Machine Learning Meets Distinctness in Variety Testing

Abstract

Distinctness is a binary trait used in variety testing to determine if a new plant variety can be considered distinct or not from a set of already existing varieties. Currently distinctness is mostly based on human visual perception. This communication considers distinctness with a machine learning perspective where distinctness is evaluated through an identification process based on information extraction from machine vision. Illustrations are provided on apple variety testing to perform distinctness based on color. An automated pipeline of image acquisition, processing and supervised learning is proposed. A feature space based on the 3D color histogram of a set of apples is built. This feature space is built using optimal transport, fractal dimension, mutual entropy and fractional anisotropy and it provides results in accordance with human expertise when applied to a set of varieties highly contrasted in color and another one with low color contrast. These results open new research directions for achieving higher-throughput, higher reproducibility and higher statistical confidence in variety testing.