Visible-near infrared (VNIR) diffuse reflectance spectroscopy has low accuracy in estimating soil phosphorus (P) and potassium (K). We used database of 1582 soil samples from 8 soils to investigate P and K content. All samples were oven dried, ground, and sieved with a 2mm screen. Each sample was divided into two subsamples. One subsample was tested by chemical method. Another subsample was scanned by an ASD FieldSpec Pro FR spectrometer. Data were collected using FieldSpec RS3 software. All spectra were recorded between 350nm and 2500nm and output at a1nm interval. Each soil sample was scanned 3 times with the sample cup rotated within the sample holder to angles of 0°, 45° and 90°. The three spectra of each sample were averaged. Spectral data at the lower visible wavelengths were removed due to their low signaltonoise ratio; and the last 50nm at the high near infrared wavelengths were also deleted for the same reason. Then spectra from 401nm to 2450nm with 1nm interval were reduced by averaging five successive wavelengths. 〖JP2〗Thus, the number of spectral variables was 410. Pretreatments of log10(1/reflectance) plus mean normalization plus median filter smoothing with or without direct orthogonal signal correction (DOSC) were investigated. Results from partial least squares regression (PLSR) with leaveoneout crossvalidation were: the root mean square error of prediction (RMSEP), the determination coefficient of prediction (R2) and the ratio of standard deviation to RMSEP (RPD) were respectively 27.343mg/kg, 0.309, 1202 for P， and 70975mg/kg, 0.421, 1313 for K when DOSC was not used. The value of RMSEP, R2 and RPD were respectively 21.464mg/kg, 0.574, 1531 for P，and 53.485mg/kg, 0.671, 1743 for K when DOSC was used. Additionally, calibrations using only those samples within the approximate range of interest for fertilizer application to field crops (P from 0 to 27mg/kg and K from 0 to 192mg/kg) were investigated. Value of RMSEP of calibration models by PLSR with DOSC decreased by 26.93%(P(0,27)) and 27.67%(K(0,192)), but R2 and RPD increased respectively by 108.31%, 36.90%(P(0,27)) and 87.01%, 38.29%(K(0,192))comparing with models by PLSR without DOSC. The results of this research showed DOSC algorithm can eliminate spectral signal noise might be caused by soil type, texture, etc. for estimating soil P and K using VNIR. DOSC might be a good pretreatment method of spectra for testing soils P and K by VNIR.