| Author | Message | |||
     Cesar Guerrero (Cesar) New member Username: Cesar Post Number: 3 Registered: 3-2006 |
Many thanks again to all!! Very nice information! The differences between calibrations in nm or in cm-1 are not too much different, but in wavelengths are always slightly better. Now, I'm working with a FT, with 8 cm-1 as resolution. When spectrum is transformed to nm (or microns as in the pdf) changes the 'shape' (sorry if some word is incorrect), being now similar that others (soil samples too) when I was working with another spectrometer with the resolution was in nm (2nm of resolution). I'm thinking that PLS could be influenced by the shape and/or size of the peaks. The software found the best regions along the spectrum, but in a region could be more or less information depending of the scale (wavenumber or wavelenght). That's only a poor hypothesis, and probably incorrect..... Thanks again to all!! Best wishes Cesar Guerrero | |||
| Tony Davies (Td) Moderator Username: Td Post Number: 120 Registered: 1-2001 |
Hello Cesar, welcome to the group and thanks for giving us the spectra. Many people do not! The question of converting bewteen scales has been one of my interests for some time. I assume that your software has done the transformation to equally spaced wavelength intervals via some sort of interpolation. This is not ideal. Tom Fearn and I have demonstrated how it can be done by FT treatment of the data but Peter Griffiths did a much, much better job. JNIRS 11,229 - 255 and NIR news 15(5) 10-11. You have to compute the correct resolution for every point in the transformed spectrum, which is different for every point! Not a job to be undertaken lightly. Your actual question is why do you get better answers with the transformed data; this is not what we would expect. My thoughts are: 1) Some studies have shown that high resolution wavenumber data (1 or 2 wavenumber) is not as good as moderate (8 wavenumber) data. What is the resolution produced by your spectrometer? 2) Is the spectrum in your .PDF one of you samples? If it is I think it shows that the wavelength spectrum has better resolution in the regions which contain most information (peaks). This will aid the calibration. Best wishes, Tony | |||
| David W. Hopkins (Dhopkins) New member Username: Dhopkins Post Number: 68 Registered: 10-2002 |
Cesar, Now, that is an interesting observation. It sounds like the transformation from cm-1 to nm and reduction in numbers of data points is giving a slight benefit to the regression. That is an unusual way of preprocessing the absorbance data. I would recommend staying with the cm-1 data and applying traditional preprocessing such as smoothing or derivatives, MSC or SNV to obtain optimal results. Thanks for sharing your observation and questions with us. Best wishes, Dave | |||
| Gabi Levin (Unregistered Guest) Unregistered guest |
Howard is right in his addition in the event that you have an instrument that produces data in say cm-1 and you want to convert to nm, or vice a versa, than the numerical treatment of data will affect the quality of the data, because numerical is only numerical, to the better or the worse. The reason I said there shall be no difference is in the theoretical aspect, and in fact, if your instrument produces data in cm-1 stick to it because you will get the best out of the numerical treatment, and so is the truth if your instrument produces data in nm - stick to it. At the same time, if we could convert the discrete data set of data points into a perfect mathemtical function f= ax +bx2 + cx3 (powers) or some sort of a continuous function, the difference between the methods shall disappear. Thanks, Gabi Levin Brimrose | |||
| hlmark (Unregistered Guest) Unregistered guest |
Cesar - what the previous commentors said was true, but incomplete. There are two ways you can get data in cm-1 or in nm: 1) You're using different types of instruments. As Gabi said, FT-based instruments inherently produce data linear (i.e., constant frequency spacing between data points) in frequency (wavenumbers), while diffraction grating-based instruments inherently produce data linear (i.e., constant wavelength spacing between data points) in wavelength (nanometers). 2) You have data from a single instrument that you've converted mathematically into the other domain. Let's consider case 2 first, since it's the simpler case. Case 1 includes all the considerations from case 2, plus some has some more of it's own. If you simply divide 10,000 by the wavenumber, you'll get the value of the corresponding wavelength, in microns. However, since division is not a linear operation, the spacing between data points is no longer constant, and therefore the wavelength scale is no longer linear. The same thing happens if you convert the other way. What happens next depends on the software you're using. Many programs that do wavelength-to-frequency conversions implicitly include some sort of interpolation routine to provide a wavelength/wavenumber (as appropriate) scale that is linear in the target domain. Therefore the results you get will depend on the type of interpolation routine used. If the conversion program is intended to mimic the data you'd obtain if you actually used the other type of instrument, then in addition it will have to simulate the other characteristic differences between the data obtained by the different instrument types. So lets go back to point 1 and consider those. Each type of instrument not only has a linear scale, but for each technology the spectral resolution is constant across the wavelength range when expressed in terms of the native units appropriate to that technology. If you compare the spectral resolution of one technology to the other, what you find is that the "other" technology has a non-constant spectral resolution, as well as a non-linear scale. If you go through the math, what you find is that the resolution varies with the square of the units in the new domain. Therefore, to simulate spectra taken with one type of instrument as being from the other type, you not only have to adjust the scale, you also have to take into account the differences in spectral resolution characteristics. This may mean doing some sort of weighted averaging at one end of the wavelength range, changing the weights as you go toward the other end, and maybe even do some sort of spectral deconvolution to enhance the spectral resolution of the measured spectrum at the other end of the range. Spectral deconvolution is a risky procedure, since it is likely to introduce artifacts in the computed spectrum. Now, when you do calibration work, and especially if you're using data transformations, the interpretation of the results is complicated by all the foregoing considerations. For example, doing a simple boxcar averaging over, say, 30 nm in wavelength space, and then doing a boxcar averaging over the same number of data points after conversion to wavenumbers, might mean that you're averaging the equivalent of 200-300 nm at one end of the scale and maybe 3-5 nm at the other end. There are lots of complications that get introduced. Howard \o/ /_\ | |||
| Cesar Guerrero (Cesar) New member Username: Cesar Post Number: 2 Registered: 3-2006 |
Dear all, Thank for the answers! I'm starting to work with NIR, and I probably making a 'big error'. Initially I thought that have the data in cm-1 or in nm (or microns) was equivalent... but, when I applied the transformation from wavenumber to wavelenght the spectrum is different... probably that is due to the way that I'm transforming: my spectrum in cm-1 was composed of 2126 points, and after transforming to wavelenght was 1000 points (and option in OPUS software). The spectra are different... The calibrations are slightly better when I work in wavelenghts... ��?? Could be due to the higher area of peaks? I attached a pdf with the spectrum in both ways (in wavenumber and in wavelenghts) Thanks again to all for helps!!
| |||
| David W. Hopkins (Dhopkins) New member Username: Dhopkins Post Number: 67 Registered: 10-2002 |
Cesar, I agree with Gabi, our calibration algorithms do not consider the wavelength values. Therefore, I am puzzled that you say you find differences. I expect differences if you select different wavelength regions, even if only slightly different. And, most transformations do not consider the wavelength values, or even whether they are evenly spaced. Even though the Savitzky-Golay algorithm computes derivatives and expects to operate on data points evenly spaced in wavelength, you will obtain the same values whether you express the wavelength in cm-1 or nm! That is because only the ordinate values are actually used in the computations. They will appear different to the eye, because of the different spacing of the points, but they will be the same amplitudes. If you take the same sample set and run them on 2 instruments that use essentially the same wavelength regions, but they employ nm or cm-1, then I certainly expect different results, due to different instrument responses. However, the experience of a large number of grating and FTIR instrument users is that they are satisfied with the accuracy of the results. Can you be more specific, under what conditions are you seeing differences, and what are the nature of the differences you are inquiring about? Regards, Dave | |||
| Gabi Levin (Unregistered Guest) Unregistered guest |
Hi Cesar, Do you expect the temperature of a given object to be different if you name it in farnheit or in Celsius? We would be in terrible situation if the way we express wavelngth affected the quality of calibration. However, if the question is what is more convenient to use, then answers will differ - the FT people are used to cm-1, and they will answer cm-1, the dispersion like grating, AOTF, etc will prefer microns or nm because it is the most straight forward for them. If you have FT, use cm-1 if you have AOTF, or grating, use nm. The Atlas of NIR spectra usually provides charts in nm and cm-1, but nm is more useful in finding reference peaks by linear interpolations from the scale. Gabi Levin Brimrose | |||
| Gabi Levin (Unregistered Guest) Unregistered guest |
Hi Cesar, Do you expect the temperature of a given object to be different if you name it in farnheit or in Celsius? We would be in terrible situation if the way we express wavelngth affected the quality of calibration. However, if the question is what is more convenient to use, then answers will differ - the FT people are used to cm-1, and they will answer cm-1, the dispersion like grating, AOTF, etc will prefer microns or nm because it is the most straight forward for them. If you have FT, use cm-1 if you have AOTF, or grating, use nm. The Atlas of NIR spectra usually provides charts in nm and cm-1, but nm is more useful in finding reference peaks by linear interpolations from the scale. Gabi Levin Brimrose | |||
| Cesar Guerrero (Cesar) New member Username: Cesar Post Number: 1 Registered: 3-2006 |
Hello! Please, if anyone can help me... What is better for a PLS calibration in the NIR region data: use data in cm-1 or in nm??? I found some differences.... Why?? Many thanks!! Cesar |
graph of the same spectrum