Yellow peaches are soft and prone to damage，and the different level of damage can directly affect the end use and sale price of yellow peaches. The reflection （R），absorption （A），and Kubelka-Munk spectra （K-M） of yellow peaches were obtained by using hyperspectral techniques and used to detect the early damage level of yellow peaches. Partial least squares discriminant analysis （PLS-DA），extreme gradient boosting （XGBoost） and random forest （RF） models based on three raw spectra and various pretreated spectra were established. The results were compared to select the model with higher correctness. The model with its characteristic wavelength was constructed and compared again. The results showed that RF models based on the three raw spectra and SG pretreated spectra were superior in discriminating，with the overall accuracy rates all above 90.00%. The wavelength screening of the raw spectra and SG pretreated spectra was performed with the competitive adaptive reweighting （CARS） and uninformative variable elimination （UVE） algorithms，and the RF models were established again. The results showed that the A-RAW-CARS-RF model and the K-M-SG-CARS-RF model were improved in discriminating compared with the RF model at full spectrum. Among the RF models established based on the characteristic wavelengths，the A-RAW-CARS-RF model had the best discriminating effect with an overall accuracy of 97.12%. The number of misclassifications for the four subcategories were 0，1，1，and 1. It is indicated that the feasibility of detecting the early damage level of yellow peaches based on absorption spectroscopy （A）. It will provide some theoretical basis for detecting fruit bruise with hyperspectral techniques in the future.