Sugar content is a crucial parameter for assessing watermelon quality, influencing watermelons marketability and commercial value. However, the natural biological characteristics of large volume and thick skin pose challenges for rapid and non-destructive evaluation of the sugar content of watermelon. In this study, 230 watermelons were selected for investigation. A custom-designed full-transmission visible-near-infrared detection system was developed. Spectral data of all samples were acquired online. Each sample spectral data comes from the equatorial part of the watermelon. The overall watermelon sugar content and the central sugar content were measured separately to provide reference values for the assessment of sugar content. In the data processing phase, the spectral data of each sample was averaged, and spectral data in the 690～1 100 nm was selected. The Monte Carlo method was implemented to remove abnormal samples, and preprocessing, such as Standard Normal Variate correction and Savitzky-Golay smoothing, was applied to optimize the spectral data. The SPXY algorithm was used to divide the calibration and prediction sets. Utilizing the optimized spectral data, linear Partial Least Squares Regression (PLSR) and non-linear Least Squares Support Vector Machine (LS-SVM) models were developed to forecast each samples center sugar content and overall sugar content. The results revealed that, Combined with standard normal variate correction and Savitzky-golay smoothing, the LS-SVM model yielded the most favorable results in predicting the overall watermelon sugar content. The calibration correlation coefficient (RC) of 0.92 and root mean square error of calibration (RMSEC) of 0.37°Brix were obtained for the calibration set. Correspondingly, the prediction correlation coefficient (RP) of 0.88 and root mean square error of prediction (RMSEP) of 0.40°Brix were obtained for the prediction set. Furthermore, feature wavelength selection algorithms (e.g., Competitive Adaptive Reweighted Sampling, Uninformative Variable Elimination, Successive Projections Algorithm) were used for variable selection. Study found that the LS-SVM model combined with Competitive Adaptive Reweighted Sampling and Uninformative Variable Elimination methods has the optimal performance in predicting the overall watermelon sugar content with a calibration correlation coefficient (RC) of 0.94 and a calibration root mean square error of 0.31°Brix. Correspondingly, the prediction correlation coefficient (RP) and the root mean square error of prediction (RMSEP) were 0.91 and 0.37 °Brix, respectively. Additionally, the number of variables was significantly reduced from 1 524 to 39. This study provides a reference for the practical application of rapid and non-destructive testing of sugar content in watermelon.