Abstract: The total acidity is an important index in the citrus internal quality assessment. In order to minimize variable collinearity effects in the calibration data set, reduce the modeling variables to alleviate the computation workload, a novel variable selection strategy-successive projections algorithm (SPA) was employed to optimize the near infrared spectrum testing model of citrus total acidity. The splice correction method was used to correct the original NIR spectra. The outlier samples were analyzed by studentized residual error and regression line. After outlier samples eliminated, The SPXY (sample set partitioning based on joint x-y distances) method was used to subset partitioning. Finally, the “Successive Projections Algorithm” (SPA) was applied to select the optimal sets of variables for calibration, and then the prediction performance comparison between the model built by the selected variables and full-spectrum-PLS model, the influence of the orange peel to the total acidity model prediction accuracy is also analyzed in this paper. As can be seen, nine and thirteen optimal effective variables were selected from full-spectrum variables by SPA, for the total acidity determination with the whole fruit samples and the flesh samples, respectively. SPA-MLR, SPA-PLS and full-spectrum-PLS were comparable in terms of prediction performance for the total acidity determination with the whole fruit samples. The prediction set correlation coefficient (Rp) of the total acidity determination with the whole fruit samples was 0.829470, 0.837095 and 0.857299, respectively. While SPA-MLR and SPA-PLS resulted in models with good prediction ability when compared to full-spectrum PLS model for the total acidity determination with the flesh samples. The Rp of the total acidity determination with the flesh samples was 0.819430, 0.825277 and 0.780146, respectively. The SPA improves the prediction ability of the total acidity determination with the flesh samples effectively.