| Author | Message | ||
     Ivelin Iliev Rizov (Ivrizov) |
Dear Colleges, Now I must choose strategy for producing of calibration model for mixed grass forages from generic models of individual grasses. In a lot of publication (especialy in JNIRS) I found that the most common algoritm for these purposes is LOCAL, but I do not have any experiance with using it. Is the method so good? What is the simpliest way for using it? What is statistic theory behind it? All advises will be highly appreciate! Best regards, Iv. Rizov | ||
| W. Fred McClure (Mcclure) |
Iv. The term "local" is little confusing. It is not a regression method - rather, it is a sample selection technique. Fundamentally, it means you calibrate ONLY over the range in which you are interested. Hence, if your process produces samples with 10% to 16%, it "may be" best to calibrate only for this region. However, you would want to look at the extended range just to satisfy yourself that local regression is better. It just may be that it is not better. Mac | ||
| Tony Davies (Td) |
I have to correct my friend! LOCAL employs a large database of analysed samples. Each new unknown sample is compared to every member of the database to find a sub-set of very similar samples. These are put into a special version of PLS to produce a calibration for that sample ONLY! The procedure is repeated for every new sample. It is currently exciting and the results appear to be good but it should be regarded as experimental. Harald Martens said that the Stistical Community regard PLS as statistically unsound; I can guess what they would say about LOCAL. P.S. I have asked John Shenk to find a new name for LOCAL; people get cofused with Locally Weighted Regression (LWR) which is something different. (How about LSPLS? [Local Samples Partial Least Squares]). |