| 31 Aug 2026
Multi-Model Comparison of Atmospheric River Flood Simulations in the Russian River Basin
Abstract. Extreme atmospheric river (AR) events are a dominant driver of flood risk in Northern California. Traditional flood studies and operational forecasts often rely on a single model; however, substantial structural differences persist across hydrologic modelling frameworks due to variations in process representations. Here, we present a multi-model comparison framework to evaluate differences in streamflow and flood inundation during an AR-driven flood event in the Russian River Basin (Sonoma County, California) in late February 2019. Three physical-based modelling systems, each forced with a common observation-based precipitation dataset, are compared: (1) Torrent, a computationally efficient overland flow model operating directly on digital elevation models (DEMs); (2) HEC-RAS, implemented using a two-dimensional shallow-water (Saint-Venant) solver in rain-on-grid mode; and (3) WRF-Hydro-HAND, a multi-scale hydrologic routing framework that combines distributed runoff generation with terrain-based inundation mapping. These systems differ fundamentally in their treatments of runoff generation, infiltration, and flow routing, enabling an explicit comparison of model behavior across frameworks. Model performance for channel water levels is evaluated against observations from U.S. Geological Survey stream gauges using the modified Kling–Gupta Efficiency (KGE′) and its components, including correlation, bias ratio, and variability ratio, together with diagnostics of peak timing. In addition, inundation extent is evaluated using the Landsat-derived Dynamic Surface Water Extent (DSWE) product, which provides spatially distributed observations of flood extent. Analysis of channel dynamics shows that variability and bias dominate model performance differences, while timing errors play a secondary role. Results indicate that all models capture the primary inundation along the mainstem Russian River, but substantial differences emerge in tributary and floodplain regions. Torrent and HEC-RAS tend to overestimate flood extent, while WRF-Hydro-HAND tends to underestimate it. These findings highlight the importance of model structure in flood simulation and demonstrate the value of multi-model frameworks for improving understanding of flood behavior in complex watershed systems.
This manuscript compares Torrent, HEC-RAS 2D Rain-on-Grid, and WRF-Hydro-HAND for the February 2019 atmospheric river flood in the Russian River basin. The models use AORC precipitation, and their water-level simulations are evaluated at 28 USGS gauges. Flood extent is also compared with one Landsat-derived Dynamic Surface Water Extent (DSWE) image. The topic is important. The manuscript also presents the model differences through maps, hydrographs, and some statistics. However, one of my concerns is that the current experiment does not isolate model structural uncertainty as claimed. The three systems receive the same precipitation product, but they use different loss/infiltration functions, initial states, spatial resolutions, parameters, and routing approaches. So it is difficult to determine whether the reported differences arise from model structure, setup, or scale. The paper is promising, but the main conclusions are stronger than the evidence supports, which need additional sensitivity tests, clearer uncertainty treatment, and more careful interpretation. I recommend major revision.
Also a careful language edit is needed. Correct some inconsistent spacing around hyphens, subject-verb agreement, singular and plural forms, and inconsistent use of modelled/modeled and modelling/modeling.
Major comments:
Specific comments:
Pages 2-3, Introduction. Define what is meant by structural uncertainty in this study and explain how a comparison of three complete model systems differs from a controlled experiment that changes one process representation at a time. End the introduction with specific research questions.
Page 3, Study Area and Event. It would be better if you could include a precipitation total map instead of just saying some locations receive more than 500 mm rainfall over 72h.
Pages 4-6, Section 3.1. Clarify the simulation start time, warm-up period, initial soil moisture, initial channel water, and downstream boundary condition for every model. These settings are essential for an event that followed earlier storms.
Page 5, lines 117-121. What’s the default value of Torrent's uniform Manning coefficient? How does rainfall losses are handled when infiltration, evapotranspiration, and subsurface processes are absent in Torrent?
Page 8, lines 206-221. Report the total DSWE comparison area, the percentage excluded by cloud or shadow, and the percentage classified as permanent water. Explain the resampling method used for each model and DSWE layer.
Page 9, lines 246-263. Why are 0.1 m inundation threshold chosen? Are TP, FP, FN, and TN pixel counts or proportions of all pixels?
Page 13, Figure 3. Check the color descriptions for the weakest KGE' components; the text and caption use different color names.
Page 14, lines 354-360. How much is the major flood threshold? Explain how the major flood threshold, which is usually based on gauge height, was converted to the same water-level reference used for each model. Exceeding a categorical threshold does not by itself demonstrate accurate magnitude or timing.
Page 15, Figure 4. Also provide a companion table with KGE', r, beta, gamma, peak bias, and timing error for the selected sites. Use 'mainstem' rather than 'mainstream.'
Page 22, Conclusion. Correct 'approache' around lines 515. Reduce the claim that HEC-RAS provides a balanced representation. Specify that it obtained the highest IoU for one post-peak DSWE scene while still overpredicting flooded area.