Data-driven model as a post-process for daily streamflow prediction in ungauged basins

Streamflow prediction in ungauged basins (PUB) remains a significant challenge in water resource planning and management. Although recent studies have proposed various approaches to reduce prediction errors using data-driven models (DDMs), further efforts are needed to improve applicability and accuracy of predictions in ungauged basins. This study proposes a framework that utilizes DDM as a post-processor to enhance the PUB performance of process-based models (PBMs) or DDMs and investigates its applicability. For this purpose, the Parsimonious EcoHydrologic Model (PEHM) was selected as a PBM, and Long Short-Term Memory (LSTM) and Random Forest (RF) were chosen as DDMs. We tested the proposed approach on 28 basins in Korea, which were assumed to be ungauged. First, PEHM and LSTM were used separately to predict streamflow in ungauged basins. Subsequently, RF was employed as the main DDM for post-processing, and the post-processing effects of LSTM were also examined. The results in this study demonstrate the potential value of various post-processing approaches in improving streamflow prediction in ungauged basins.