SpectrumImager uses an iterative NIPALS algorithm to extract principal components. The Convergence parameter under Settings → Proceed → PCA → Iteration parameters defines the maximum difference between two consecutive loading vectors. To avoid infinite loops, you can also set the Max iterations number.
SpectrumImager employs a proprietary TEMDM algorithm to filter outliers that can distort PCA. Outliers are detected on the fly within the NIPALS loops and interpolated using neighboring pixels. The detection threshold is given in standard deviations from the mean (default: 4.0). TEMDM aö also limits the maximal number of outliers removed during extraction of each principal component. Adjust Threshold and Max to remove under Settings → Proceed → Outliers parameters.
PCA performs best with Gaussian noise, whereas real spectrum-images exhibit mainly Poisson noise. Enable noise equalization under Settings → Proceed → PCA → EELS/EDX Weighting and choose between Keenan (M.R. Keenan and P.G. Kotula, Surf. Interface Anal. 36 (2004) 203–212 ) or Anscombe (F.J. Anscombe. Biometrik, 35(1948) 246–254) methods. This can be set separately for EELS and EDX datasets.
EELS spectrum-images acquired with 4-quadrant CCD cameras may show gain mismatches, producing a step in the middle of spectrum. SpectrumImager corrects this on the fly within the NIPALS loops. Enable the correction under Settings → Proceed → PCA → EELS Quadrants.
EDX datasets can be highly sparse, degrading PCA accuracy. If sparsity (fraction of non-zero elements) is below a certain threshold, SpectrumImager automatically applies Gaussian blurring prior PCA. Configure the Sparsity threshold and blurring Sigma under Settings → Proceed → PCA → EDX → Blur.