DeepFuse: Towards Frequent Flood Inundation Monitoring using AI and EO

Despite the increase in the number of Earth Observation satellites with active microwave sensors suitable for flood mapping, the frequency of observations still limits adequate characterization of inundation dynamics. Particularly, capturing the flood peak or maximum inundation extent, still remains elusive and a major research gap in the remote sensing of floods. Rapidly growing archives of multimodal satellite hydrology datasets combined with the recent deep learning revolution provide an opportunity to solve this problem adequate observation frequency. DeepFuse is a scalable data fusion methodology, leveraging deep learning (DL) and Earth Observation data, to estimate daily flood inundation at scale with a high spatial resolution. In this proof-of-concept study, the potential of Convolutional Neural Networks (CNN) to simulate flood inundation at the Sentinel-1 (S1) spatial resolution is demonstrated. Leveraging coarse resolution but temporally frequent datasets such as soil moisture/accumulated precipitation data from NASA’s SMAP/GPM missions and static topographical/land-use predictors, a CNN was trained on flood maps derived from S1 to predict high-resolution flood inundation. The proposed methodology was tested in southwest France at the confluence of its two main rivers, Adour and Luy, for the December 2019 flood event. The predicted high-resolution maps were independently evaluated against flood masks derived from Sentinel-2 using the Random Forest Classifier. First results confirm that the CNN can generalize some hydrological/hydraulic relationships leading to inundation based on the provided inputs, even for some rather complex topographies. However, further tests in catchments with strongly divergent land-use, hydrological, and elevation profiles is necessary to evaluate model sensitivity towards different land surface conditions. Achieving daily cadence for flood monitoring will enable an improved understanding of spatial inundation dynamics, as well as help develop better parametric hazard re/insurance products to effectively bridge the flood protection gap.