In the present work, recursive variable selection methods (updating both the model coefficients and effective variables during the prediction process) were applied to maintain the predictive abilities of calibration models. This work compared the performances of partial least squares (PLS), recursive PLS (RPLS) and three recursive variable selection methods, namely variable importance in the projection combined with RPLS (VIP-RPLS), VIP-PLS, and uninformative variable elimination combined with PLS (UVE-PLS) for the measurement of soil total nitrogen (TN) and organic matter (OM) using Vis-NIR spectroscopy. The dataset consisted of 195 soil samples collected from eight towns in Wencheng County, Zhejiang Province, China. The entire data set was split randomly into calibration set and prediction set. The calibration set was composed of 120 samples, while the prediction set included 75 samples. The best prediction results were obtained by the VIP-RPLS model. The coefficient of determination (R2) and residual prediction deviation (RPD) were respectively 0.85, 0.86 and 2.6%, 2.7% for soil TN and OM. The results indicate that VIP-RPLS is able to capture the effective information from the latest modeling sample by recursively updating the effective variables. The proposed method VIP-RPLS has the advantages of better performance for Vis-NIR prediction of soil N and OM compared with other methods in this work.