A robust ensemble chemometric strategy for near-infrared determination of ash content in fishmeal

Near-infrared (NIR) spectroscopy combined with machine learning algorithms has been widely used for rapid and non-destructive quantitative analysis, but accurate quantitative modeling remains challenging due to the high dimensionality, collinearity, and redundancy of spectral variables, as well as the limited size of calibration datasets. In this study, a customized ensemble framework termed RFXGB-DA was developed for predicting fishmeal ash content. The proposed method combines the bootstrap sampling and random feature selection mechanisms of Random Forest (RF) with the nonlinear regression capability of eXtreme Gradient Boosting (XGBoost), and incorporates a random mixed data augmentation strategy to simulate realistic spectral perturbations. This framework aims to reduce the influence of redundant variables, enhance model diversity, achieve a better balance between bias and variance, and improve robustness against sample heterogeneity and measurement variability. The performance of RFXGB-DA was evaluated on a fishmeal NIR dataset and compared with Partial Least Squares (PLS), RF, XGBoost, and RFXGB. The results showed that RFXGB-DA achieved the best predictive performance, with R_T² = 0.9375 ± 0.0146, RMSET = 0.9317 ± 0.1067 and RPD = 4.0766 ± 0.4527. In addition, ablation analysis demonstrated that both the ensemble mechanism and data augmentation played significant roles in improving calibration performance. SHAP analysis further indicated that the model captured chemically meaningful spectral regions associated with indirect responses from the organic matrix and water related to minerals. Overall, the proposed method offers a robust and interpretable chemometric strategy for the rapid determination of ash content in fishmeal using NIR spectroscopy.