The sugar content of kiwifruit is an important measure of its internal quality. Traditional sugar content detection is time-consuming and destructive sampling, and it is of great significance to non-destructive detect the sugar content of kiwifruit effectively for its quality classification, storage and sales. The common non-destructive detection methods of fruit and vegetable quality based on hyperspectral imaging technology mostly use a single algorithm of competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), principal component analysis (PCA) and iteratively retains informative variables (IRIV) to extract features. However, using these algorithms alone will lead to insufficient stability of prediction results. This study designs a non-destructive detection method for kiwifruit sugar content based on hyperspectral imaging technology. The “Red Sun” kiwifruit samples in Ya'an city of Sichuan province were numbered, their hyperspectral images in the wavelength range of 400~1 000 nm were collected, and the average spectrum of the region of interest was calculated as the effective spectral information of the samples. Then, three spectral data preprocessing methods including Multiplicative Scatter Correction (MSC), Standard Normal Variate (SNV), and Direct Orthogonal Signal Correction (DOSC), were used to analyze the influence on the accuracy of the prediction models, respectively. The comparison results showed that DOSC had the best preprocess effect. Further, for the preprocessed spectrum, 7 dimensionality reduction methods including CARS, SPA and IRIV from one-time dimensional-reduction algorithms, CARS+SPA and CARS+IRIV from the first-order combined dimensional reduction algorithms, and (CARS+SPA)-SPA, (CARS+IRIV)-SPA from the second-order combined dimensional-reduction algorithms respectively, were used to extract characteristic spectral variables, and three models for predicting the sugar content of kiwifruit were constructed i. e. Support Vector Regression (SVR), Least Square Support Vector Machine (LSSVM) and Extreme Learning Machine (ELM) models. Finally, the prediction accuracy of the three models based on different feature extraction methods was compared through experiments. This study shows that the ELM model has the best prediction performance, while the SVR model has the worst prediction performance. When the characteristic spectral variables extracted by (CARS+IRIV)-SPA were input into LSSVM and ELM models, respectively, the prediction results are better than those obtained by other methods. Then (CARS+IRIV)-SPA is verified to be effective in improving the prediction accuracy of the models. Comparing the prediction results of these methods, the prediction performance of (CARS+IRIV)-SPA-ELM is better than other methods, with the correlation coefficient R_C=0.945 1, R_P=0.839 0, RMSEC=0.450 3, RMSEP=0.598 3, and RPD=2.535 1, which will provide reliable theoretical basis and technical support for the non-destructive, precise and intelligent development of kiwifruit sugar content detection.