Share
Export Citation
Aerial imagery segmentation of natural disaster-affected areas using deep convolutional networks for disaster assessment
Nugraha D.W.
Aip Conference Proceedings
Abstract
After a natural disaster, aerial imagery from an unmanned aerial vehicle (UAV) is an invaluable resource for impact assessment and disaster management. Traditionally, the current disaster assessment process is highly dependent on humans, requires considerable resources, and is very slow. Manual interpretation with ordinary vision is highly challenging for humans, especially given the sheer volume of aerial images that need to be processed quickly. This study proposes a technique to reduce human intervention using deep convolutional networks for aerial imagery semantic segmentation, focusing on disaster-affected areas. Semantic segmentation is used for object detection and part identification in high-quality post-disaster aerial imagery from UAVs. Four networks were trained, validated, and tested for the best outcomes on annotated aerial images: Tiramisu, Capsule Network (CapsNet), U-Net, and Pyramid Scene Parsing Network (PSPNet). The best semantic segmentation results are shown by the PSPNet model, which can detect and identify objects with irregular shapes and sizes, essential objects affected by natural disasters, and objects with small shapes. PSPNet successfully segmented areas affected by natural disasters using post-disaster aerial imagery with the complexity and diversity of objects appearing in the image. The segmentation results are then integrated into the Fuzzy C-Means clustering algorithm to automatically assess areas affected by natural disasters with four categories: areas not affected, lightly affected, moderately affected, and heavily affected. This study can perform an accurate disaster assessment from the segmentation results.