Benchmark‐Guided Catchment Attribute‐Based Selection of Hydrological Models Across CAMELS ‐ US Catchments

Accurate simulation of streamflow in a catchment depends on selecting an appropriate model or model structure and equations that effectively represent the underlying hydrological processes. Further, model performance in a catchment is determined using performance metrics without explicit lower benchmarks. This study compares the effectiveness of conceptual lumped rainfall‐runoff (PB) and deep learning models compared to benchmark models across various catchment and climatic conditions for a benchmark‐informed, catchment‐attribute‐based model selection framework. Streamflow simulation was carried out using three PB models (GR4J, HBV, and SACSMA), as well as one deep learning model (LSTM). Model performance was assessed primarily using the Modified Kling–Gupta efficiency (KGE) metric and compared against catchment‐specific benchmark performance critical ranges derived from 17 benchmark models. In addition to modified KGE, four efficiency performance criteria such as high flow, low flow, water balance, and streamflow variability were evaluated to determine model robustness. A one‐dimensional analysis was conducted to explore how model efficiency varies with individual catchment characteristics. Additionally, a Classification and Regression Tree (CART) analysis was performed for each model to guide benchmark‐informed catchment‐specific model selection. The results suggest that PB models performed better in 70 catchments, achieving an average modified KGE of 0.82 in low aridity (< 0.7) and snow fraction (< 0.19), whereas the LSTM model was more effective in 82 catchments, achieving an average modified KGE of 0.68 in regions with a high snow fraction (> 0.22). The analysis offers significant understanding into the choice of models for streamflow simulation depending on catchment features/characteristics.