Abstract:
Change detection is a sub-field of remote sensing that aims to detect surface differences between two images taken at different times. It plays a significant part in detecting disaster damage and planning rescue operations. The advent of deep learning has lead to the development of many change detection solutions. Convolutional neural networks are at the core of recent approaches. As with most geographical phenomena, the spread of urban damage is more similar with buildings that in proximity. However, these networks only rely on local features and ignore the interactions and similarities between neighboring buildings. Also, it is important to map damage quickly whenever a new disaster occurs for an efficient response. Therefore it is not practical to wait for data to be annotated and models to be trained. Additionally, many structural building properties that are not based on proximity may impact the degree to which each building is damaged such as age and height. These properties can be very diverse especially in dense and irregularly urbanized cities and are not discernible from overhead imagery. In this work, we present a graph formulation for building damage change detection which enables learning relationships and representations from both local patterns and non-stationary neighborhoods that cannot be captured by traditional neural networks. We propose a novel architecture combining a Siamese convolutional neural network and a graph convolutional network which we train in a semi-supervised framework allowing the task to be performed with a small number of annotations and reducing the time and effort needed to obtain damage assessment. We also investigate a supervised variant and evaluate the possibility of generalizing to unseen disasters. We train and validate this approach on the xBD dataset. We also demonstrate this method on the Beirut Port Explosion and show that performance is improved by incorporating domain knowledge from building meta-features.
