Affordable Phenotyping at the Edge for High-Throughput Detection of Hypersensitive Reaction Involving Cotyledon Loss.

Abstract

The use of low-cost depth imaging sensors is investigated to automate plant pathology tests. Spatial evolution is explored to discriminate plant resistance through the hypersensitive reaction involving cotyledon loss. A high temporal frame rate and a protocol operating with batches of plants enable to compensate for the low spatial resolution of depth cameras. Despite the high density of plants, a spatial drop of the depth is observed when the cotyledon loss occurs. We introduce a small and simple spatiotemporal feature space which is shown to carry enough information to automate the discrimination between batches of resistant (loss of cotyledons) and susceptible plants (no loss of cotyledons) with 97% accuracy and with a timing 30 times faster than for human annotation. The robustness of the method-in terms of density of plants in the batch and possible internal batch desynchronization-is assessed successfully with hundreds of varieties of Pepper in various environments. A study on the generalizability of the method suggests that it can be extended to other pathosystems and also to segregating plants, i.e., intermediate state with batches composed of resistant and susceptible plants. The imaging system developed, combined with the feature extraction method and classification model, provides a full pipeline with unequaled throughput and cost efficiency by comparison with the state-of-the-art one. This system can be deployed as a decision-support tool but is also compatible with a standalone technology where computation is done at the edge in real time.

Plant phenotyping relevance

低コスト深度画像による植物病害表現型（子葉脱落・抵抗性）の自動抽出、分類パイプラインを開発し、多数品種・環境で性能と頑健性を検証しており、表現型取得手法が中心である。

Code and data availability

The paper describes a pepper-TSWV depth/RGB image dataset (208 resistant and 50 susceptible batches) and a classification pipeline, but no public deposit of the dataset, images, or analysis code is stated in the supplied blocks. The only code availability mention refers to a prior book chapter [54] for acquisition pil,

No evidence-backed public reproduction asset is currently recorded.