Leveraging satellite soil moisture observations for deep learning methods for flood inundation mapping

Various studies have been done to apply state-of-the-art deep learning techniques such as UNET, Encoder-Decoder LSTM, or stacked autoencoder recurrent neural network (SAE- RNN) for flood extent and depth mapping. Current methods, while found to be successful for their specific study areas and flood events, still have several limitations in generalization for unseen case studies and the lack of dynamics features/variables usage. Indeed, the most common features/input data for current deep learning flood mapping studies are static variables (e.g., slope, elevation data, land use). Soil moisture is one dynamic and readily available variable that has not been utilized for input variables for flood mapping deep learning models. As soil moisture was found to show distinctly anomalous behavior prior to flooding and could provide information on soil saturation and future runoff, satellite soil moisture observations prior to a flood event can greatly benefit flood modelling and forecasting. In this study, we will evaluate the potential of downscaled satellite soil moisture as an input for current deep learning models for flood extent mapping and depth estimation.