Hi Hanno,

I am not familiar with those acronyms in the context of this type of analysis. What are MLR, PLS or PCR?

Hi Hanno,

Thanks for the acronym clarification.

I actually tried doing a PCA of soil samples in spectral space. I then used those PCA axes as predictors of the soil attributes in a multiple regression, where the the laboratory measured values for soil chemical attributes were the dependent variables. However, the PCA axes (I used 6 axes from the PCA, despite the fact that axes 4-6 explained little of the variation) did not produce as well-fitting model as when I let the forward variable selection select amongst the raw, spectral predictors (eg. reflectance). I interpreted this as the spectral features associated with each soil attribute being very narrow. The PCA dilutes this small scale signal. I suppose I could also try other forms of multivariate data reduction but PCA would be the logical choice as reflectance between nearby 10 nm bands are strongly linearly correlated.

So, my problem of how to resample the continuous function of reflectance as a function of wavelength still remains. I made semivarigrams of reflectance, which showed no sill where the correlation between reflectance values of nearby bands obviously dropped off. The semivariogram just continued upwards in a monotonic fashion.  Based on the semivariograms, I tried using reflectance at 100 nm intervals as predictors in the multiple regression but there was no improvement in the model fit. Again, I think that there could be finer scale features lost when I resample at courses spectral scales.

I was left thinking that if one wants to capture the multiple scales of a spectral signature to use as predictors of a soil sample, one should use multiple sets of reflectance values (eg. bandwidths) in a single predictor matrix.