Geo Disaster AI Net: AI Framework for Risk Detection and Resilience Analysis

Abstract

To ensure resilient societies, correct and timely classification is essential as urban and rural areas continue to feel the impacts of natural disasters such as floods, cyclones, earthquakes and wildfires. The traditional approaches to the detection of disasters are manually based or rely on the prediction ability of only one model, and are hard to be generalized to various environments and types of disasters. GeoDisasterAINet is a multi-stage ensemble method and uses the Natural Disaster Image Dataset from Kaggle, comprising of tagged images of various disaster events. The first models (ERI2025, DRI-2025 and DE-2025) are trained without using any kind of further pre-processing of the data. SMOTE and conventional scaling are used for the advanced models, ERI2025 + XGBoost, DRI-2025 + XGBoost and DE-2025 + XGBoost for balanced training. In the last stage, multiclass SVM is paired with XGBoost improved models, where just standard scaling is employed, to improve classification accuracy. Convolutional neural networks, Xception, ResNet50 and Hybrid Ensemble of Xception and ResNet50 are also used for extracting deep features. The accuracy, precision, recall and F1-score assessment indicates that ResNet50 is the best model with an accuracy of 99.2% on the test set. The methods of explainable AI, like Grad-CAM and LIME, offer heatmaps and feature importance visualizations to gain insight into the model's decisions efficiently. A flask based interface is made available to submit photos and receive real time predictions with confidence scores with ease. The framework provides, in general, a high accuracy, interpretable and deployable disaster classification for resilient urban and rural communities. “Keywords— Natural Disaster Classification, Deep Learning, Hybrid Ensemble, Explainable AI, XGBoost, Multiclass Support Vector Machine (SVM)”.