DCT2PC: Direct Calibration Transfer to Principal Components Using Ensemble Extreme Learning Machine for Near-Infrared Spectroscopy (NIRS)

Near-infrared (NIR) spectroscopy is a rapid, nondestructive technique for quality assessment. A persistent challenge, however, is that optical component variations across different instruments introduce systematic distortions in spectral responses. As a result, calibration models developed on a master instrument typically lose predictive accuracy when applied to slave instruments. Calibration transfer techniques aim to correct such instrument-induced spectral differences without requiring the laborious recalibration work. DCT2PC (direct calibration transfer to principal components) is a framework that directly maps slave spectra to a principal component (PC) space derived from the master spectra via singular value decomposition (SVD). An ensemble of extreme learning machines (ELMs) is trained to learn the nonlinear mapping from slave spectra to the corresponding master PCs. Transferred spectra are subsequently reconstructed by multiplying the transferred PC scores with the right singular vectors obtained from SVD. DCT2PC requires no intermediate linear calibration steps and relies on a PC space independent of instrument response, which enables the model to accurately and efficiently transfer key information. Another ELM-based calibration model was developed to predict the component contents. The proposed method was validated on three public NIR benchmark datasets. Across all datasets, DCT2PC-ELM consistently achieved lower root mean square errors of prediction (RMSEP) compared to several established methods, including CTCCA, SST, PDS, TEAM, and a PLS-based variant (DCT2PC-PLS). Statistical significance testing confirmed the superiority of the proposed approach. This framework offers a practical solution for cross-instrument NIR calibration without costly spectral recalibration.