A Unified Neural Network and XGBoost Learning Model for Disaster Prediction and Management

Abstract

Accurate disaster prediction plays a critical role in disaster management and mitigation by reducing the loss of life and property. Traditional machine learning approaches often struggle to handle complex disaster patterns, class imbalance in datasets, and challenges in multi-class disaster classification, which reduces prediction reliability. To address these issues, a hybrid disaster prediction framework based on Neural Networks and XGBoost is proposed for the multi-class classification of natural disasters. In the proposed framework, neural networks are employed for high-level feature extraction from real-world disaster data, while XGBoost (Extreme Gradient Boosting) performs efficient classification using gradient-boosted decision trees. The dataset, sourced from the EM-DAT International Disaster Database, comprises 4,543 samples with 31 features encompassing wildfires, floods, and earthquakes. The Synthetic Minority Oversampling Technique (SMOTE) is applied exclusively to the training data to address class imbalance and improve the model's generalization capability. Experimental results demonstrate improved performance, with the proposed Neural-XGBoost (N-XGB) model achieving 94.72% accuracy, 98.57% ROC-AUC, 90.49% F1-score, and 89.65% precision, outperforming traditional machine learning models. The proposed system provides a reliable and efficient disaster prediction approach that supports better disaster preparedness and decisionmaking.