Greengage acidity detection is very important in refining and deeply processing greengage. However, traditional greengage acidity detection methods based on physicochemical analysis are destructive, time-consuming and not detective online. The fast and non-destructive method based on hyperspectral imaging system was proposed to predict greengage acidity. Hyperspectral images of 487 greengage specimens between wavelengths of 550nm and 1000nm were captured. Three spectral dimensional reduction methods such as successive projection algorithm (SPA), genetic algorithm (GA) and SPA combined with GA (SPA+GA) were explored after spectrum relative reflectivity was calibrated and the images were filtered in six different ways. The featured wavelengths of the spectrum were extracted which reflected the internal acidity information of greengage. Partial least squares (PLS) prediction model was built between wavelength, and pH value and prediction precision were compared among different methods of filters and dimensionality reductions. The results showed that the model smoothly filtered by Savitzky-Golay (S-G) had the highest prediction accuracy. The model smoothly filtered by five points and then dimensionally reduced by both SPA and GA can reduce its complexity and improve its prediction accuracy compared with the ones only using SPA or GA. The root mean square error of prediction set was 0.0706, and the correlation coefficient of prediction set was 0.7925. This model based on the selected wavelength was practical to predict the greengage acidity, which would lay the foundation for further developing actual greengage multispectral image system.