The prediction of sugar content (SC) in citrus by near-infrared spectroscopy (NIRS) and sensory test was investigated the validation whether the result of non-destructive determination methods by NIRS can meet the request of consumers’ sensory or not, and the simplification of the prediction model of NIRS for citrus’s SC with variables selection on the basis of meeting their demands. Result of the latter analyzed by one-way ANOVA shows that there was a significant difference influenced by individual diversity, but not by gender. After excluding the sensuous outliers, root mean standard error of deviation (RMSED) of every participator was calculated and the minimum equaled to 0.633, which was chosen as borderline of NIR model’s RMSEP to meet the sensory request. Then, combined with spectral preprocessing and variables selection methods, SPA-MLR model was obtained by its robustness with Rp=0.86, as well as RMSEP=0.567 for prediction set, furthermore, prediction time just costs 6.8 ms. The achievement that not only meets the customers’ sensory, but also simplifies the prediction model can be a good reference for real time application in future.