Disaster Response Coordination and Resource Management: An AI-Based Ensemble Approach for Natural Disaster Classification

Abstract

Natural disasters including floods, wildfires, and earthquakes cause immense human and economic losses annually. Timely and accurate disaster classification is vital for enabling emergency agencies to pre-position resources and coordinate response efforts. This paper presents an intelligent machine-learning ensemble system that classifies natural disaster events into three categories—Flood, Wildfire, and Earthquake—using the EM-DAT global disaster database comprising 10,431 records spanning 1900 to 2023. A dedicated feature-engineering pipeline derives 19 semantically meaningful features from six raw numerical inputs. The severe class imbalance (Flood: 3,837; Earthquake: 1,087; Wildfire: 374) is corrected using SMOTE, producing a balanced training set of 9,204 samples. A soft-voting ensemble of a Multi-Layer Perceptron (MLP), XGBoost, and Random Forest achieves a test accuracy of 79.4% on a held-out set of 1,061 samples with strict no-data-leakage methodology, confirmed by 5-fold crossvalidation yielding 77.1%–81.8%. The system is deployed as an interactive Streamlit web application providing real-time predictions, probability distribution charts, feature radar visualisations, and natural-language explanations, serving as a decision-support tool for emergency management agencies.