The document presents a PhD thesis defending research on inferring the geolocation of non-geotagged tweets at a fine-grained level to enhance applications like traffic incident detection. It argues that previous methods were limited by working only at coarse levels and proposes a novel approach using individual tweet analysis, majority voting, and learning to rank methods to improve accuracy. The findings demonstrate that these approaches significantly improve the geolocation accuracy of tweets, which is vital for effective emergency management and traffic monitoring.