Damage Evaluation Following Natural Disasters Using Deep Learning

Natural catastrophes including flooding, tornadoes, earthquakes, and wildfires have been occurring more frequently over the past few decades as a result of global warming and climate change. Therefore, it is more crucial than ever to give emergency response workers accurate and timely information to enable them to respond to crises effectively. Among the many pieces of information required for disaster response and management, it is crucial that rescue workers are promptly notified of the location and extent of a building's destruction in order to maximise the effectiveness of their efforts. Nevertheless, despite significant efforts, problems with picture classification for disaster response still exist. In this study, a potential deep learning-based method is put forth for identifying damaged buildings in high-resolution satellite photos. It solves the issue of limited training data common in many remote sensing applications by using generic data augmentation. It is suggested that a pretrained model be used in conjunction with transfer learning as a fine-tuning method for the relevant task. The trials with images of Port-au-Prince, Haiti showed that the suggested strategy works well with sparse training data. With enriched training data, the Convolutional Neural Network (CNN) model can detect damaged buildings with an accuracy of 83%, compared to only 53% with the original training data. The focus of future study will be on investigating automated ways to obtain larger training datasets and model generalisation by researching more reliable data augmentation strategies.