Abstract:The study aimed to realize the rapid nondestructive detection of the sugar content of Kaiyang loquat.The UV/Vis fiber-optic spectrometer was used to obtain the reflectance spectra of Kaiyang loquat the effects of standard normal variation（SNV）and multi-scatter calibration（MSC）on preprocessing the original spectra were explored and compared.The competitive adaptive reweighted sampling（CARS），successive projection algorithm（SPA）and combination of the two methods were employed to select characteristic variables separately.Then the back propagation（BP）neural network model was built up to detect the sugar content of Kaiyang loquat based on the characteristic variables.The results showed that SNV was a desirable spectral preprocessing method.A total of 49 characteristic variables were chosen by CARS from 835 full variables，which evidently improved the working efficiency of the model.Among the established models for predicting the sugar content of Kaiyang loquat，CARS-BP neural network model had the highest detection ability（rp=0.91，RMSEP=0.56%，RPD=2.42）.Therefore，it was possible to detect the sugar content of Kaiyang loquat by UV/Vis spectroscopy and BP neural network in a rapid and nondestructive way，which provided theoretical guidance for the research and development of online nondestructive detection equipment.