Objective: In order to eliminate a large amount of redundant information in near infrared spectrum, and improve the accuracy of sugar degree in apple of the prediction model, a method for fast and non-destructive testing of the sugar content of apples was established. Methods: A prediction model of apple sugar content based on wavelet packet transform and bottle ascidian algorithm was proposed. Firstly, according to the characteristics of apple spectral data with high and complex dimensions, the spectral data was reduced, and the characteristic wavelength screening method was determined by comparing the results of full band and partial least squares method, continuous projection method and wavelet packet transform. Secondly, in view of the influence of extreme learning machine (ELM), the model performance was affected by its initial weight and hidden layer bias selection. The bottle salp swarm algorithm was used to optimize the initial weight and hidden layer bias of ELM model, and an prediction model for sugar content in apple was proposed based on bottle ascidian algorithm improved extreme learning machine. Results: Compared with GA-ELM, PSO-ELM and ELM, the prediction model based on SSA-ELM had the highest accuracy. Conclusion: The parameters of ELM model optimized by intelligent algorithm can effectively improve the prediction accuracy of ELM, and provide a new method for the prediction of sugar content in apple.