Soil available phosphorus (P) and available potassium (K) don’t possess direct spectral response in the near infrared (NIR) region. They are predictable because of their correlation with spectrally active constituents (organic matter, carbonates, clays, water, etc. ). Such correlation may of course differ between the soil sample sets. Therefore, the NIR calibration models with fixed structure are difficult to achieve good prediction performances for soil P and K. In this work, the method of recursive partial least squares (RPLS), which is able to update the model coefficients recursively during the prediction process, has been applied to improve the predictive abilities of calibration models. This work compared the performance of partial least squares regression (PLS), locally weighted PLS (LW-PLS), moving window LW-PLS (LW-PLS2) and RPLS for the measurement of soil P and K. The entire data set of 194 soil samples was split into calibration set and prediction set based on soil types. The calibration set was composed of 120 Anthrosols samples, while the prediction set included 29 Ferralsols samples, 23 Anthrosols samples and 22 Primarosols samples. The best prediction results were obtained by the RPLS model. The coefficient of determination (R2) and residual prediction deviation (RPD) were respectively 0.61, 0.76 and 1.60, 2.05 for soil P and K. The results indicate that RPLS is able to learn the information from the latest modeling sample by recursively updating the model coefficients. The proposed method RPLS has the advantages of wider applicability and better performance for NIR prediction of soil P and K compared with other methods in this work.