• home
• current students
• graduate
• events

Presenter: Yifeng Bie
Supervisor:

Date: Fri, September 11, 2026
Time: 09:30:00 - 00:00:00
Place: Zoom - see below.

ABSTRACT

https://uvic.zoom.us/j/8472200636?pwd=NTVCbHhZcnJYSU80VXlkSGlsMlF3Zz09&omn=88271444099

Meeting ID: 847 220 0636
Password: 700022

Note: Please log in to Zoom via SSO and your UVic Netlink ID

Abstract: 

Spectroscopic quantification often depends on large labeled datasets and computationally intensive models, which can limit practical deployment in real-time sensing applications. This work investigates the intrinsic structure of spectral data and shows that, when the signal dominates noise, a mixture containing chemical components can be effectively represented by approximately essential principal components. This low-dimensional structure indicates that much of the observed spectral dimensionality is redundant or noise-related. By retaining only the dominant functional principal components, the most informative spectral variation can be captured while substantially reducing the dimensionality of the original data.

Based on this observation, we develop an fPCA-based linear regression model and a nearly training-free quantification approach that uses known single-component extinction spectra to construct the spectral basis directly. This reduces dependence on large training datasets and simplifies the learning process to a compact concentration-mapping problem. The proposed methods achieve competitive quantification accuracy compared with conventional approaches such as PLSR and XGBoost, with the training-free method showing particular advantages in low-sample regimes. The framework is evaluated using simulated nine-gas infrared mixtures and further validated experimentally with Orange G and crystal violet solutions, demonstrating its effectiveness under spectral overlap, noise, and experimental variability.