Benchmarking data-driven rainfall-runoff modeling across 54 catchments in the Yellow River Basin: Overfitting, calibration length, dry frequency

Study region: Yellow River Basin, China. Study focus: The rainfall-runoff modeling performance of Data-Driven Models (DDM) in the Yellow River Basin (YRB) at a large scale is unclear such that the DDM research in the YRB lacks essential reference which can be critical for model development research. Understanding the advantages and disadvantages of DDMs by comparing them to Process Based Models (PBM) helps model selection in practice, especially when benchmarking is performed at large scales. We benchmarked three DDMs, namely SVM, LSTM and CNN-LSTM, and a PBM, namely the Xinanjiang (XAJ) model, across 54 basins of the YRB. Factors affecting DDM performance are identified and the sensitivity of PBM to these factors is also discussed. New hydrological insights for the region: DDM performs the best in the upper reaches, the worst in the middle reaches. PBM demonstrates a wider applicability when calibration data is limited, whereas DDM generally outperforms PBM for areas where data limitation is not a problem. The most important catchment attribute affecting PBM and DDM is a high frequency of dry days (<1 mm d⁻¹). However, DDM is more vulnerable to this factor. In addition, DDM performance depends heavily on the introduced lagged streamflow when data is insufficient. We conclude that the rainfall-runoff modeling relationship in catchments with high drought frequencies is more complex, resulting in DDM requiring more data, but PBM is less affected by these factors, indicating that PBM has better applicability in the case of limited data.