Improved National-Scale Flood Prediction for Gauged and Ungauged Basins using a Spatio-temporal Hierarchical Model

Abstract

Floods cause hundreds of fatalities and billions of dollars of economic loss each year in the United States. To mitigate these damages, accurate flood prediction is needed for issuing early warnings to the public. This situation is exacerbated in larger model domains for high flows, particularly in ungauged basins. To improve flood prediction for both gauged and ungauged basins, we propose a spatio-temporal hierarchical model (STHM) to improve high flow estimation using a 10-day window of modeled National Water Model (NWM) streamflow and a variety of catchment characteristics as input. The STHM is calibrated (1993-2008) and validated (2009-2018) in controlled, natural, and coastal basins over three broad groups, and shows significant improvement for the first two basin types. A seasonal analysis shows the most influential predictors are the previous 3-day average streamflow and the aridity index for controlled and natural basins, respectively. To evaluate the STHM in improving streamflow in ungauged basins, 20-fold cross-validation is performed by leaving 5% of sites. Results show that the STHM increases predictive skill in over 50% of sites by 0.1 Nash-Sutcliffe efficiency (NSE) and improves over 65% of sites’ streamflow prediction to an NSE>0.67, which demonstrates that the STHM is one of the first of its kind and could be employed for flood prediction in both gauged and ungauged basins.

Leveraging Synthetic Aperture Radar (SAR) to improve above-normal flow prediction in ungauged basins

“…The correlation between SAR-derived and observed discharge varies across the CONUS is important to understand when looking to use SAR-derived products as a proxy for flow prediction (Figure 2), but Figure 2 shows improvement in estimating 3-day streamflow using the SAR-derived streamflow when compared with the 3-day streamflow estimated using the simple depth of streamflow. The correlation between SAR-derived streamflow and observed above average streamflow is notably strong, as depicted in The analysis conducted in Fang et al (2024), comparing the R-squared values between GFDS SAR-derived streamflow and the 3-day average streamflow estimated from the simple depth method, demonstrates that GFDS-SAR derived streamflow consistently surpasses the performance of the simple depth approach in explaining the variability observed in streamflow (see Figure 3a). This suggests that GFDS-SAR derived streamflow better captures the underlying variability in streamflow dynamics compared to the traditional 3-day flow approach.…”

“…Figure 3a also compares the NSE from the GFDS-SAR with the NSE (X-axis) between the 3-day average streamflow estimated based on the simple depth of streamflow (i.e., without using the GFDS-SAR stage estimates) with the observed streamflow. We then conducted an analysis of model performance using SAR-derived data (Figure 4) and present an examination of contributing factors in the STHM-SAR compared to the base STHM (Fang et al, 2024).…”

“…Moreover, Figure 3b shows that the improved performance of SAR derived streamflow decreased as the number of gauges increases in the same HUC08. The relationship underscores that these performance enhancements are notably more significant in basins with limited gauged locations, where 29% of these basins with ≤ 2 gauges often located within coastal basins (<150km along coastal line, Fang et al 2024).…”

“…Since we are interested in assessing the performance of the model for estimating flows in ungauged locations, we use both spatial and temporal validation procedures similar to that of Fang et al, (2024). For spatial validation, we used a k(20)-fold cross-validation method (Browne, 2000) treating 5% of locations as ungauged within each hierarchical group and fit the remaining 95% of stations for the period 1993 and 2018.…”

“…Postprocessing model outputs has been shown to enhance the forecast skill, however, NSEs of forecasts for ANF conditions near gauged locations are only skilled in 50% of evaluated basins (Fang et al, 2024;Frame et al, 2021;Johnson et al, 2023). The ability to predict abovenormal flows is likely worse in ungauged basins highlighting the need for improved ANF prediction and the opportunity for utilizing using the unprecedented availability of NWM forecasts as a starting point.…”