Exploring the capability of Noah-MP LSM in predicting fractional flooded area using U-Net architecture
David F. Drainer1, Aniket Gupta1, Ali Behrangi1, and Guo-Yue Niu1
1Department of Hydrology and Atmospheric Sciences
The University of Arizona, Tucson, AZ
Operational meteorologists and hydrologists commonly use physics-based models for flood inundation predictions. However, at regional scales, their coarse resolution limits their hydrological applications like flood inundation mapping (FIM). Previous research has relied upon various forms of satellite imagery to analyze flooded regions and create target maps that are used to train and validate deep learning models. Improvement to FIM has been shown through hybrid approaches to overcome limitations of satellite-based methods, problems such as satellite revisit times, and the presence of cloud cover obscuring the view of any surface flood water that may be present. This research introduces an approach to utilize continuous output from the Noah-MP land surface model (LSM) as the primary input to a U-Net architecture to predict flood water over the Sacramento Watershed. Noah-MP has been driven by a downscaled NLDAS-2 dataset at 1 km resolution using the WRF-Hydro Meteorological Forcing Engine. Four terrain inputs and the output variables from Noah-MP for soil moisture, snowmelt, surface-subsurface runoff, and ponding depth have been used to train the U-Net. Target flood maps are from the RAPID NRT flood inundation archive and have been adjusted to show fractional water coverage at 1 km to match the resolution of Noah-MP. K-fold cross-validation was utilized on 4,780 samples created from 31 flood events from 2017-2019 with 5 folds, holding out 15% for testing. Results show that this approach effectively identifies surface water versus dry land with an RMSE of 0.04 and a bias of -0.09.