Abstract: The Loess Plateau is one of the regions with the most severe water loss and soil erosion, which has the important ecological and economic significance for the prediction of the soil and water loss in the region. Integrating such factors as the climate, early soil moisture, soil type, and underlying surface, and featuring advantages of easy calculation, few parameters and easy access to the data needed, the SCS-CN (Soil Conservation Service Curve Number) has been extensively applied in the prediction of the surface runoff yield in the small watershed. Considering the specific climate and underlying surface of the Loess Plateau, the typical small watershed of Jiuyuangou, Suide County, Yulin, Shanxi Province was taken as the study region, then by virtue of rainfall runoff data in the Jiuyuangou area, the corresponding parameters (initial abstraction rate and rainfall intensity) affecting the runoff generation relation of the rainfall were optimized. The result showed that: ①The optimization of initial abstraction rate was conducted by use of the back-calculation method to confirm that λ was 0.075 and the coefficient of efficiency for the model was 0.208; ②The further optimization of initial abstraction rate was conducted by using MATLAB combined with the particle swarm optimization, and it was confirmed that λ was 0.13 and the coefficient of efficiency for the model was 0.504, increasing by 142% compared with the back-calculation method; and the prediction accuracy of this model was greatly improved; ③The factors of rainfall intensity were introduced in the loess hilly and gully region for the amendment of rainfall function, and after the improvement, the coefficient of efficiency for the model was 0.652 and the deterministic coefficient was 0.753. Compared with the standard SCS model, after the use of the modification function of rainfall intensity the deterministic coefficient R2 and the coefficient of efficiency of the model were increased by 101% and 534%, respectively. The comparison between the predicted and the measured runoff depths in the watershed shows that the simulated accuracy for the model is relatively ideal, and the model can be used for the prediction of the runoff generation of the rainfall in different small watersheds of the Loess Plateau.