Data Augmentation Strategies for Terahertz Spectral Material Classification: A
Systematic Study on Model-Strategy Interaction
Zhangchong Ye,
Lifeng Zhu,
Luoyi Li,
Jiangyang Zhan,
Lei Zhang,
Jie Cai,
Jing Shen,
Chong Tian,
Xiaofeng Zhang and
Xianliang Li
Small sample sizes in terahertz (THz) spectroscopy constrain the generalization of machine learning classifiers, motivating data augmentation as a low-cost remedy. However, researchers rarely evaluated the effectiveness of augmentation strategies across diverse classifier architectures. This study cross-evaluates five augmentation strategies (Gaussian noise injection, amplitude scaling, baseline drift, a combined version of the three perturbations, and intra-class Mixup interpolation) on a five-material THz insulating material dataset (129 spectra). Eleven classifiers spanning gradient boosting, bagging ensembles, sequential boosting, kernel methods, linear models, generative models, and neural networks were benchmarked under each augmentation strategy plus a no-augmentation baseline control, yielding 66 independent experiments (6 conditions × 11 models), with hyperparameters tuned via 5-fold stratified cross-validation. On the single fixed partition of the main experiment, the interaction appears large: Mixup improves multi-layer perceptron macro F1 by 83.7% (from 0.416 to 0.765) while AdaBoost degrades by 56.9% (from 0.743 to 0.321), and the best combination, Extra Trees with baseline drift, reaches macro F1 of 0.871. Because such effect sizes may depend on the partition drawn, every cell was reevaluated across 10 independent stratified splits with hyperparameters held fixed. Single-partition magnitudes prove substantially inflated, and the two extreme cases diverge: the MLP gain persists in direction but shrinks to +18.0% (paired ΔF1 = +0.099 ± 0.133), whereas the AdaBoost degradation does not reproduce (+0.064 ± 0.119, 5 of 10 splits negative), its apparent size traceable to a baseline of 0.743 that no repeated split recovered (mean 0.540). What survives is the family structure rather than the individual percentages: augmentation helps neural network (+0.110), kernel (+0.042), and bagging (+0.021 to +0.026) classifiers and harms linear-generative (-0.087) and generative (-0.034) ones. Baseline drift remains the most consistent strategy, its mean gain falling from 0.038 to 0.011. These findings demonstrate that augmentation strategy selection must be guided by classifier architecture, and that effect sizes from a single small-sample partition should be treated as upper bounds. The defensible framework is at the family level: augment neural network, kernel, and bagging classifiers, and avoid augmenting linear-generative and generative ones. Independent-dataset validation is required before generalizing these recommendations.
