Urban Flood Mapping With Bitemporal Multispectral Imagery Via a Self-Supervised Learning Framework

Near realtime flood mapping in densely populated urban areas is critical for emergency response. The strong heterogeneity of urban areas poses a big challenge for accurate near realtime flood mapping. However, previous studies on automatic methods for urban flood mapping perform infeasible in near realtime or fail to generalize well to other floods, for several reasons. First, multitemporal pixel-wise flood mapping requires accurate image registration, hindering the efficiency of large-scale processing. Although automatic image registration has been investigated, precisely coregistered multitemporal image sequence requires time-consuming fine tuning. Additionally, the floods may lead to the loss of many corresponding image points across multitemporal images for accurate coregistration. Second, existing unsupervised methods generally rely on hand-crafted features for floodwater detection. Such features may not well represent the patterns of floodwaters in different areas due to inconsistent weather conditions, illumination, and floodwater spectra. This article proposes a self-supervised learning framework for patch-wise urban flood mapping using bitemporal multispectral satellite imagery. Patch-wise change vector analysis is used with patch features learned through a self-supervised autoencoder to produce patch-wise change maps showing potentially flood-affected areas. Postprocessing including spectral and spatial filtering is applied to these patch-wise change maps to remove nonflood related changes. Final flood maps and parameter sensitivities were evaluated using several performance metrics. Two flood events from areas with differing degrees of urbanization were considered: Hurricane Harvey flood (2017) in Houston, Texas, and Hurricane Florence flood (2018) in Lumberton, North Carolina. The proposed method shows strong performance for self-supervised urban flood mapping.