Flood prediction in an ungauged watershed: A case study of the naengcheon watershed, Korea

Over the decades, a variety of physically-based, data-driven, and hybrid flood prediction models have been developed, demonstrating high performance by leveraging extensive datasets, including topography, surface roughness, and hydrological data. Despite advancements, the use of these models remains largely limited to gauged watersheds due to their reliance on measured discharge data for accurately modeling infiltration processes, and the use of consistent infiltration parameters further restricts prediction accuracy. To overcome this limitation, this study introduces a straightforward technique that utilizes parameters from an adjacent watershed with comprehensive data collection to account for rainfall loss due to infiltration, employing event-by-event infiltration parameters to enhance prediction accuracy. By integrating diverse watershed data, including rainfall-runoff and GLDAS-Noah data, with the Long Short-Term Memory (LSTM) network—a high-performance deep learning model widely used in hydrology—we achieved reliable parameter estimation in adjacent gauged watersheds. By transferring these estimated parameter values to the target ungauged watershed, leveraging the spatial proximity between adjacent watersheds, we effectively estimated rainfall loss and simulated a physically-based model for streamflow prediction, even in an ungauged watershed. For a gauged watershed, we achieved an accuracy with a Nash-Sutcliffe Efficiency (NSE) of 0.7–0.97 and a coefficient of determination (R²) of 0.7–0.98 in predicting flood hydrographs for most events. For an ungauged watershed, where observed runoff data do not exist, we employed indirect verification methods using alternative data sources such as news reports and flood inundation maps. We first generated flood maps based on our channel routing results and then compared them with a flood inundation map from a news report. In this case, we quantified the accuracy using the confusion matrix, and the metrics were measured as Accuracy = 0.72 and Recall = 0.58. This technique offers significant potential by integrating watershed data with deep learning models like LSTM to enhance flood prediction in data-scarce regions, making it applicable for ungauged watersheds through spatial proximity. By improving disaster preparedness and management with more reliable flood forecasting and efficient resource allocation, it helps mitigate the impact of floods and inspires further advancements in flood prediction models.