Near infrared (NIR) spectroscopy was investigated to determine the protein content of Auricularia auricula (commonly called black woody ear or tree ear) using partial least-squares (PLS), multiple linear regression (MLR), and least-squares-support vector machine (LS-SVM). The performances of different preprocessing were compared including Savitzky-Golay (SG) smoothing, standard normal variate, multiplicative scatter correction (MSC), first derivative, second derivative, and direct orthogonal signal correction. A successive projections algorithm (SPA) was applied for relevant effective wavelengths selection. The combinations of various pretreatment and calibration methods were compared based on the prediction performance. The optimal full-spectrum PLS model was achieved by raw spectra, whereas the optimal SPA-MLR, SPA-PLS, and SPA-LS-SVM models were achieved by MSC spectra. The best prediction performance was achieved by the SPA-LS-SVM model, with correlation coefficients (r) = 0.9839 and a root mean squares error of prediction (RMSEP) = 0.16. The results indicated that NIR spectroscopy combined with SPA-LS-SVM was the most successful to determine the protein content of A. auricula.