Coupled Physics-Based and Machine Learning Streamflow Modeling with High-Resolution Flood Inundation and Impact Assessment for Hurricane Florence (2018): A Case Study of the Neuse River at Goldsboro, North Carolina

H M Ashaduzzaman Asad 

Hurricane Florence made landfall in North Carolina in September 2018 and produced record or near-record river stages across the Neuse River basin, including at United States Geological Survey (USGS) gauge 02089000 near Goldsboro. This study presents an end-to-end hydrological and floodmapping workflow that couples a semi-distributed, physics-based rainfall-runoff model (Physics V3.5) and a gradient-boosted machine learning model (ML V4, based on Extreme Gradient Boosting) to predict discharge at the Goldsboro gauge, converts predicted and observed discharge to stage through an event-effective, hysteresis-aware stage-discharge (H-Q) surrogate, and produces spatially explicit flood-extent and flood-depth maps using a 1 m light detection and ranging (lidar) digital elevation model (DEM). The upstream contributing watershed (≈6,232 km2 ) was delineated using the USGS Network-Linked Data Index and the National Hydrography Dataset Plus High Resolution and discretized into 37 routed subcatchments. Physics V3.5 was developed using five historical calibration events (2015–2016) and three temporal-check events (2017), with Hurricane Florence withheld from parameter optimization; ML V4 was trained on 2015–2016 data and validated on the same 2017 events using only exogenous rainfall and routing features, with no observed-discharge lag or recursive predicted-discharge inputs. The stage-discharge surrogate uses isotonic (monotonic) regression fitted independently to each limb after robust aggregation of repeated stage values, and ML V4 draws on 40 engineered, purely exogenous rainfall-routing features selected from a ten-candidate hyperparameter grid using a composite skill score. Over the 14–23 September 2018 evaluation window, Physics V3.5 achieved a lower mean absolute error than ML V4 for both daily discharge (82.3 vs. 103.6 m3 s -1) and daily stage (1.03 vs. 2.01 ft), and both models reproduced the timing of the observed discharge peak on 18 September 2018; the retained Physics V3.5 artifact additionally reports an approximate whole-event Nash–Sutcliffe efficiency of 0.887 and Kling–Gupta efficiency of 0.831, although – as the artifact’s own provenance record makes explicit – Florence’s observed discharge contributed to that model’s parameter selection, so this figure is not an independent validation statistic in the way the ML V4 result is. The validated discharge series were converted to a spatial water-surface elevation profile along the Neuse River mainstem using a robust, iteratively re-weighted longitudinal regression, compared against the lidar DEM under a morphological hydraulic-connectivity constraint, and combined with WorldPop, National Land Cover Database, and OpenStreetMap layers – overlaid as fractional, sub-pixel coverage rather than a binary mask – to estimate population, land-cover, road, and facility exposure in the Goldsboro area of interest. The results illustrate both the value and the limitations of combining conceptual hydrology with data-driven discharge prediction for extreme, data-scarce flood events, and the importance of transparent gap-handling, extrapolation, and validation-independence rules in operational flood-mapping workflows.