Loblolly pine is the backbone of the southeastern US forestry industry – but its productivity is constantly threatened by fusiform rust, a complex fungal disease whose early symptoms often go unnoticed. Traditional approaches to identifying resistant trees rely on time-consuming genetic and physiological assessments. But could vibrational spectroscopy offer a faster, more holistic alternative?
To find out, researchers turned to handheld near-infrared (NIR) and benchtop FT-IR spectroscopy to capture the “chemical fingerprint” of resistant and susceptible pines. Despite pandemic-related delays, challenging field logistics, and less-than-ideal sampling conditions, the team uncovered surprisingly robust spectral indicators of resistance – and even found that the simpler NIR tool outperformed Fourier transform infrared spectroscopy (FT-IR) in their models.
We spoke with lead author and Assistant Research Professor at North Carolina State University, USA, Simone Lim-Hing, about the inspiration behind the project, the analytical challenges involved, the unexpected results, and how spectroscopy, chemometrics, and machine learning could help foster a new era of precision forestry.
What inspired this research?
Identifying resistance to disease in forest tree systems is quite complicated, since their resistance mechanism is often unclear. In many of these systems, we see a defensive response expressed as a suite of genetic responses and traits, particularly with chemical signaling. However, characterizing each and every chemical signal (i.e., secondary metabolites, hormones, etc.) is time-consuming and costly and, at the end of the day, may not be fruitful since it’s so targeted. We wanted to use something that could take a snapshot of a sample, analyze it in its totality, and relate it back to its response to disease. There were a few studies that had used vibrational spectroscopy to categorize disease resistance in trees, and they had promising results. We therefore wanted to apply it to loblolly pine and fusiform rust disease, which are both very economically important in the southeastern US.
What was the biggest challenge you had to overcome?
As with many investigations in recent years, the pandemic really grounded everything to a halt. The portable spectroscopy tool we used in this research was a brand-new product, but the supply chain issues during and after the height of the pandemic really altered the original plan. Once we finally got the device, we had to get moving quickly and coordinate site visits since they were all on private land. We wanted to collect spectra at around a similar time frame since we know that the chemical profiles change depending on the season. However, time constraints pushed us to collect in February, which isn’t ideal as trees are physiologically in flux – they are coming out of dormancy and getting ready to set seed. We definitely see these phenological influences in the results, but nevertheless we got pretty robust profiles associated with resistance.
Results in Brief
Across eight field sites, phloem and needle samples were scanned to capture each tree’s chemical fingerprint, and the resulting spectra were modeled using machine-learning approaches including SVM and sPLS-DA. The top NIR model – built using spectra from the 30 most resistant and 30 most susceptible trees – achieved 81.5 percent training accuracy and 68.7 percent testing accuracy, outperforming FT-IR models, which reached up to 65 percent testing accuracy. Phloem tissue provided clearer discrimination than needles, and several NIR bands repeatedly emerged as markers of resistance-associated chemistry.
How could this approach change how scientists and foresters assess plant health or disease resistance in the future?
There is so much potential in using spectroscopy-based tools in plant health and it is definitely gaining some momentum in the field. As these tools decrease in cost and increase in accessibility, it will provide a more efficient and biologically holistic approach. Identifying disease resistance can be so tricky, especially in forest trees, and adding a layer of confirmation through tools like these will be important for breeding. Additionally, the tool we use is handheld, user-friendly, and pairs with cloud-based software that allows you to upload your own algorithms and models. So, the research-to-application pipeline is quite smooth.
What would it take to bring this technology into widespread use?
I think that more research with this particular tool and more collaborations with the forestry sector will push this approach forward into widespread use. There are a few ongoing projects that I am a part of using this same framework. The results are promising and the interest from companies is still there.
My current research has shifted to focus more on larger-scale forest health monitoring. We are seeing some research come out where they are using drones with hyperspectral sensors and satellite data to pick up signals associated with stressed trees.
How do you see the relationship between the chemometrics and AI fields evolving?
The spectral data you’re working with is only as good as the modeling approach. There are so many ways you can analyze this type of data, and I’ve learned it can be quite easy to skew the results in your favor (which is why we opted to use a bootstrapping sampling approach for our predictive accuracies). I think the rise of AI and machine learning can be a double-edged sword. On one hand, they are incredibly powerful tools that can pick up minute patterns in spectra, resulting in a successful model. On the other, it can be a black box that obscures the biological relevancy of the results. Nevertheless, I think chemometrics has become way more accessible, which is a big step in using these types of tools.
Newsletters
Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.
