Machine Learning Evaluation and Enhancement of a Low-Complexity Flood Inundation Model using Reach-Specific Geomorphic Parameters

Climate change and human activity place increasing stress on river networks that can lead to negative outcomes such as increased erosion and transport of sediment and nutrients. Low complexity models such as those based on Height Above Nearest Drainage (HAND), are useful to planners because they have relatively low computational costs and require inputs that are easier to acquire such as remotely acquired topographic data. This enables modeling of larger-scale drainage networks with less execution time than physics-based models (e.g., 2D HEC-RAS). Monte Carlo analysis has recently been incorporated into the HAND synthetic rating curve approach to improve probabilistic estimates of inundation, in a framework called probHAND. HAND-based approaches often underrepresent channel bathymetry because of their reliance on lidar-derived topography, and do not well represent inundation heights in regions where channels and floodplains have been substantially modified. We evaluate how the integration of geomorphic data into the HAND map and probHAND model influences the depth and extent of inundation and channel specific stream power at the reach scale by comparing reach-averaged outputs to calibrated 2D HEC-RAS models of the Mad River and Black Creek in Vermont. We leverage data from Vermonts Stream Geomorphic Assessment program, which provides estimates of channel width, valley confinement, entrenchment, and channel incision at the scale of geomorphically-consistent reaches. We use a new unsupervised clustering method designed for time series data, SOMTimeS, to cluster reaches based on patterns exhibited by probHAND data outputs as depth of inundation is varied. We show that SOMTimeS clusters are correlated with sediment regime types generated from the SGA field measurements. Finally, we incorporate the SGA data to improve probHAND predictions for reach types where probHAND predictions deviate more from reference conditions predicted by 2D HEC-RAS models. By identifying the geomorphic conditions where probHAND performs relatively poorly, we characterize uncertainty and give planners a tool to evaluate the degree to which they can rely upon probHAND results. Using relatively easy to obtain inputs we enhance the predictive ability of probHAND to a greater variety of geomorphic conditions.