To develop a model for rapid, accurate grading of lumbers based on knots, near infrared spectroscopy was used to detect knots on coniferous wood surface. We explored the effects of spectral preprocess methods and modelling methods on knot detection, and investigated the feasibility of using a model built within one species to discriminate the samples from other species. Successive projections algorithm (SPA) was used to select effective wavelengths. The results showed that least squares-support vector machines (LS-SVM) coupled with first derivative preprocessed spectra achieved the best performance for both single and mixed models. Fifteen effective wavelengths, only 0.87% of the full wavelengths, were selected by SPA to build an LS-SVM model, and the sensitivity, specificity and accuracy in validation set were 0.990, 0.954, 97.44%. The results showed that near infrared spectroscopy combined with SPA and LS-SVM can be used to detect surface knots on different coniferous wood varieties. SPA is a powerful tool to select the efficient variables, and it can simplify model and improve model prediction precision.