4.1.3 The proposed method (Methods 1-3)
Predicted runoff using the Method 1 are closer to the observed values, with a regression line slope of 0.90 and intercept of -0.30 (Table 3). The regression line of the Method 1 is closer to the 1:1 line, and the coefficient of determination R2 increases to 0.79 from 0.21 (the SCS-CN method). Moreover, the Method 1 performed satisfactorily as most values lie closer to the line of perfect prediction, even for the large runoff events, exhibiting a good match with the observed runoff (Fig. 3a). In the calibration case, the performance of the Method 1 was better than the above three methods because it yielded a higher NSE value of 75.84% and a lower RMSE value of 3.26 mm (Table 3). In the validation case, the value of NSE increases to 77.45% for the Method 1 from -182.60% for the original SCS-CN method, whereas the value of RMSE decreases to 3.72 mm from 13.18 mm. Thus we can conclude that the Method 1 performed best of the four methods for both calibration and validation datasets.
However, the points of Method 1 lie around the 1:1 line still scatter, which might be inherent to the standard SCS-CN method due to the adoption of the 0.2 value for the initial abstraction ratio λ. The assumption of λ = 0.2 has frequently been questioned for its pragmatic applications (Pronce and Hawkins, 1996; Baltas et al., 2007). Therefore, in order to improve the SCS-CN method, we introduced Method 2 with optimized λ based on the Method 1 to test the effect of initial abstraction on the proposed method. Moreover, we also optimized parameters in Eq.(6) (a1 anda2 ) (Method 3) based on Method 2 to test the slope factor on the proposed method as compared with the fixed values obtained from Huang et al.(2006).
A comparison of measured and estimated runoff depths by Methods 2 and 3, presented in Fig. 3b and Fig. 3c, shows that the runoff estimation is much improved over Method 1, with observation points close to the 1 : 1 line especially for runoff depths between 10 and 20 mm. The statistics in Table 3 confirm the predictive capacity of Methods 2 and 3 with higher model efficiency of 80.50% and 80.95%, respectively. However, when compared between Methods 3 and 2, it can be found that there is slightly improvement in Method 3 which using optimized slope parameter (a1 =213.99 and a2 =25.38) than Methods 2 with NSE values of 80.73% vs. 80.44% in calibration and 81.21% vs. 80.58% in validation. The results indicated that the proposed method with optimized λ (Methods 2-3) could further improved the SCS-CN method for runoff prediction in this study area.