| Author | Message | ||
     Chia (xaivier) Member Username: xaivier Post Number: 14 Registered: 4-2010 |
Hi Howard, Your explanations are very helpful for me to move forward with my NIR project. Thank you. Best wishes, Xaivier Chia | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 343 Registered: 9-2001 |
Xaivier - What Dave said is correct, but is only part of the story. In the early days of NIR analysis, MLR was essentially the only algorithm available for developing calibrations, and since it used a small set of wavelengths, it was well-suited to the filter instruments that were available at the time. When full-spectrum instruments became available, it quickly became apparent that they brought with them a new set of problems to be solved: not only was the question of how many wavelengths were needed at issue (similar to the question of how many factors to use for PCR or PLS) but also, for any given number of wavelengths to include in a calibration model, the question arose: WHICH wavelengths were best to use, and how can they be found? It turned out that it was well-nigh impossible to define a set of "best" wavelengths for any given analysis, and worse, selecting even a small number out of a full spectrum was a combinatorial problem. The problem was solved in a practical sense by forgetting about trying to find the "best" wavelengths, and simply seeking sets of wavelengths that worked "adequately" well, for the desired analysis. Some work was done at that time on investigating effects of wavelength selection; you might want to read the following articles: Spectroscopy, 3(11), p.28-36 (1988) Applied Spectroscopy, 42(8), p.1427-1440 (1988) Articles like these made it apparent that theory didn't help to much. In practice, therefore, in order to achieve useful results, several algorithms were developed that could do a less-than exhaustive search, while still being able to find useful wavelength sets that gave adequate predictions. Some of these were step-up methods, step-down methods, step-sideways methods (this last being my semi-humorous term for algorithms that attempted to improve search efficiency by changing several wavelengths at a time; these include genetic algorithms, simplex optimization, and the like. When full-spectrum algorithms such as PCR and PLS were developed, they were immediately grabbed up. There were several reasons for this, including the "pizzaz" factor, but one very real and very important benefit was that it completely sidestepped the need to perform wavelength searches, with their long running times and inherent uncertainties. So yes, there are known methods for doing what you are asking for, but they are not "magic answers to all your calibration problems". Depending on what you are doing you may want to investigate them, however, keeping in mind that no algorithm is without it's problems. Finding descriptions of the PLS and PCR algorithms is not at all difficult; almost any book on chemometrics contains description of those algorithms. One book that is oriented toward novices is Richard Kramer's "Chemometric Techniques for Quantitative Analysis" but almost any book with "Chemoemtrics" in the title will describe those algorithms. Howard \o/ /_\ | ||
| Chia (xaivier) Member Username: xaivier Post Number: 13 Registered: 4-2010 |
Hi Dave, Thanks for your suggestion. This is very comprehensive and useful for me. Regarding the use of full wavelength methods to develop instruments that employ selected wavelengths for routine measurements, would you mind to suggest few literature? Best wishes, Xaivier Chia | ||
| David W. Hopkins (dhopkins) Senior Member Username: dhopkins Post Number: 161 Registered: 10-2002 |
Hi Xaivier, The literature has many examples of using using studies using full wavelength methods to develop instruments that employ selected wavelengths for routine measurements. In fact, that is precisely how many filter NIR instruments were developed. However, if you intend to continue to use a full wavelength instrument for further use with the samples, then I would recommend that you stick with the full wavelength calibration methods. The advantage is not so much with the calibrations, as with the error checking that is possible. You can (and should) qualify that a sample is a member of the population used for training the calibrations, for example. Some pretreatments to remove spectral variability due to sample physical properties such as variation in granularity, can best be applied to multivariate data, rather than a selection of fewer wavelengths. Good observation, and there is much to say, depending on your goals for the measurements. Best regards, Dave | ||
| Chia (xaivier) Member Username: xaivier Post Number: 12 Registered: 4-2010 |
Hi all, After studying about dimension reduction via PCA, I was wondering is it possible to use PCA or other methods to select the effective wavelengths of spectrum in such ways that the performance of calibration model can be improved without using PCA pre-processing. In other words, I was thinking to use the effective wavelengths of a spectrum as the input of calibration model instead of whole spectrum. In this case, I only have input spectra data and SSC reference value. I am looking forward to your suggestion, comment, or recommendation. Thank you very much. Best wishes, Xaivier Chia |