We have seen in a previous post how to get the best configuration for a local calibration (Min and Max number of samples and factors). See: LOCAL Equations with Win ISI 4 (Part: 1)
We have a RED file which has the math-treatments applied, and a reduction of the number of data points if necessary. This RED file can have also a certain range of wavelengths (NIR, NIR + VIS,..., or selected ranges into these areas). The better prepared the RED file, better and faster results we will get.
Of course the predictions will change depending of the RED file, so it is important to check with trial and error the best configuration for this file before to implement it in routine.
Hi José Ramón,
ResponderEliminarCongratulations for your useful post about local calibrations.
After watching your video, I have found a similarity with my own experiences, as the number of samples increases the SECV and SEP decreases. The question is, when do we stop the process???
Hi Antonio,
EliminarIn case of a good database, normally when the number of sample increases the SEP decrease. But of course there is a limit. Local select the samples based in a special correlation method and as soon as the samples start to have lower correlation, the SEP will be worse. The software can give the option to configure a correlation cutoff in order to be sure that we don´t enter in the samples selected samples not similar to the unknown.
For the number of terms LOCAL does not use the cross validation, LOCAL uses all the terms you configure in a sequence:
Prediction with 1 (model 1), with 1 and 2(model 2), with 1, 2 and 3 (model 3),……….But it use a weighted average of all the predicted values generated with them.
The weigh will depend of how well the different models builds the unknown spectrum, if the residual is large they will have a weigh close to cero and their prediction is not considered. So the final prediction is the average of the models which build the unknown spectrum better.