The long year’s siltation of Tsengwen Reservoir affects the normal water supply severely. This study is aimed to evaluate soil erosions and sediments at corresponding watersheds, in which the management can be easily accessed. Regression analysis was conducted to predict the soil erosions where the hydrological and physiographic factors were considered for a specific rainfall site and/or the cases of continuous raining. Additionally with the presumed slope dimensions and the application of empirical formula and physical mode (WEPP) to predict soil changes along the slope, the probabilistic model of soil erosion can be established. Furthermore, GeoWEPP program was conducted to simulate climate change. With land reclamation, digital topographic map and soil files, the amount of sand transports at the watersheds can be predicted based on the information of rainfall and transports of soil formations. Both slope and watershed levels’ assessment can be suggested. 1. Soil erosions: (1) Empirical formulas are conservative compared to the physical mode WEPP. There difference is not very significant. When using the regression formulas, one standard deviations of plus and/or minus are suggested. When using physical model, the dimensions of slope is better to be large in order to provide suitable predictions.. (2) The probabilistic soil erosion model can be used to provide soil erosion in correspondence with the annual exceednace of rainfall. It can be used to predict the sediment yield at the reservoir inflow station. It is suggested that the relationships of the return period and the soil erosion depth is two years-0.34cm、five years 0.37cm、ten years-0.39cm、twenty years-0.41cm and one hundred years-0.44cm. 2. Sand transports at watercourse: (1) Using the sand prediction program GeoWEPP to predict the sediment yields by typhoon Nanmadol in 2011 and Typhoon Tembin in 2012, with the flowpath model which predicts each watershed, river, natural and protected slopes and total amount of soil lost, as well as proper source area and channel length, an error of 6.7% can be found in the prediction. (2) Use sand sediment balance system, the soil loss of the reservoir watershed at the inflow stations can be computed by multiplying the soil loss of the slope with the sediment delivery ratio.