Is the method 'partial least squares' too much of a black box to be accepted by the remote sensing community? Similar to neural networks, it has great promise but as it is not physically based, may not be appropriate for creating new products and tools. How transferable are the results across space and time?
PLS is meant to deal with large predictor dataset, so it has been used in hyperspectral RS. I specially rememeber some works by Lin Li and Susan Austin. I would not call it a black box, as you can go as far in the algorithm as you wish. I however experienced once that shifting the order of predictors may provide a differing result. Perhaps, the problem may be that algorithms usually have a starting point from which to iterate. I would suggest to then test convergence to a solution. About not being physically based....depends on the objectives of your research. Your statement is right if you just use it for prediction. PLS may nevertheless be used in search cause-effect patterns underneath may effects or noisy data (not uncommon in RS). You may see an example of deducting physical relations from PLS results in (oncoming) Valbuena et al. (2012) Forest Science. My advice is to nevertheless do the predictions out from PLS later on; they usually overfit.
Ruben, I will definitely want to see your paper.
I have worked with Lin Li and have a paper published in SPIE Proceedings that compared models in estimating chlorophyll and phycocyanin (cyanobacterial pigments) using PLS with empirical models. We found that PLS can bring an improvement over simple band ratio models; however, it fairs no better than models that have taken into consideration the behavior of reflectance (e.g. pigment absorption, cell scattering). The problem with PLS is that few studies fully explain the PLS weights (especially the 2nd and 3rd weights), which generally are related to some physical basis. To me, the PLS weights should be the highlight of these studies because they provide insight into which spectral bands might be the best in simpler empirical models.
erratum on my comment: "may effects" meant "main effects"
I absolutely agree with the comment above. That was exactly what we did. We even found one same parameter showing opposite effects (which actually were against our own prior hypothesis). This is how PLS can assit in improving physical models. If the right person is in place to interpret the 2nd and 3rd weights, there are great possibilities to deal with physically complex situations.
The problem is that the results from the weights require the researcher to have deep knowledge on the physics involved, in order to infer these relations. Otherwise the 2nd weights may just look messy. I gave a try once with some hyperspectral data in seach for soil components, finding virtually nothing. I finally deducted that I lacked of expertise in that certain matter, or maybe the data was truly too noisy.
Hi
With regards to the spectral bands that provide insight to the model, you might consider calculating VIP (Variable Importance in the Projection) scores. for a R implementation have a look at Mevik, B.-H., Wehrens, R. (2007); The pls Package: Principal Component and Partial Least Squares Regression in R; Journal of Statistical Software 18(2), 1—24. Also the paper by "Chong, Il-Gyo & Jun, Chi-Hyuck, 2005, Performance of some variable selection methods when multicollinearity is present, Chemometrics and Intelligent Laboratory Systems 78, 103–112" might be of interest
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing