Machine Learning Models for Mining Social Media Data for Effective Natural Disaster Assessment

Abstract

Satellite technology has emerged as a key tool for effective management and assessment of natural disasters. However, the challenge of accurately estimating impacted populations and assessing building damage, often obscured from aerial views, persists. To address this, the integration of imagery and textual data from social networks offers a promising solution. This study employs Twitter and Flickr datasets, using SVM, CNN, XGBoost, Logistic Regression, and Gradient Boost to extract insights. The sentiment analysis component categorizes disaster-affected individuals' emotions as panic, neutral, or non-panic. The Logistic Regression model excels in text classification, boasting an impressive 88.99% accuracy on the test dataset and 83.45% on training. The framework introduces an Aid analysis model, which gives us an impressive accuracy of 83.16% for the classification of tweets based on the aid sought by people through tweets. Image classification employs CNN, achieving 83.29% accuracy to comprehend disaster impact visually. Given real-time social media responses, the system assists government and organisations promptly, prioritising assistance. It serves as a dependable aid resource, enabling efficient responses tailored to affected communities. Thus, this approach holds potential to significantly enhance disaster relief efficacy.