Detecting the attack of the fall armyworm (Spodoptera frugiperda) in cotton plants with machine learning and spectral measurements

The Spodoptera frugiperda (i.e., fall armyworm) causes irreversible damage in cotton cultivars, and its visual inspection on plants is a burdensome task for humans. A recent strategy to automatically do similar tasks is processing hyperspectral reflectance measurements with machine learning algorithms. Herein, its proposed a framework for modeling the spectral response of cotton plants under the fall armyworm attacks using machine learning algorithms, culminating in a theoretical model creation based on the band simulation process. A controlled experiment was conducted to collect hyperspectral radiance measurements from health and damage cotton plants over eight days. A hand-held spectroradiometer operating from 350 to 2500 nm was used. Several algorithms were evaluated, and a ranking approach was adopted to identify the most contributive wavelengths for detecting the damage. The Self-Organizing Map method was applied to organize the spectral wavelengths into groups, favoring the theoretical model creation for two sensors: OLI (Landsat-8) and MSI (Sentinel-2). It was found that the Random Forest algorithm produced the most suitable model, and the last day of analysis was better to separate healthy and damaged plants (F-measure: 0.912). The best spectral regions range from the red to near-infrared (650 to 1350 nm) and the shortwave infrared (1570 to 1640 nm). The theoretical model returned accurate results using both sensors (OLI, F-Measure = 0.865, and MSI, F-Measure = 0.886). In conclusion, the proposed framework contributes to accurately identifying cotton plants under the Spodoptera frugiperda attack for both hyperspectral and multispectral scales.

Automated summary

This study proposes a framework for modeling the spectral response of cotton plants under fall armyworm attacks using machine learning algorithms. Through organizing spectral wavelengths, the Random Forest algorithm was found to be the most suitable model. The best spectral regions range from 650 to 1350 nm and 1570 to 1640 nm.