which struggle to generalize across diverse environments and disaster types. Using the Natural Disaster Image Dataset from Kaggle, comprising labeled images of multiple disaster events, GeoDisasterAINet employs a multistage ensemble framework. Initial models, ERI-2025, DRI2025, and DE-2025, are trained without additional preprocessing, while SMOTE and standard scaling are applied to enhanced models, ERI-2025 + XGBoost, DRI-2025 + XGBoost, and DE-2025 + XGBoost, for balanced training. In the final stage, multiclass SVM is combined with XGBoostenhanced models, with only standard scaling applied, to improve classification accuracy. Convolutional neural networks, Xception, ResNet50, and a Hybrid Ensemble of Xception and ResNet50, further extract deep features. Evaluation using accuracy, precision, recall, and F1-score highlights ResNet50 as the best-performing model with 99.2% test accuracy. Explainable AI techniques, including GradCAM and LIME, generate heatmaps and feature importance visualizations to interpret model decisions effectively. A Flaskbased interface enables users to upload images and receive real-time predictions with confidence scores. Overall, the framework provides high-accuracy, interpretable, and deployable disaster classification for resilient urban and rural communities.