Flood Water Segmentation from Crowdsourced Images

In the United States, 176 people were killed by flooding in 2015. Along with the loss of human lives is the economic cost which is estimated to be $4.5 billion per flood event. Urban flooding has become a recent concern due to the increase in population, urbanization, and global warming. As more and more people are moving into towns and cities with infrastructure incapable of coping with floods, there is a need for more scalable solutions for urban flood management.The proliferation of camera-equipped mobile devices have led to a new source of information for flood research. In-situ photographs captured by people provide information at the local level that remotely sensed images fail to capture. Applications of crowdsourced images to flood research required understanding the content of the image without the need for user input. This paper addresses the problem of how to automatically segment a flooded and non-flooded region in crowdsourced images. Previous works require two images taken at similar angle and perspective of the location when it is flooded and when it is not flooded. We examine three different algorithms from the computer vision literature that are able to perform segmentation using a single flood image without these assumptions. The performance of each algorithm is evaluated on a collection of labeled crowdsourced flood images. We show that it is possible to achieve a segmentation accuracy of 80% using just a single image.