Improving the SCS-CN method based on adjusting the main parameters using rainfall-runoff data

The Paskouhak representative basin in southwestern Iran, lies in elevations ranging from 2007 to 2931 m, and a semi-arid climate, the basin drains into Maharlou Lake. Rainfall–runoff records from 2011 to 2023, including training events (2011–2020) and validation events (2021–2023), were analyzed. This study evaluates and refines the USDA Soil Conservation Service Curve Number (SCS-CN) method by calibrating both the CN and initial abstraction coefficient (λ) using observed rainfall-runoff data. Tabulated CN values (USDA lookup) were compared against event-based CN estimates across three antecedent moisture classes. Four λ adjustment approaches- constant (λ= 0.2), discrete (mean/median), rainfall-dependent regression, and simultaneous model fitting- were applied. Model performance was assessed on the validation dataset using Nash-Sutcliffe Efficiency (NSE), root-mean-square error (RMSE), and mean absolute error (MAE). Calculated CN values were consistently lower than standard tables, reflecting local soil–land use conditions, and simultaneous model fitting (CN = 59.4; λ = 0.018) delivered the best performance (NSE = 0.78; RMSE = 1.68 mm; MAE = 1.30 mm) compared to conventional settings. A targeted Monte Carlo analysis of this optimal scenario confirmed that keeping parameter uncertainty within ± 5 % preserves high skill (mean NSE ≈ 0.76; 95 % CI (Confidence Interval): 0.70–0.78), whereas larger perturbations degrade accuracy. These findings underscore the importance of site-specific, joint calibration of CN and λ with tightly constrained uncertainty for dependable runoff prediction in semi-arid basins. • CN and λ calibrated using 2011–2023 rainfall–runoff data in a semi-arid basin. • Tabulated CNs significantly overestimate runoff compared to event-based CNs. • Four λ-adjustment methods tested; model-fitting yielded best accuracy. • NSE reached 0.78 with CN = 59.4 and λ = 0.018 in validation phase. • Monte Carlo uncertainty analysis confirmed model robustness under ±5–15 % variation.