Prior Information in Support of Deep Learning Methods to Map Floodwater in Urbanized Areas

Abstract

Due to the complexity of urban environments, the synthetic aperture radar (SAR) based mapping of floodwater is impacted by different factors such as water depth, building orientation and the density of built-up areas. Several studies have proven that both SAR multitemporal intensity and interferometric SAR (InSAR) coherence data acquired in VV and VH polarizations support the urban flood mapping. We propose a deep learning (DL) based method using dual-polarization Sentinel-1 multitemporal intensity and coherence data combined with prior information to map floodwater in urbanized areas. The proposed method aims at mapping flooded areas in urbanized regions and bare soils/sparsely vegetated areas within the entire frame of a Sentinel-1 image. In this paper, our method is evaluated for the Houston (US) urban flood event in 2017 via a qualitative and quantitative comparison with two established DL models. The proposed method has the lowest number of false alarms in flooded urban areas, indicating that the prior information from the probabilistic urban mask is valuable.