The separation of coal and gangue is a crucial step in coal mining, but the existing methods such as manual selection and mechanical separation are ineffective and environmentally hazardous. This study aimed to explore the feasibility of the accurate classification of coal and gangue with black background based on the visible and near-infrared hyperspectral imaging technology, simplify classification models using feature selection methods, and provide a reference for constructing a multispectral system for coal and gangue separation. Hyperspectral imaging technology is a fast and non-destructive detection method without sample pretreatment and environmental contamination. Firstly, a hyperspectral imaging system was developed to collect hyperspectral data of 85 coal samples and 83 gangue samples in the range of 400~1 000 nm (Vis/NIR) and 1 000~2 500 nm (NIR) from the XiMing mine. After removing background information of hyperspectral images, the average spectra in the randomly selected regions of 100 pixel×100 pixels in 400~1 000 nm and 50 pixel×50 pixels in 1 000~2 500 nm were extracted. After repeating 10 times, 850 coal spectra and 830 gangue spectra were obtained in each of the two bands. Savitzky-Golay smoothing and standard normal variate transformation were performed successively to reduce the impact of errors and noise on the spectra. Three models, including support vector machine (SVM), k-nearest neighbor (KNN), partial least squares discrimination analysis (PLS-DA), were established based on full-band spectra. The classification accuracy rate of each model for the prediction set was greater than 0.95, which revealed that coal and gangue could be distinguished by spectral information. Subsequently, competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) were employed to select characteristic wavelengths to simplify models. Considering factors such as accuracy and cost, the SVM model based on the 3 characteristic wavelengths screened by SPA in the Vis/NIR range had the best performance, that not only effectively reduced the number of wavelengths, but also improved the classification capacity and the corresponding sensitivity, specificity, accuracy was: 1.000 0, 0.965 2, 0.983 3, respectively. Based on the discriminant model and the average spectra of the samples, the classification and visualization of coal and gangue can also be realized. The research results have great potential for developing a low-cost and multi-spectral separation system for coal and gangue based on the characteristic wavelengths to achieve fast and accurate non-destructive separation.