| Author | Message | ||
     Scott Parsons (aussiecologist) Junior Member Username: aussiecologist Post Number: 6 Registered: 4-2007 |
Pierre, many thanks for your great advice! I will take this info back to the drawing board  | ||
| Pierre Dardenne (dardenne) Senior Member Username: dardenne Post Number: 35 Registered: 3-2002 |
Hi Scott, The first thing to know before starting a NIR calibration is the SEL (Standard error of the laboratory method). The SEL can be the repeatability error but I prefer the reproducibility error (if you have more than one lab, one can use the SEL of ring tests). Knowing the SD (standard deviation) of the reference values of the population (i.e. all the soils in your area) and knowing that R2=(SD^2-SEC^2)/SD^2 , if you replace SEC by SEL (you consider the NIR error =0), you have already the maximum R2 you can expect before starting modeling. Of course your model statistics (SEC-SECV) will never be better than the SEL. It is always more efficient to obtain very precise reference values to calibrate. But this has a cost. Generally I prefer 2n samples analyzed in single than n samples analyzed in duplicates. The final test set must be very precise and deserves replicates for the wet chemistry. Anyway link the spectra with the more precise ref values (average X and Y). If you scan the sample more than once, do the calibration with the averaged spectra otherwise you will have problems with the cross validation (with WinISI, other software take care of that). With WinISI, an efficient trick is the use of the duplicated or replicated scans within the repfile. Pierre | ||
| Scott Parsons (aussiecologist) New member Username: aussiecologist Post Number: 5 Registered: 4-2007 |
Hi Pierre, Many thanks for your response! I have additional spectra that I can add to the calibration set that may help with the cross validation problems, although further wet chem is needed on these. In addition to what you say it does also appear that I may need to reassess the laboratory values I have used for my spectra. For the soils there is obviously some variability between repeated wet chem readings (and the repeated NIR scans), which I may or may not have quantified adequately. Is there an accepted way of using this variability in the calibrations/validations? Maybe by testing cross validation on the different wet chem values representing the variability in the soil matrix. Or is the best option to run the NIR analysis with the means of the wet chem and the spectra performing cv therein? Scott | ||
| Pierre Dardenne (dardenne) Senior Member Username: dardenne Post Number: 33 Registered: 3-2002 |
Scott, Your observations are common : when starting an application the difference between sec and secv is generally quite large. It means that the calibration set (when some samples are removed) is unable to cover the whole variation. Then you need more samples (each sample does not have enough neighbors). I would not spend more than 2 hours to do the first models. I would use these models on new samples and retained those with NH>1 (with series of 5 or 10 depending of the resources), do the wet chemistry, add the spectra to the data base and restart the process of calibration, scan new independent samples and repeat the loop until 95% of the new samples show NH<1. Dr Shenk set the limit for the NH at 0.6: this is a good limit but requiring more samples. I suggest you to use the NH based on the PLS scores (PL1 instead of PCA). The LOCAL procedure gave good results on soils but it also requires larger data sets. RPDc = SD/SEC or =1/SQRT(1-R2) thus the same conclusions must be extracted from R2 vs 1-VR and SEC vs SECV. Pierre | ||
| Scott Parsons (aussiecologist) New member Username: aussiecologist Post Number: 4 Registered: 4-2007 |
Hello, I am postgrad doing a study attempting to calibrate spectra taken from range of soils with variables such as nutrients and texture. I'm fairly new to NIRS. Generally speaking I am getting good Rsq coefficient of determinations for variables you'd you'd expect to work, but substantially lower 1-vr coefficient of determinations in cross validation. For instance, for one variable I have an rsq of around 80%, but a 1-vr of ~60%. However, my SEC and SECV are all pretty close. Additionally, if I take the rsq/SEC and 1-vr/SECV, the latter are generally less than 2x 1-vr/SECV. I'm using WINISI II (developing cals with the full spectrum) and I have about 70 samples in the calibration. My method has been following the recommendations in the ISI help re: the number of cross validation groups and number of terms. i.e. it says: "if the calibration contains 50 or more samples 4 cv groups should be adequate", and the no of terms is the no of samples divided by 10 + 2 or 3, so I've used 9. The RPD values are reasonable from what the literature has told me for soil. The problem seems to be in cross validation. Upping the number of cv groups does not seem to make any great difference to the 1-vr, so but I am stumped for ways to better this. I am even confused if the rsq and 1-vr difference denotes great problems with the models when the SEC and SECV values are close. Can anyone say if this suggests a great problem with my data, or suggest a different approach to better this that might be applicable using ISI? Many thanks to a great resource |