Transformer-Based Flood Detection Using Multiclass Segmentation

Joo-Chan Park, Dong-Geon Kim, Ji-Ro Yang, Kyo-Seok Kang

Abstract: Natural disasters such as hurricanes, tornadoes, earthquakes and floods cause enormous human and economic damages every year. In order to provide fast and accurate useful information to first responders (victims and rescuers), technology that can predict the extent of damage is required. We propose an enhanced transformer-based multiclass flood detection model that can predict flood occurrence while classifying roads and buildings. We also introduce our loss function and road noise removal algorithm that can achieve high performance in road evaluation metrics APLS. The proposed methods show superiority through the SpaceNet8 Challenge data.