Neural-XGBoost A Hybrid Approach for Disaster Prediction and Management Using Machine Learning

Aim

          To develop a four-class disaster prediction system that uses SMOTE for class balancing, evaluates four advanced machine learning models, selects the best-performing classifier, and deploys it through an interactive web interface

Abstract

         Accurate disaster prediction is essential for mitigating risks and improving emergency response. This study proposes a SMOTE-enhanced multi-model machine learning framework that classifies disaster events into four distinct categories. The system integrates four advanced algorithms—LightGBM, Random Forest, Extra Trees, and CatBoost—combined with comprehensive preprocessing and class balancing using the Synthetic Minority Oversampling Technique (SMOTE). Each model is trained on the balanced dataset and evaluated using accuracy and F1-score, after which the best-performing model is automatically selected as the final predictor. A lightweight web interface is developed to allow users to input disaster-related parameters and receive real-time predictions. By combining SMOTE with ensemble-based learning, the proposed system significantly improves classification performance on minority disaster classes and provides a practical, deployable solution for real-world disaster analytics.

Proposed System

         The proposed system trains four advanced machine learning models—LightGBM, Random Forest, Extra Trees, and CatBoost—on a four-class disaster dataset enhanced with SMOTE. SMOTE generates synthetic samples for minority classes, ensuring equal representation across all categories. After preprocessing and balancing, each model is evaluated, and the best-performing model is selected for final deployment. A user-friendly web interface enables real-time predictions using this optimized model.

Advantage

• SMOTE ensures equal representation of all four disaster classes
• Improved accuracy and F1-score, especially for minority classes
• Multi-model comparison ensures the best model is selected
• Fast and efficient predictions
• Easy-to-use web interface
• Reliable classification even with originally imbalanced datasets