| Author | Message | ||
     Dr. Stephan Helary (Stephan) |
Hi, my name is Stephan and I am conducting a research (PhD) on the nutritional value of black rhino diet in different ecosystems in Southern Africa in order to identify nutritional limiting factors that affect reproduction sucess. The lab I am working with for NIR is concerned about the fact that the morphological structures (from woody plant to succulent to forbs) of the different plant species to be analyzed as well as the difference between seasons (2000 samples in total) might be a problem for the calibration. They suggested that I do calibration for each morphological type and for each season, which will considerably increase the costs and time. They are also concerned that I might not pick up differences between species for certain parameters such as bound proteins and minerals. Please, if you have any thoughts or suggestions that can help! | ||
| Bruce H. Campbell (Campclan) |
Stephan, You need not scan all 2000 samples to discern if the differences you list are important or not. You can use subsets. Then by comparing results for each subset you should be able to verify if morphological differences are important or not. If the differences are important, you should ensure there are enough samples in each subset to have more robust calibrations. If morphological differences are not important, you can combine the subsets into one overall sample set. There are at least two other factors you should also be wary of. The first is the time elapsed from harvesting to gathering of the spectra and the second is the location in which the black rhino was grown. There also is the recommended repeating of the measurements on at least two different days to build in at least some resistance to time variation of the spectrometer system. Bruce | ||
| hlmark |
Stephan - this sounds like a good situation for Indicator Variables. These are discussed in "Principles and Practice of Spectroscopic Calibration", Wiley (1991), p.78-80 in the second edition. There are descriptions in some of the books on NIR, as well. Howard \o/ /_\ | ||
| NIRman |
Stephan - you might also read chapter 13 entitled "Indicator Variables: How They May Save Time and Money in NIR Analysis" in the "Handbook of Near-Infrared Analysis," Marcel Dekker, 2nd Edition (2001). Don Burns | ||
| NIRman |
Hi again Stephen, Years ago Dave Honigs wrote a program for Technicon called PICKS. Using only spectra of too many samples, it picked out the most significant ones. Someone reading this might know if the program still exists (perhaps under a different name) and is available from one of the current vendors. Don Burns | ||
| hlmark |
Stephan - PICKS was available as an optional addition to IDAS, the program package that Technicon sold along with their instruments. It selected both the most significant wavelengths and most significant samples from a data set. As far as I know, IDAS is no longer available, nor has it been for a good number of years, from the manufacturer. Some legacy copies of IDAS exist, but copyright restrictions prevent creating new ones. I know how the PICKS program worked, because I worked with Dave Honigs at Technicon at the time he wrote that software. I could rewrite the algorithm (in a more modern language - the original was written in FORTRAN), but that would be a custom piece of software that I would have to charge for, and that would make it a commercial project that would not be appropriate to discuss on this discussion group. Due to that, Stephan, if you're interested in pursuing this approach, please contact me privately, off the discussion group. Howard \o/ /_\ | ||
| Pierre Dardenne (Dardenne) |
Stephen, To do a optimal subset selection, the package from John Shenk (WinISI) contains a routine based on PC distances (SELECT). If you have access to Matlab, special routines exist also. One based on Kennard & Stone algorithm works fine to obtain representative samples. These procedures will select samples on X matrix. On the other hand, a good way to select calibration samples is to predict them with available models from cattle feces or even forages (most similar product). Even if there are biases and poor accuracies, the ranking would be good for protein, fibre, ndf and even OMD. This procedure will add information to select samples on unexisting Y matrix. Pierre |