PLSR regression with cross validation(LOOCV) was used to establish the optimal model using the dataset of a long-term plot experiment in Suzhou and dataset of Zhihugang river watershed aimed to discuss the influence of different pretreatment methods on prediction of total phosphorus(TP) in paddy soil and explor the feasibility of total phosphorus prediction with hyperspectrum. Important spectrum bands for total phosphorus prediction were also explored with correlation analysis and PLSR regression coefficient analysis. Results showed that effect of pretreatment methods of hyperspectral data on the precision of PLSR model was little. The precision of optimum PLSR model was good with the coefficient of determination (R2) of 0.85 for calibration and 0.70 for validation, and RPD of 1.8. 440–740 nm was important for TP prediction with hyperspectrum. It is feasible to predict soil TP using visible-near infrared reflectance spectroscopy with PLSR regression method.