Classification of Plant Endogenous States Using Machine Learning-Derived Agricultural Indices

Abstract

Leaf color patterns vary depending on leaf age, pathogen infection, and environmental and nutritional stresses; thus, they are widely used to diagnose plant health statuses in agricultural fields. The visible-near infrared-shortwave infrared (VIS-NIR-SWIR) sensor measures the leaf color pattern from a wide spectral range with high spectral resolution. However, spectral information has only been employed to understand general plant health statuses (e.g., vegetation index) or phytopigment contents, rather than pinpointing defects of specific metabolic or signaling pathways in plants. Here, we report feature engineering and machine learning methods that utilize VIS-NIR-SWIR leaf reflectance for robust plant health diagnostics, pinpointing physiological alterations associated with the stress hormone, abscisic acid (ABA). Leaf reflectance spectra of wild-type, ABA2-overexpression, and deficient plants were collected under watered and drought conditions. Drought- and ABA-associated normalized reflectance indices (NRIs) were screened from all possible pairs of wavelength bands. Drought associated NRIs showed only a partial overlap with those related to ABA deficiency, but more NRIs were associated with drought due to additional spectral changes within the NIR wavelength range. Interpretable support vector machine classifiers built with 20 NRIs predicted treatment or genotype groups with an accuracy greater than those with conventional vegetation indices. Major selected NRIs were independent from leaf water content and chlorophyll content, 2 well-characterized physiological changes under drought. The screening of NRIs, streamlined with the development of simple classifiers, serves as the most efficient means of detecting reflectance bands that are highly relevant to characteristics of interest.