Human-guided Flood Mapping on Satellite Images

Flooding is responsible for substantial loss of life and economy. Flood mapping, the process of distinguishing flooded areas from non-flooded areas during and after a disaster, can be very useful in guiding first response resources in a disaster situation, and in assessing flood risk in future disaster scenarios. This paper involves the use of image segmentation methods and human guidance to provide a mechanism for flood mapping. Previous image segmentation methods do not work well in flood mapping because they are designed to segment objects out of an image, where there are only a few objects, e.g., foreground-background segmentation. However, satellite images of flooded areas often contain hundreds to thousands of large and small water areas that need to be identified. Therefore, we design a semi-supervised learning algorithm specifically to tackle the flood mapping problem. We first divide the satellite image into patches using a graph-based approach depending on the proximity and intensity of pixels. We then classify each of the patches in an interactive and incremental way, where each time the user is asked to label a few patches and we learn a classifier to automatically classify other patches into water area or land area. We run our algorithm on satellite images of Chennai, India during the 2015 Chennai flood period. The results show that our algorithm can robustly and correctly detect water areas compared to baseline methods. We compare the segmentation results of post-flood with pre-flood and conduct an effective flood evolution analysis.