PEAMATL: A Strategy for Developing Near-Infrared Spectral Prediction Models Under Domain Shift Using Self-Supervised Transfer Learning

Near-infrared (NIR) spectroscopy combined with spectra prediction models have been widely employed as quick and cost-effective analytical techniques in the pharmaceutical, chemical, and food industries. However, calibration has to be conducted for a prediction model being constructed on data from the source domain if we want to apply the model to a new target domain. Most deep transfer learning methods, which extract domain-invariant features from the source domain samples and transfer these features to enhance the representation ability for target domain data, are available to calibrate prediction models. However, due to the difficulty of measuring samples' reference values (label), the reliance on labeled samples for supervised techniques to extract domain-invariant features remains a major bottleneck. In this study, we propose a novel self-supervised transfer learning approach named PEAMATL for learning and transferring generalized domain-invariant features from samples' spectra, aiming to accurately predict unseen samples' reference values. PEAMATL firstly trains a pyramid encoder consisting of three external attention modules to extract multi-scale features from unlabeled source domain samples using a self-supervised learning framework; secondly, it transfers the pre-trained spectra encoder followed by an initialized prediction head network to build a prediction model; finally, PEAMATL refines the model parameters using a portion of the labeled target domain samples to adapt to unseen target domain samples. The calibration analysis is tested on tablet, melamine, and apple datasets for predicting active pharmaceutical ingredient, turbidity point, soluble solid content, and firmness. Compared with three existing supervised and two self-supervised transfer learning methods, the proposed PEAMATL method achieves at least 3.32 % ~ 30.88 % prediction error reduction on 19 out of 20 scenarios involving three types of domain shift. Therefore, PEAMATL has the potential to be a generic framework for tackling the common problem of domain shift-induced performance degradation of prediction models in the domain of NIR-based quantitative analysis.