Deep learning-based detection of affected areas and buildings from aerial images in disaster zones

Natural disasters are usually defined as sudden and unpredictable natural events that have a serious impact on human life and the environment. Earthquakes, landslides, floods, forest fires, storms, hurricanes and tsunamis are among the most important types of disasters that have a profound impact on society. Search and rescue operations following such events are critical to the protection of human life and the safety of emergency responders. These operations involve several important phases, including accessing the disaster area, assessing structural damage, rescuing trapped victims, providing urgent medical assistance and meeting basic needs. In recent years, the use of remote sensing imagery has become an indispensable tool for post-disaster damage assessment. This work aims to accurately recognise the regions affected by a disaster by developing novel segmentation and classification models tailored to different types of natural disasters. The FASegNet model achieved mean Intersection over Union (mIoU) values of 84.3% and 84.5% for the "Flood Area" and"Water Bodies" datasets, respectively, demonstrating high performance in flood and tsunami scenarios. The LandslideSegNet model achieved an overall accuracy of 97.60% and an mIoU value of 73.65% for the Landslide4Sense dataset, enabling precise delineation of landslide areas. For building detection, the ABDSegNet model achieved mIoU values of 90.94% and 79.59% for the "WHU Building" and "Inria Aerial Image Labelling" datasets, respectively, and delivered successful results in the damage assessment after the 2023 Kahramanmaraş earthquake. The WBSegNet model, which was developed to study the impact of climate change on water resources, performed strongly with an accuracy of 95.36% and an mIoU value of 85.51%, proving its applicability in monitoring lakes in Turkey. A hybrid model based on Vision Transformer (ViT) was proposed for early detection of forest fires. By integrating features extracted from three different ViT architectures and filtering them using Minimum Redundancy Maximum Relevance (mRMR) optimisation, the model achieved 100% accuracy. The results show that the developed segmentation and classification models provide very accurate and efficient results in different situations.