https://www.euro-online.org/enog/inoc2007/Papers/mac-slots.html https://www.euro-online.org/enog/inoc2007/Papers/m https://www.euro-online.org/enog/inoc2007/Papers/mac-slots.html

NIR Model Calibration for Wood Chemistry

sabki's picture
Forums: 

Dear All,
I used XDS RCA Foss NIR for Wood Properties Analyses. When do NIR Model calibration for 1 species of Eucalypt wood chemistry especilly klason lignin and S/G lignin are very difficult to find result for validation R2 above 0.80 (>0.80).
I have 70 references data, 80% for calibration and 20% for validation after that I find 5 samples of outlier and i have exclude outlier data from the model. for pre-treatment used 2nd Derivative, wave length 1100-2500 nm. PLS Regression with Vision software by FOSS
From Total 70 samples ( 51 samples Eucalypt 3 years old and 19 samples Eucalypt 4 years old).
Myquestions :
1. What is the solution to increase validation R2 above 0.80 for this case?
2. Can I do transfer Calibration to unscramble software 10.3?
Many Thanks for helping.
Best Regards,
Sabki

shileyda's picture

Hello Sabki,
Lignin is a difficult assay that is not very reproducible.. Have you submitted blind replicates to the lab to determine the SEL? My guess is that at 0.8 you are running into the limits of the lab method. I would start there. You might try submitting each sample in blind replication over several days, then compare these results and resubmit again if they exceed some threshold. Just remember that if you submit the calibration samples in replicate you will also need to submit the validation samples in replicate.
Good luck,
Dan

sabki's picture

Hello Dan,
Thanks for your suggest, I will consider for submit blind samples to the lab to find SEL and also for independent validation of NIR model. The reference data from lab 2 replicate analyses for every samples. 20 samples It's enough for independent validation (include 5 blind replicate)?
Regards,
Sabki

hlmark's picture

Sabki - blind samples are best sent at different times, and preferably on different days, to maximize the potential variations that  the lab might introduce into their measurement (different standard samples, instrument drift, environmental changes, etc) that would affect their results.
Howard
\o/
/_\
 

ianm's picture

Sabki, I suggest that you verify the reproducibility of the reference method for lignin, and that you determine the reproducibility of the spectral data, which will be affected by the sample preparation method, which, in turn, affects the particle characteristics. In Unscrambler there is a list of spectra that you can import. It is possible to import NSAS spectral data, so it should be possible to import spectral data developed using Vision.

You should get the same order of magnitude for the statistics anyway, no matter what software you use, because they will all be affected to more or less the same extent by the reproducibility of your reference method and your spectral data.

Have a good day,

Phil.
(posted on behalf of Phil Williams by Ian Michael)

jrodrigues's picture

Hello Sabki
Plase check if the range for validation and calibration sets are about the same? What about the errors, are they similar?

shileyda's picture

Sabki,
Depending on what you learn from the blind submission you may need to measure each calibration and validation set in triplicate. Years ago when working with data from a particularly bad reference assay we had to submit each sample for 6 individual analyses. It was a painfully long and expensive process, but in the end a model was developed that would otherwise not have been possible.
When faced with such an assay you can either try to work with the lab so that they discover and reduce the number of factors contributing to the poor SEL, or you can accept that they assay is poor and measure each with multiple submissions of each sample using the average of these analyses OR you can go find a more competent laboratory. No amount of chemometric "magic" will make bad reference data good. Some techniques are better at dealing with reference data noise but the best outcomes are generally from the reference data with the lowest possible analytic noise.
Best regards,
Dan

dangdk's picture

Hi,
What is the range for your lignin content?
Thanks
Dan

dangdk's picture

Hi,
What is the range for your lignin content?
Thanks
Dan

sabki's picture

Hi Dangdk,
The range of lignin content is 27,4% - 38,3% (with standard deviation 2,4).
Regards,
Sabki

dangdk's picture

OK, so a better r2 should be possible, depending on the quality of the reference analysis method. We had a similar range of 11.6% (12.9-24.4%) for KL in Miscanthus and got an R2(pred) of 0.95 and an SEP of 0.634. That is with an SEL of 0.25.
More details here:
http://www.sciencedirect.com/science/article/pii/S0960852412009017