Unlocking Comparative Plant Scoring with Siamese Neural Networks and Pairwise Pseudo Labelling

Phenotypic assessment of plants for herbicide discovery is a complex visual task and involves the comparison of a non-treated plant to those treated with herbicides to assign a phytotoxicity score. It is often subjective and difficult to quantify by human observers. Employing novel computer vision approaches using neural networks in order to be non-subjective and truly quantitative offers advantages for data quality, leading to improved decision making. In this paper we present a deep learning approach for comparative plant assessment using Siamese neural networks, an architecture that takes pairs of images as inputs, and we overcome the hurdles of data collection by proposing a novel pseudo-labelling approach for combining different pairs of input images. We demonstrate a high level of accuracy with this method, comparable to human scoring, and present a series of experiments grading Amaranthus retroflexus weeds using our trained model.