An unsupervised domain adaptation regression method in kernel partial least squares subspace embedded with joint statistical and manifold alignment for Fourier-transform infrared spectroscopy in agri-food analysis

Abstract

Within the agri-food sector, the precise measurement of essential ingredients in samples across different measurement contexts using Fourier Transform Infrared spectroscopy (FTIR) is crucial, underscoring the need for advanced calibration methods with extensive generalizability. Domain adaptation (DA) in machine learning is a pivotal area of research focused on training models to be adaptable to both source and target domains with differing data distributions. This paper delves into the application of unsupervised domain adaptation (UDA) for FTIR analysis in agri-food products, utilizing unlabeled data from the target domain to address the challenge of limited reference samples. To realize complex nonlinear adaptation, combining the advantages of statistical alignment and nonlinear ability from domain-invariant iterative partial least squares (DIPALS) and kernel domain adaptive partial least squares (da-PLS) respectively, a novel UDA regression method in kernel partial least squares subspace embedded with joint statistical and manifold alignment (JSMKPLS) is present by further integrating a manifold alignment strategy that could incorporate geometric nonlinear structure into the adaptation process. The framework simultaneously exploits the statistical and geometrical properties in reproducing kernel Hilbert space (RKHS) and extract the domain invariant features. Experimental results of corn, rice, γ-PGA fermentation and wheat datasets confirm the effectiveness of JSMKPLS for FTIR analysis.