Re-learning observed flood extents can improve remote sensing products

Abstract

Flood extent data is critical for both response action and subsequent impact modelling. Satellite-based emergency mapping products typically exhibit low sensitivity in vegetated and urban areas, though. As the information content of spaceborne sensor data is limited in these areas, e.g. due to shadow and overlay effects, advances in processing can primarily reduce overdetection, but hardly avoid underdetection. We present a novel procedure of re-learning the observed flood extent from different features (i.e. not derived from sensor data) using a one-class classifier (OCC). This allows us to treat the entire flood mask as training region, and extrapolate into unobservable areas. The approach was tested for hurricane Harvey in Houston, where three satellite-based flood masks of varying quality were available to us, as well as a 50 cm resolution aerial image for validation. An assessment of the initial masks showed that the standard emergency mapping product effectively detected only open water, while the high quality products detected about 40% flooded urban areas. All of these products could be improved by our presented modelling approach, with the best models raising the k score by 0.2 (high quality product) to 0.7 (standard product).