I have not used it , but ICA (Independent component analysis) may help !
It really depends on the objective.
PCA, PLS regression are most common.
If you cannot pick up difference straight away, PCA/PLS on first and second derivative spectra is next option.
Which method for what? Following N.N. Misra, the objective narrows down the possible list of statistical methods. For just visualizing data, PLS regression might work fine, as it realized both subspace projection and data fit in a joint way and offers some way to estimate relevant wavenumbers. PCA, on the other hand, is completely unsupervised, but PC regression can be used in similar vain for supervised problems. When doing classification with NIR data, PLS discriminant analysis if often employed in chemometrics. Other methods from the pattern recognition domain, like svm, however, are becoming more and more popular for this objective.
It depends on for which purpose you are using MA in NIRS. Easiest and mostly used statistical approach is PCA for grouping samples based on spectra. If you are interested in developing calibration model and doing prediction of a parameter using spectra--PLS regression, MLR, are a few of them. If you are interested in discriminating your samples then LDA, SIMCA, KNN, PLSDA are the useful approaches. Before that you have to know some statistical terms used in MA such are R2, SEP, SEC, RMSCECV, REMSEP, Bias, SDR, slope etc. My advise is better get a good book on Multivariate Analysis on Chemometrics and stat playing with your data. Best of luck.
Classification and pattern recognition technique is best for nir spectroscopy.
Principal component analysis with nir spectroscopy is perfect match.
On the other way classification method K - nearest neighbour ( K- NN is bench mark to measure other method) partial least square (PlS most popular for linear discriminant analysis),Soft independent modelling class analogy (SIMCA robust discriminant analysis).
Statistical classifications were made on basis of variance ratio F .
Nir global calibration really challenging.
Pattern recognition is usually four distinct step . 1 ) data preprocessing, feature selection, mapping and display or clustering and classification.
Finally problem are solved dependent upon nature of the problem .
- Badges
- Science topic