Hi all,
I have 1000 spectra of different geological materials. These spectra contain reflectance value for each 1nm ranging from 350nm to 2500nm. I want to train my neural network with these data as input and their corresponding label (name of the mineral/rock) as output. Please suggest me how to extract various features from spectra which can be used to define each mineral/rock?
Dear Mr. Dhamotharan,
Your spectrum covers the VIS spectrum (with a small incursion into UV) and part of the IR spectrum. First and foremost it might be possible to use this spectrum to some extent for the identification of some minerals but not for aggregates of them, called rocks. You need to investigate these aggregates by quantity and quality and as such simple spectra do not comply with your demands. As far as the minerals are concerned I suggest having a closer look to non-destructive methods, particularly those used in gemology. There are some tables showing besides optical data also spectra which can be useful for you.
With kind regards
H.G.Dill
Hi Sundara,
If the spectra from the different materials possess specific signal peaks or maximums that seem to be recurrent, I'd attempt a model where you select those signals, as they might be related to the composition of your material. You could try two models: using the value of the signal at the maximum of these bands as inputs, or calculate their areas under the curve. In the following paper of my research team, this approach was carried out (you can find it in my RG profile):
Aroca-Santos et al. Identifying and Quantifying Adulterants in Extra Virgin Olive Oil of the Picual Varietal by Absorption Spectroscopy and Nonlinear Modeling. Journal of Agricultural and Food Chemistry, 2015.
On the other hand, you could employ several feature selection methods, such as Fisher's discriminant ratio, information gain test, or one that I have seen to present a good synergy with neural networks called Relief-F algorithm. This way, you can include all of your signals from 350 nm to 2500 nm, and these feature selection algorithms will rank the independent variables, according to different mathematical criteria, suggesting the best variables for a successive model (e.g. neural network). This was done in the following work (also in my RG profile):
Shehada et al. Silicon Nanowire Sensors Enable Diagnosis of Patients via Exhaled Breath. ACS Nano, 2016.
Hope this helps and best of luck,
John
Hi!
It's difficult to answer this without knowing what kind of materials you want to analyze. Generally, it probably makes sense to look at the reflection spectra and deconvolute them to identify recurring signal peaks. Mr. Cancilla already made some very good suggestions as to how to approach this. However, while this approach might enable you to identify single key components / minerals (at least those showing distinctive peaks), I don't think the reflection spectrum alone contains enough information to actually name a rock. What type of rock you're dealing with not only depends on the mineral composition but also largely on things like grain size and texture of the material. Maybe you'd want to combine your approach with additional textural information (segmentation), so your neural network can learn whether it's dealing with, e.g., a plutonic or metamorphic rock with identical mineral composition.
Good luck with your project!
Ricarda
You need a reliable spectral library representative of rock-forming minerals including variations thereof. One starting point could be the USGS spectral library freely available.
regards
Brian
- Badges
- Science topic