Statistical consequences of applying a PCA noise filter on EELS spectrum images

Area of Science:

• Materials Science
• Spectroscopy
• Data Analysis

Background:

• Principal Component Analysis (PCA) is widely used for noise reduction in Electron Energy Loss Spectroscopy (EELS) spectrum images.
• Accurate noise filtering is crucial for reliable interpretation of experimental EELS data.
• Understanding the statistical behavior of PCA is essential for its effective application.

Purpose of the Study:

• To statistically investigate the behavior of PCA noise filtering on simulated EELS data.
• To identify and quantify the bias introduced by PCA noise filtering.
• To develop criteria for assessing the suitability of PCA filtering in EELS analysis.

Main Methods:

• Application of PCA noise filtering to a simulated EELS dataset with realistic noise.
• Utilizing least squares fitting and parameter estimation theory.
• Analysis of singular values and their bias/precision using established literature expressions.

Main Results:

• PCA filtering, while potentially improving precision beyond the Cramér-Rao lower bound, introduces significant bias.
• Bias originates from inaccurate retrieval of principal loadings in noisy data.
• An evaluation criterion for singular values based on noise level and data information content was developed.
• PCA filtering does not guarantee improved signal-to-noise ratios in elemental mapping due to spectral data correlation.

Conclusions:

• PCA noise filtering in EELS can lead to biased results, potentially altering scientific conclusions.
• A novel criterion can guide decisions on when to avoid PCA filtering for EELS data.
• The effectiveness of PCA filtering depends critically on noise levels and data characteristics.