Over the last decade, disasters have impacted more than 1.5 billion people worldwide, causing more than half a million deaths and resulting in major economic losses for the affected areas. The most impacted are those residing in the least developed countries and small island developing states. Accessing relevant information in a timely manner is key for providing an accurate and efficient response to these situations. Traditionally, satellite imagery has been a common approach, yet analyzing these images is both costly and time-consuming. In contrast, recent advancements indicate that social media plays a pivotal role in delivering timely information. On these platforms, users share critical updates about disasters, including the extent of damage to infrastructure, severity of the damage, and the condition of victims. Despite its potential, the utility of social media data in operational contexts is limited by challenges such as filtering irrelevant information and the lack of location in the content.

 

This thesis addresses the problem to get relevant information and retrieve location from social media content. It investigates the role of advanced machine learning in two directions. Firstly, it seeks to enhance information classification by applying sophisticated machine learning techniques to improve the accuracy of categorizing social media data into vital response categories, such as information relevance, damage severity, and urgent needs. Secondly, it involves predicting the geolocation of social media content to aid disaster response. This thesis employs advanced machine learning algorithms to infer the geographical origins of social media posts lacking explicit location data, aiming to accurately identify areas in need during disaster situations, thereby optimizing response efforts. This problem-driven research leverages and analyzes real-world datasets from social media platforms. Motivated by real-world problems, the thesis is conducted in collaboration with various practitioners and institutions, including the Joint Research Center of the European Commission, Digital Humanitarian Organization VOST Portugal, Artificial Intelligence Research Center CSIC Spain, and Politecnico di Milano, Italy. As a contribution to the field, this thesis offers an approach, design, and implementation of advanced machine learning to analyse social media data. The results are expected to assist policymakers in understanding the technical approach to applying social media analysis to enhance disaster response.