The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Publications Copernicus
Articles | Volume XLVIII-3/W1-2022
27 Oct 2022
 | 27 Oct 2022


Z. Wang, F. Zhang, and C. Wu

Keywords: Remote Sensing, Disaster, Transfer Learning, Building Damage, Sample Library, U-Net

Abstract. In recent years, the analysis of satellite remote sensing data using deep learning methods has become an important solution in the field of disaster response. The sample library can provide knowledge support for deep learning tasks in a research area, and its sample size and labeling accuracy directly determine the model effectiveness. However, the traditional sample library is often constructed by manual annotation, which is time-consuming and less automated. This research explores a transfer learning method to automate the construction of a sample library of building damage. We used models instead of manual work to automatically annotate the damage information of disaster images, combined with an auxiliary manual inspection to achieve high quality and dynamic optimization of the sample library. Following are three main aspects of research work. (1) This study extensively collected remote sensing images related to disaster events, and carried out data pre-processing work. (2) This study built a building damage information identification model based on the U-Net framework, which applied pre-trained backbone model and was formally trained on xBD dataset. (3) With CycleGAN to implement color style transfer between RS images collected from two different data sources, we constructed a high-quality building damage sample library, mainly by automatic machine labeling and supplemented by manual sampling verification. Finally, we successfully constructed a sample library of disaster-damaged building images, and conducted several rounds of manual precision sampling work and dynamic optimization work. The number of image slices in the sample library reached 12,834, and the overall accuracy is higher than 0.80 under the manual inspection of 200 samples.