Multimodal Visible and Near-Infrared Spectral Fusion Enables Nondestructive Determination of Soluble Solids Content in Shatangju Mandarin

Optical acquisition geometry determines the tissue regions represented in visible–near-infrared (Vis–NIR) spectra, yet its interactions with spectral preprocessing and multimodal fusion remain insufficiently understood for intact citrus fruit. This study systematically compared diffuse reflectance (DR), diffuse transmittance (DT), and transmittance (T) for the nondestructive prediction of soluble solids content (SSC) in Shatangju mandarins. For each optical mode, 180 modeling pipelines combining six preprocessing methods, five wavelength-selection strategies, and six regression algorithms were evaluated using repeated nested cross-validation. Three multimodal fusion strategies were subsequently investigated: direct feature concatenation, sequential orthogonalized partial least squares (SO-PLS), and stacked generalization (Stacking). Marked differences in predictive performance were observed among the three optical modes. T achieved the strongest single-mode performance, followed by DT, whereas DR performed substantially worse, with mean coefficient of determination () values of 0.832, 0.815, and 0.448 and corresponding root mean square errors (RMSEs) of 0.808, 0.849, and 1.463 °Brix, respectively. Preprocessing effectiveness was strongly mode-dependent: Savitzky–Golay smoothing provided the greatest improvement for DT, whereas preprocessing yielded only limited gains for DR. Wavelength selection did not consistently improve prediction over the full valid wavelength range, although competitive adaptive reweighted sampling substantially reduced spectral dimensionality in some pipelines. Neither direct feature concatenation nor SO-PLS surpassed the best single-mode model. In contrast, prediction-level Stacking consistently improved predictive performance. Stacking-6, combining the two highest-ranked base models from each optical mode, achieved the best overall results, with a mean of 0.861, RMSE of 0.735 °Brix, mean absolute error of 0.576 °Brix, concordance correlation coefficient of 0.924, residual predictive deviation of 2.700, and ratio of performance to interquartile range of 3.931. Relative to the best T-mode model, Stacking-6 reduced mean RMSE by approximately 9.0% and achieved lower RMSE in all 10 repetitions. These findings demonstrate that acquisition geometry governs not only single-mode predictive capability but also the effectiveness of preprocessing and multimodal integration. Under the conditions evaluated, transmittance provided the most informative single-mode spectra, whereas prediction-level stacking more effectively exploited complementary information across optical modes than feature-level concatenation or SO-PLS fusion.