PCR vs. PLS (part 10) Meat Meal

In an application for meat meal I proceed the same way as in the case of the soy meal in the Infratec. This time the instrument is a DA1650 NIR with a wavelength range from 1100 to 1650 nm. The spectra was math treated with SNV, Detrend and first derivative with data points every 2 nm. The parameter interested in for this post is the Ash.

 

Two calibration approaches had been done with the PCR and the PLS regressions. Spectra has been sorted by Ash reference value, so we can see in this plot ll the samples treated with SNV-DT and first derivative.

 

matplot(wavelength,t(NIR),type="l",
        ylab="Reflectance",
        xlab="Wavelength",
        col=AshLev)
abline(0,0)
legend("bottomleft",legend=c("0-10% Ash",
       "10-20% Ash","20-30% Ash","30-40% Ash"),
       col=c("blue","green","orange","red"),
       lty=c(1,1), cex=1)

I split the samples into odd and even , the same than in other examples and keep the even outside for external validation. The odd samples were used to develop the PCR and PLS Regressions with CV using 10 groups. Notice that the external validation with the even samples is like if we would use a CV with 2 groups.

 

Cross Validation (CV) help us to decide how many terms we would use for the development of the regression looking to the CV RMSEP graph, and in this case the graph is:

 

plot(mm_da_odd_pcr,"validation",estimate="CV",

     ylim=c(0.5,6.0),xlim=c(1,30),col="red")
par(new=TRUE)
plot(mm_da_odd_pls,"validation",estimate="CV",

     ylim=c(0.5,6.0),xlim=c(1,30),col="blue")
grid(nx = 30, ny = 6, col = "lightgray", lty = "dotted",
     lwd = par("lwd"), equilogs = TRUE)
legend("bottomleft",legend=c("PCR RMSEP","PLS RMSEP"),
       col=c("red","blue"),lty=c(1,1),cex=1)

The graph is really confuse in this case for the PCR, and we get worse results than for the PLSR. In the case of the PLSR, we see that 15 terms can be enough, but we can try with less  (10 for example) with the even validation set.

 

As we can see 15t works better than 10 for the even samples.

I have to say that in these sets there are samples from different origins, labs,...etc., and outliers have not been removed, so there is room for improvement.