| Author | Message | ||
     Hoang Buu Quoc (oilman_vn) New member Username: oilman_vn Post Number: 1 Registered: 8-2007 |
Thank you for your suggestion, Could you give me more about the using of NIR for fuel analyzer, I am using a PetroSpec instrument for quick fuel analysis in the field, my email: [email protected] Thanks, | ||
| jo |
Hello, (sorry if yhis message is repeated) I'm an honours student trying to develop models using NIRS for various leaf components. I'm using a Bruker MPA and i'm having trouble using it. All of the literature i have read talks about scanning many samples prior to chemical analysis to get the best outcome using the fewest samples(avoiding unnecessary chemistry on samples that are "too similar or different" Apparently i can't scan samples and compare the mahalanobis distance without first having a rudimentary model at least-but this makes no sense as i thought the whole purpose was to scan alot of samples prior to chemistry. I was told if i made a rudimentary model, then i would be given the mah. distance on any other samples i scanned and then i could choose which samples would be best to add to the model. I'm getting confused though. I thought the mah. distance was calculated using the whole NIR region(because you do it before knowing anything about where you may look spectrally for each component), but it seems with the bruker software that the mah. dist. is calculated after the model has been developed and uses the range that a given component was modelled from. This is getting lengthy..............can anyone help me with this-i would appreciate it very much jon | ||
| hlmark |
Jo - the dichotomy has to do with the measurement of the lab (reference) values, which is usually more time-consuming and costly than collecting the optical (spectral) data. Therefore, the procedure that has been developed has been to use the NIR instrument to scan all the samples available and calculate the Mahalanobis Distances for those samples, based on only this spectral data. Then you can use those to select samples to be analyzed by the reference lab method, thereby minimizing the time and cost that step requires. If your software doesn't permit this, then you'll have to obtain software that can calculate Mahalanobis Distances without having to perform a calibration first. If you contact me off the discussion I can make a suggestion. Alternatively, you can simply measure all the samples and do the calibrations. This will take more overall time but should provide a satisfactory calibration. Howard \o/ /_\ | ||
| jon |
hello Howard, i've been unable to email you at the address that comes up when i click on hlmark@nir.... the message gets sent back. do you have another address or would you mind emailing me at [email protected] and i can send a reply thanks jon(jo was a typo) | ||
| Nuno Matos (Nmatos) |
Jon, The Mahalanobis Distance (MD) should not be applied to the spectra because there would be a lot of variables. I think the best way to apply MD is to apply it to PCA scores space. Furthermore, you would have the spectral residual as a second discriminant parameter. If you want to calculate "distances" prior to chemical analysis you should try correlating each spectra to an average spectra. Other methods are possible. I don't know if I've answered you. Best regards Nuno | ||
| jon |
hello, I am wondering when people do the mahalanobis distance and other spectral averaging techniques to derive an average spectra(which other spectra can be compared to).........what spectral region do people do this analysis on? For example, is the whole NIR region(750-2500nm) used or should the range be limited to a more narrow region. In my case, there seems to be very little variation below 1000nm or above 2400nm. I want to scan a large number of samples and then do chemistry on a smaller amount based on their mahalanobis score(ie use samples >0.6 and <3 standard deviations from the mean of all spectra) Does anyone have any thoughts on the spectral range i should use for calculating spectral averages? I would greatly appreciate any suggestions jon | ||
| Eric LALOUM |
Jon, You can differentiate two types of analysis where H value is involved: 1. exploratory one to find out some "special" samples for which it can be useful to measure the Ref 2. modelling where a calibration is build between spectra and chemical values. A high H value means that the model might not apply well to the sample and it is useful to validate the chemical value by the Ref method It seems from your question that you are in the first case, so there are no reason to reduce the spectral range and I suggest not to pre-treat data in any kind (neither wavelenght selection nor derivation or wathezver). When you are in the second case, then you have to compute H in the same space where you are building the model, so you might even compute H from PCR or PLS factors. Eric | ||
| jon |
thanks for your reply Eric-that was helpful jon |