【Objective】Soil is the important carrier of agricultural production, while the potassium in soil is one of the nutrient elements essential for plant growth, so it is very important to quickly and accurately assess soil potassium content in farmland. Conventional soil potassium content determination methods are expensive and time-consuming. The visible and near infrared reflectance spectroscopy (VIS–NIR), which possess the advantages of non-destructive and rapid detection, has been a useful tool for quantitative analysis of soils of numerous attributes. The object of this study is to investigate feasibility to use the visible and near infrared reflectance spectroscopy in estimating soil potassium contents in the Weibei Rainfed Highland. 【Method】A total of 120 loessal soil samples were collected from the farmfields in Qian County of Shaanxi Province for analysis of total potassium (TK) and readily available potassium (AK) contents in lab with conventional chemical methods. Reflectance spectroscopic data of the soil samples were acquired with the SVC HR–1024i spectroradiometer. Three types of pretreatments, including First-order differential of soil spectral reflectance (DSSR), band depth (BD) and First-order differential of band depth (DBD), were adopted to amplify the weak absorption characteristics, eliminate noises in the system and external disturbances. The continuum-removal method was used to extracted band-depths (BD) of the soil reflectance spectra and based on correlation analysis a model was built up for prediction of TK and AK contents in the loessal soil, using the multiple linear regression (MLR) and partial least squares regression (PLSR) methods and validated with independent samples. 【Result】Results show that the multiple linear regression model based on DSSR as independent variable could accurately estimate TK contents while the other multivariate linear regression models could not do TK and AK contents so accurately. Comparison shows that the PLSR models were generally higher than the MLR models in prediction accuracy. The models built up with the PLSR method incorporating the 4 spectral variables, could estimate TK content accurately, specially the PLSR model based on DBD as independent variable, of which coefficient of determination (R²_pre) of the validation set was > 0.90, root mean square error of the prediction (RMSE_pre) was 0.054, and residual predictive deviation (RPD_pre) was 3.310. The PLSR model based on BD followed. However, the PLSR models based on the 4 spectral variables were relatively low in calibration and validation accuracy and only some models could roughly predict AK contents. The PLSR model based on DSSR as independent variable, with RPD_prebeing 1.619 and 1.572 was the best for accurate prediction of soil AK contents. 【Conclusion】All the finding in this study demonstrate that the visible and near infrared spectroscopy coupled with MLR and PLSR can be used to predict rapidly and accurately TK content in the loessal soil of the Weibei Rainfed Highland, but only roughly AK contents therein.