A deep transfer learning-based damage assessment on post-event very high-resolution orthophotos

Abstract

Post-disaster building damage assessment is an important application of remote sensing. The increasing resolution of remote sensing imaging systems and state-of-the-art deep learning networks has facilitated damage assessment. However, most existing methods in the literature concentrate on damage/non-damage classification only in specific disaster types/areas using pre- and post-event images. Furthermore, site visits are inevitable to assess the level of damage to structures. Therefore, the main objective of this study was to utilize deep transfer learning over a pre-trained network and extend it to a damage assessment framework. The network is fine-tuned to identify four different damage levels: non-damage, minor damage, major damage, and collapsed, using only post-event images taken from different disaster types/areas. To evaluate the proposed framework, we carried out three experiments on Hurricane Irma in Sint Maarten, Hurricane Dorian in Abaco Islands, and Woolsey Fire using post-event orthophotos derived from unmanned aerial vehicle (UAV) images. The results of over 80% overall accuracy confirm that with a structured learning scenario, it is possible to use transfer learning on very high-resolution remote sensing images to classify the level of structural damage.