High-resolution flood mapping improvements on global events

Abstract:

Public medium-resolution satellite data such as, Sentinel-1, Sentinel-2, and Landsat have been vital for observing global flood events. However, with a revisit time of 2-3 days utilizing medium-resolution satellites, flood events can be missed or obscured from cloud cover. Commercial satellites constellations, either tasked or orbiting, have the potential to fill gaps in observation of global hazards such as flooding. In Sri Lanka, a difficult location to capture flood imagery due to cloud cover, PlanetScope data was able to increase daily cloud-free coverage of the country an additional 10%. With the majority of optical satellite-based flood mapping utilizing shortwave infrared for flood detection, which the majority of smallsat missions lack, new methods for flood mapping from high resolution commercial satellite imagery are needed. Cloud to Street has focused on flood observations from high resolution commercial high resolution imagery in order to capture urban flooding, as well as fill in the gaps between public observation. Using PlanetScope, SkySat, and WorldView-3, high resolution imagery was captured during multiple urban and rural flood events in Mozambique, Sudan, Niger, Democratic Republic of Congo, and Gambia. Additionally, using the Sen1Floods11 dataset (Bonafilia et al, 2020), we extended the flood events used in this study to US, Cambodia, Ghana, and Spain to include overlapping PlanetScope, Sentinel-2, and Sentinel-1 scenes. Using a random forest classifier we trained a local model for water probability prediction for each flood event, before refining it to a global model which was applied across all test locations. Additionally, we will present the limitations of high resolution flood mapping that we encountered, including building shadows in urban areas, cloud and cloud shadow masking, and well as solutions we took to overcome them.