A New Vibrational Analysis with Near-Infrared Spectroscopy using Supervised Machine Learning

Previous attempts using near-infrared techniques to spatially characterize chemical composition of plant biomass perform well within defined ecosystem or plant type but fail to address key challenges of accurate prediction and quantification of the underlying fundamental vibrational signatures.

In this report, we use a supervised machine learning method that analyzes the fundamental relationships within the infrared spectroscopy to create a robust predictive model to generate mid-infrared (MIR, 4000-400 cm^-1) spectra using solely the near-infrared (NIR, 12800-4000 cm^-1) spectral data. The mid-IR spectra contain more distinctive and unique signatures related to the strong fundamental vibrational modes that have been widely used to classify and quantify biochemical components of lignocellulosic biomass. However, these measurements are not suitable for characterizing chemical compositions with high spatial resolution.

Specifically, we collected near-infrared and mid-infrared spectra on senesced leaves in a mountainous mixed shrub system in Crested Butte, Colorado, and then trained a partial least squares regression model (PLSR) with mid infrared and near infrared spectrum to predict mid-infrared spectra using only the observed near-infrared spectra. We then used these predicted mid-infrared spectra to classify species composition of these samples, and further quantify the relative concentrations of macronutrients in the leaves which provide important sources in Nitrogen/Phosphorus/Carbon cycling. Using spatial near-infrared data in another system collected by the Airborne visible/infrared imaging spectrometer (AVIRIS) we were able to recreate spatial concentrations of these compounds. We validated our results from the original leaves collected from the ground with direct mid-infrared measurements, and together they revealed the accuracy of the model.

We hope to continue to develop this method as a consistent tool to characterize landscapes using remote near-infrared spectral data.