Reducing high-dimensional feature set of hyperspectral measurements for plant phenotype classification

Abstract

Quantifying the drought resistance of potato cultivars plays a key role in precision agriculture, and it may lead to the development of new varieties of plants that are more resistant to harsh environmental conditions. In this work, we tackle the issue of extracting such information in a non-invasive way by acquiring in-field hyperspectral measurements of the potato leaves. Then, we exploit an array of machine learning models to classify plants into three wilting classes based on such data, with those classes corresponding to their drought resistance. We show that evolutionary band selection can dramatically reduce the dimensionality of hyperspectral data while improving classification accuracy. Our experimental study revealed that the evolutionarily-optimized models offer high-quality performance with the impartial rϕ reaching 0.784, accuracy: 0.867, and a 30% improvement over the baseline models which do not benefit from band selection.