Selecting informative spectral bands using machine learning techniques to detect Fusarium head blight in wheat

Fusarium head blight (FHB, also known as scab), is among the most widespread and devastating diseases attacking small grain crops, mainly wheat and barley. Besides its negative impact on yield and grain quality, the FHB pathogens produce mycotoxins, such as deoxynivalenol (DON), that contaminate the grain, making it unsuitable for human and livestock consumption. Development of FHB resistant cultivars through breeding is the most effective solution in response to the FHB threat. The bottleneck of FHB breeding for disease resistance is the visual assessment of spikes, which is extremely demanding, laborious, time-consuming, and most importantly subject to human error. The objective of this study was to identify the most informative spectral bands to develop an automated phenotyping framework for discriminating FHB infected spikes from healthy ones. For this purpose, hyperspectral images from a known susceptible wheat variety, Wheaton, were captured in a laboratory where we could control lighting, background, and the orientation of spikes with respect to the imager. Subsequent to the required preprocessing, spectral bands were ranked by an ensemble feature selection pipeline. Afterwards, a density estimator with Gaussian kernel was used to cluster the most informative bands and find the center of broad multispectral bands. Kernel estimator identified five bands, including 766, 868, 696, 591, and 540 nm (listed in order) as the cluster centers to develop a multispectral sensor. The classification results using the selected band revealed that two spectral bands are sufficient to achieve the maximum classification accuracy, demonstrating the redundancy and high correlation between hyperspectral bands for this application.