Semi-Supervised Learning for Improved Post-disaster Damage Assessment from Satellite Imagery

摘要

The devastating aftermath of a natural disaster is often challenging to assess, and inaccuracies are bound to occur when an assessment is done manually due to the inevitable human-in-the-loop errors. Timely and accurate evaluation of the extent of damages is often needed to effectively deploy resources to hard-hit areas, save lives, and facilitate adequate planning towards disaster recovery. The commonly used supervised learning approaches have made a considerable improvement in assessing natural disasters. However, quickly implementing supervised classification is still challenging due to the complexity of acquiring many labeled samples in the aftermath of disasters. In this paper, we propose a: ⅰ) two-stream high-resolution network (HRNet) that takes a pair of pre- and post-disaster images and ⅱ) semi-supervised framework for improving the generalizability of current methods to other housing styles. The proposed method comprises of two parts: a multi-class deep learning model, and a pseudo-label generator and refinement module. By harnessing information from a large amount of unlabeled data and aerial imagery, our approach can outperform its base model. Experimental results on the xView2 dataset demonstrate that the proposed framework improves the performance of our two-stream model for unseen satellite images depicting a scene before and after a disaster.