The Soil and Water Assessment Tool (SWAT) model is diffusely applied to areas with uniform precipitation distribution and gentle slopes. However, the SWAT has apparent defects in the process of ecohydrological simulation in mid-high latitudes and arid regions. This model can merely be used in agricultural catchments with gentle slopes and uniform rainfall, and the simulation is not suitable for high-altitude alpine catchment. To this point, this paper makes appropriate improvements to the runoff module, and specifically corrects the key sensitive parameters. The global sensitivity and performance assessment of the model's parameters were carried out before and after the improvement using the monthly data of runoff and evaporation from 1990 to 2018 in the Kaidu River basin in northwestern China. It is indicated that the Modified SWAT simulates the performance of monthly runoff and evapotranspiration better than the original SWAT, which overestimates the runoff volume throughout hydrological changes in different periods. Compared with the monthly water production under different land use/cover conditions, the highest water production was found in forest and grassland. By modifying the SWAT algorithm, the structural non-determinacy of the Modified SWAT is significantly reduced, and the adaptability of the model is improved. Thus, the Modified SWAT shows a better simulation performance in arid areas with considerable terrain variation and scarce rainfall data.