Author |
Message |
Karl Norris (knnirs)
Advanced Member Username: knnirs
Post Number: 23 Registered: 8-2009
| Posted on Wednesday, April 21, 2010 - 6:54 pm: | |
Hi S�verine, If I knew your spectra, it is possible I could help. I would be glad to give my advice if you send me your spectra. [email protected] |
Cesar Guerrero (cesar)
Intermediate Member Username: cesar
Post Number: 17 Registered: 3-2006
| Posted on Wednesday, April 21, 2010 - 7:33 am: | |
Hi S�verine, Have you tried to modify the distribution of the y-term of the matrix (in this case the firmness data)?? Maybe the firmness data follows a normal distribution, but sometimes I have observed that certain pre-processing in the y-term (such as root squares or other exponential transformation) gives better results. I suggest to see the histogram of the y-term, and maybe you can "improve" the distribution. Please note that after calibration you'll need to do a back-transformation to check if you are really improving your calibration. For example, if you are calibrating using the firmness as root square (firmness^0.5), then you'll need to apply ^2 (firmness^2). Of course, the same during prediction stage. Sometimes it works. As disadvantage of this suggestion, the residuals will not remain constant along the range of values (after the back-transformation). You should choose what problem is more important for you. Best regards, C�sar |
S�verine Gabioud Rebeaud (gab)
New member Username: gab
Post Number: 2 Registered: 4-2010
| Posted on Wednesday, April 21, 2010 - 7:04 am: | |
Dear Evgeny, Thanks for your response. It's true that including a lot of variation in calibration data (different origins, years, cultivars...) has to be done in order to obtain a reliable models. But I was anyway wonderig if there is any kind of "non-linear" calibration because I'm not sure that including more variation in my models will solve the problem. I observed this many times for different types of fruit. |
Evgeny (evgeny)
New member Username: evgeny
Post Number: 5 Registered: 3-2010
| Posted on Tuesday, April 20, 2010 - 9:25 am: | |
Hello, Firmness, a texture property, is related to scattering effects and with SNV and second derivative preprocessing a lot of scattering information can be removed from the spectra. I recommend to have a look on the article where 6000 apple samples were used in PLS models (NIR data) and where different effects of cultivars, origins, shelf-life exposure time and seasons were considered. Reference: Els Bobelyn, Anca-Sabina Serban, Mihai Nicu, Jeroen Lammertyn, Bart M. Nicolai, Wouter Saeys, Postharvest quality of apple predicted by NIR-spectroscopy: Study of the effect of biological variability on spectra and model performance, Postharvest Biology and Technology, Volume 55, Issue 3, March 2010, Pages 133-143, ISSN 0925-5214, DOI: 10.1016/j.postharvbio.2009.09.006. |
S�verine Gabioud Rebeaud (gab)
New member Username: gab
Post Number: 1 Registered: 4-2010
| Posted on Tuesday, April 20, 2010 - 7:21 am: | |
Hi everybody, I�m currently working on a project using NIR-technology as a tool for determining fruit quality. I tried to calibrate firmness for different types of fruit using PLS but the models do not fit very well for high values while fitting quite well for low values (the curve looks like a quadratic function). One reason may be that firmness is rather based on physical than chemical characteristics of the fruit. My question is: does anyone know how I could solve this problem? Is there any kind of �non-linear� calibration which has been proven to be useful? Thank you in advance for your help Best regard S�verine |
|