Automatic damaged building detection and damage assessment from satellite images

Abstract

Natural disasters cause severe damage to the buildings and injuries to people in the affected areas. After a disaster, one of the initial steps is to give a quick response and relief. To identify the extent of damages to the buildings is useful for High Availability Disaster Recovery (HADR). The affected areas need a quick response, so it is time productive to identify these areas remotely. In recent years, remote sensing techniques use satellite images or drone to collect information about the damaged area. We need to automate this process so we can benefit from computer vision approaches introduces in recent years. The existing computer vision techniques are used for building damage detection, and assessment works in two stages. First, detect the buildings using the object detection model and assess the building damage using classification. The multi-task methods do not end to end trainable and suffer from poor results. We intend to use the Mask RCNN model, which simultaneously detects building in the image and classify them and outputs class label with scores. It is an extension of Faster RCNN by adding a mask branch that predicts segmentation masks for each image’s objects. So All these tasks work simultaneously in a single network, and we do not need to compute localization and classification separately. We use the XBD dataset, which provides building polygons, Joint damage scale, and damage labels and it is heavily skewed. It covers a wide variety of disasters type, which is lacking in the available literature. We tested Mask RCNN model on a large scale XBD dataset and achieved better results then the baseline