Near infrared spectroscopy (NIR) was used to develop a method to rapidly determine the contents of polysaccharides and proteins in Lentinula edodes. The NIR spectra of 124 L. edodes samples were acquired. After outliers were identified and removed using the Mahalanobis distance, the samples were divided into calibration and prediction sets by the Kennard-Stone (KS) algorithm. The NIR spectra were preprocessed using different methods, and the optimal preprocessing method was selected. Based on the preprocessed spectra, characteristic wavelengths were selected using two different methods, competitive adaptive reweighted sampling (CARS) and variable combination population analysis-genetic algorithm (VCPA-GA). Partial least squares regression (PLSR), support vector regression (SVR), and crested porcupine optimizer-least squares support vector machine (CPO-LSSVM) were employed to establish six quantitative calibration models. The predictive performance of the developed models was comparatively evaluated. The results indicated that the optimal model for polysaccharide content prediction was Savitzky-Golay (SG) smoothing-multiplicative scatter correction (MSC) + CARS + CPO-LSSVM, yielding a prediction coefficient of determination (R_p²) of 0.9489, a root mean square error of prediction (RMSEP) of 0.0102 g/g, and a ratio of performance to deviation (RPD) of 4.4238. The optimal model for protein content prediction was standard normal variate (SNV) + CARS + SVR, achieving an R_p² of 0.9280, an RMSEP of 0.0125 g/g, and an RPD of 3.8056. No significant difference was observed between the measured values by conventional chemical methods and the NIR predicted values. These findings demonstrate that NIR is a feasible and effective technique for the rapid determination of polysaccharide and protein contents in L. edodes and can be applied for its quality evaluation.