To achieve efficient and non-destructive detection of baking quality in Wuyi Shuixian tea, optimization of baking process parameters,and improvement of standardization in rock tea production,a near-infrared spectroscopy-based model for predicting baking-related color quality was developed. A total of 90 Wuyi Shuixian raw tea samples were collected from the Wuyi tea-producing region, covering different baking intensities (light, medium, and heavy fire) and time intervals (30–300 min, at 30-min intervals). Full-range near-infrared spectra (400–4000 cm-1) were acquired using a Thermo Scientific Nicolet IS 5 spectrometer, while the L, a, and b color parameters of tea infusions were measured by a CIE Lab colorimeter. The spectral data were preprocessed using methods such as Savitzky-Golay (SG) convolution smoothing. Key characteristic wavelengths were selected using genetic algorithm combined with partial least squares (GA-PLS),and predictive models were established by coupling support vector machine (SVM),back propagation neural network (BP), and random forest (RF).It was found that the L value decreased exponentially with increasing baking intensity,the a value showed a positive correlation with baking time,and the b value continuously decreased. The GA-PLS-SVM model was determined to be optimal,with calibration set correlation coefficients of 0.963, 0.956, and 0.902 for L, a, and b, respectively,and Root Mean Square Error below 1.25.Validation set correlation coefficients were all above 0.85, indicating good generalization ability.The model is capable of accurately inverting baking parameters and identifying process deviations.It is concluded that the developed model can precisely predict color changes during the baking process of Wuyi Shuixian tea, providing theoretical and technical support for intelligent control and quality standardization in rock tea processing.