Visible and near-infrared reflectance spectroscopy (Vis-NIRS) was applied to non-invasively measurement of water content in engine lubricant. Based on measured spectra, several spectral calibration algorithms were adopted to improve accuracy and simply calculation. Principal component analysis (PCA) and successive projections algorithm (SPA) were separately used to reduce variables of spectral model. Nine effective variables, 476, 483, 544, 925, 933, 938, 952, 970 and 974 nm, were selected by SPA, and were inputted into partial least square regression (PLSR) and multivariable linear regression (MLR) models. Both the two models obtained better results than full-spectra-PLSR model and PCA-PLSR model. It shows that SPA does not select uninformative but effective variables from full-spectrum. Least-square support vector machine (LS-SVM) was operated to improve Vis-NIRS's ability based on full-spectrum and SPA, separately. High coefficients of determination for prediction set (Rp(2)) up to 0.9 were obtained by both full-spectrum-LS-SVM and SPA-LS-SVM models. SPA-LS-SVM is better than full-spectrum-LS-SVM. The value of Rp(2) of SPA-LS-SVM is 0.983 and residual predictive deviation (RPD) is 6.963. It is concluded that Vis-NIRS can be used in the non-invasive measurement of water content in engine lubricant, and SPA is a feasible and efficient algorithm for the spectral variable selection.