Grass is greener for NIR spectroscopy

A field of grass may just look like grass, but it is obviously made of up of lots of different species of grass and other plants all growing together. Determining exactly which species are present has tended to be a fairly time-consuming exercise, requiring each species in a sample to be hand separated and then identified.

NIR spectroscopy potentially offers a much faster way to do the same thing, in which case the grass sample is dried, ground up and then analysed. But when scientists have tried this, they’ve found that it doesn’t work as well as they expected.

The problem is that the first step in this process involves producing a model relating spectral information to specific grass species. One way to do this is to separate a grass sample into its constituent species and identify them, before mixing the species up again, drying them, grinding them up and then analysing them with NIR spectroscopy, but obviously this is quite time-consuming.

An easier way is to create an artificial grass sample by mixing together known species and then drying them, grinding them up and analysing them with NIR spectroscopy, which does away with the time-consuming separation and identification step required for a real sample. You can even mix the spectra rather than the actual species, meaning that you analyse individual plant species by NIR spectroscopy and then simply mix the spectra together.

Unfortunately, the models produced from these artificial samples are far less accurate when tested on real samples than those produced from real samples, even though the samples contain exactly the same mix of species. To find out why this is the case, a team of Belgian plant scientists, led by Mathias Cougnon at Gent University, compared NIR spectra producing by analysing real and artificial grass samples made up of three plant species: the grass species tall fescue and ryegrass, and white clover.

As they report in Grass and Forage Science, it turns out that the spectra from the real samples contain much more variation than the spectra from the artificial samples. As a result, the models derived from the artificial samples are based on a much more limited set of spectra, explaining why they are less accurate. Interestingly, though, Cougnon and his team found that they could increase the accuracy of these models by simply determining the difference between the spectra produced by the species in real and artificial samples and including this difference in the models.

What’s still not clear is why the spectra are so different between the real and artificial samples, seeing as they contain exactly the same mix of plant species. On explanation could be that plants have a more varied chemical composition when grown in a mixture with other species than when grown as a single, isolated species.

Blog tags: