Mining large Raman spectroscopic data beyond the shot noise limit

• Main Authors: Shimada, Rintaro, Ozawa, Takeaki
• Other Authors: Asian Spectroscopy Conference 2020
• Format: Conference or Workshop Item
• Language: English
• Published: 2020
• Subjects: Science::Chemistry; Chemometrics; Hyperspectral Data Analysis
• Online Access: https://hdl.handle.net/10356/144412
• Institution: Nanyang Technological University

Various multivariate chemometric techniques have been proven to be robust in analyzing complex Raman hyperspectral datasets obtained from chemically diverse biological samples for classification, quantification, and exploratory studies. Among various techniques, singular value decomposition (SVD) and its mathematical akin, principal component analysis (PCA), are particularly useful because they provide objective summarization of a given dataset by approximation of observed spectra into linear combinations of a finite number of significant basis spectra along with effective signal-noise separation. The number of significant components retained by SVD essentially indicates the wealth of information retracted from the spectroscopic dataset; therefore, increasing the number is deemed important for improving analytical performance, especially when minor species are under investigation. Generally, the use of a larger dataset is considered to yield more significant components. However, the dataset size relationship to the efficacy of SVD summarization has not been extensively studied. In this presentation, we will report results from the systematic study on SVD analysis of Raman hyperspectral datasets of various size, from which the presence of fundamental limitation which prevents recovery of minor signals is unraveled. Furthermore, a possible workaround using variance stabilization transform is demonstrated to overcome the limitation.