Hyperspectral imaging for cross-formulation prediction of spatial concentration distributions and non-destructive evaluation of mixing uniformity

Hyperspectral imaging (HSI) provides a non-destructive means of acquiring both spectral and spatial information, enabling spatial concentration mapping and mixing uniformity evaluation. However, formulation variations can cause spectral response shifts that significantly degrade the performance of concentration prediction models when transferred across formulations. To address this challenge, a spectral-spatial feature fusion and domain adaptation framework (SSFT-DC) is proposed for cross-formulation concentration prediction using HSI. The proposed framework utilizes both the average spectrum and local spatial-spectral patch of each sample as complementary inputs. Concentration-related features are extracted through spectral and spatial branches and adaptively integrated using a gated fusion mechanism. To improve transferability across formulations, domain adversarial learning and conditional correlation alignment (CORAL) are employed to reduce inter-formulation feature distribution discrepancies while preserving concentration-related information. SSFT-DC achieved better concentration prediction performance than the comparison methods in both formulation systems, which differed in plasticizer type. On the target-domain test set, the proposed method achieved a coefficient of determination (R²) of 0.9836, a root mean square error (RMSE) of 0.7522, and a mean absolute error (MAE) of 0.5864, outperforming five representative transfer learning baselines, namely BiLSTM-FT, DL-DA, LSLT, TRN-DA-R, and DAN-MFAC-R. Ablation studies and feature visualization further confirm the effectiveness of spectral-spatial fusion and cross-domain alignment. Furthermore, under the investigated concentration conditions, the coefficient of variation (CV) calculated from model-estimated concentration maps exhibits an overall decreasing trend with increasing mixing time, indicating a gradual reduction in the dispersion of estimated concentrations. These results demonstrate that SSFT-DC enables cross-formulation concentration prediction and has potential for non-destructive, auxiliary characterization of changes in mixing state.