Improved FTIR-based classification for food authentication using a topological ensemble framework.

Ensuring the authenticity and quality of food products is a critical challenge in food science and technology, particularly when compositional differences are subtle and not visually distinguishable. Fourier-transform infrared (FTIR) spectroscopy provides a promising analytical tool for food authentication, yet practical spectra often exhibit baseline drift, scattering effects, and overlapping absorbance bands that complicate reliable discrimination. As a result, these nonlinear and overlapping spectral signals can limit the performance of traditional chemometric approaches such as principal component analysis and partial least squares regression, which are widely used for dimensionality reduction and regression but often fail to generalize to nonlinear spectral data. To address these limitations, this study introduces a framework that integrates sliding-window embeddings with topological data analysis (TDA) to capture local, meso-scale, and global spectral structures, followed by bagging-based ensemble learning for robust prediction using transparent base learners. Six food-related FTIR benchmark datasets from the UCR Time Series Classification Archive were comprehensively evaluated under the official fixed train-test splits, without additional dataset-specific preprocessing. Results show that the proposed method demonstrates consistently robust performance relative to 20 widely used time-series classifiers across distance-based, feature-based, ensemble, and deep learning families under the standardized UCR benchmark setting, achieving high classification accuracy and high Macro-F1 across diverse food categories, based on reported results in the literature. The analysis further illustrates how multi-scale topological descriptors enhance interpretability and provide insights into spectral variability. Overall, the framework balances transparency, robustness, and computational efficiency, providing a reproducible benchmark-based reference for subsequent studies that evaluate measurement-specific effects under controlled acquisition settings.