I had the opportunity today to check the performance of a calibration (moisture in intact sunflower seed in reflectance).
This is always a exciting moment:
Does the performance of the calibration for the new validation set is as expected during the calibration development?.
Can I add with the new set more variability to the calibration and improve it?
Sunflower intact seed is not an easy product to analyze by NIR, especially for fat. Much better improvements can be getting grounding the sample, but the option to eliminate the grinding process is of course very attractive.
I use some functions I have created with R software for the validation (see previous posts and videos about Monitor function) and to create some comments about the results.>monitor15(sflwseed$Moi_IX,sflwseed$Moi_Lab,2046,12,0.95,0.60)
Nº Validation Samples = 1298
Nº Calibration Samples = 2046
Nº Calibration Terms = 12
Calibration SECV = 0.6
RMSEP : 0.6473
Bias : 0.4687
SEP : 0.4466
UECLs : 0.6252
***SEP is bellow BCLs (O.K)***
Corr : 0.8977
RSQ : 0.8058
Slope : 0.8975
RER : 16.1 Fair
RPD : 2.21 Very Poor
***Bias adjustment is recommended***
Residual Std Dev is : 0.4351
***Slope adjustment is recommended***
This is a quite big validation set, see the plots:
As we can see the residuals histogram has a normal distribution.
RMSEP is quite similar to the Standard error of Cross Validation ( a little bit higher), and the SEP (Validation Error corrected by the Bias) is much better (an improvement of 0,2). So the Bias adjustment is recommended. There is a deviation of the slope, but the plots show that this should be treated with caution.
Anyway this is a long validation set, and the best option is to merge this validation data with the calibration data and recalibrate in order to improve the error and make a more robust calibration which probably will improve the statistics for the SECV and for the RMSEP in the next validation.