Human-Guided Flood Mapping: From Experts to the Crowd

Abstract

Hurricane-induced flooding can lead to substantial loss of life and huge damage to infrastructure. Mapping flood extent from satellite or aerial imagery is essential for prioritizing relief efforts and for assessing future flood risk. Identification of water extent in such images can be challenging considering the heterogeneity in water body size and shape, cloud cover, and natural variations in land cover. In this effort, we introduce a novel cognitive framework based on a semi-supervised learning algorithm, called HUman-Guided Flood Mapping (HUG-FM), specifically designed to tackle the flood mapping problem. Our framework first divides the satellite or aerial image into patches leveraging a graph-based clustering approach. A domain expert is then asked to provide labels for a few patches (as opposed to pixels which are harder to discern). Subsequently, we learn a classifier based on the provided labels to map flood extent. We test the efficacy and efficiency of our framework on imagery from several recent flood-induced emergencies and results show that our algorithm can robustly and correctly detect water areas compared to the state-of-the-art. We then evaluate whether expert guidance can be replaced by the wisdom of a crowd (e.g., crisis volunteers). We design an online crowdsourcing platform based on HUG-FM and propose a novel ensemble method to leverage crowdsourcing efforts. We conduct an experiment with over $50$ participants and show that crowdsourced HUG-FM (CHUG-FM) can approach or even exceed the performance of a single expert providing guidance (HUG-FM).