A High‐Throughput Phenotyping Pipeline for Quinoa ( Chenopodium quinoa ) Panicles Using Image Analysis With Convolutional Neural Networks

Abstract

ABSTRACT Quinoa is a grain crop with excellent nutritional properties that has attracted global attention for its potential contribution to future food security in a changing climate. Despite its long history of cultivation, quinoa has been improved little by modern breeding and is a niche crop outside its native cultivation area. Grain yield is strongly affected by panicle traits, whose phenotypic analysis is time consuming and prone to error because of their complex architecture, and automated image analysis is an efficient alternative. We designed a panicle phenotyping pipeline implemented in Python via mask R‐convolutional neural networks for panicle segmentation and classification. After model training, we analysed 5151 images of quinoa panicles collected over three consecutive seasons from a breeding programme in the Peruvian highlands. The pipeline follows a stagewise approach, which first selects the optimal segmentation model and then another model that best classifies panicle shape. The best segmentation model achieved a mean average precision (mAP) score of 83.16 and successfully extracted the panicle length, width, area and RGB values. The classification model achieved 95% prediction accuracy for the amarantiform and glomerulate panicle types. A comparison with manual trait measurements using ImageJ revealed a high correlation for panicle traits (r > 0.94, p < 0.001). We used the pipeline with images from multilocation trials to estimate genetic variance components of an index on the basis of panicle length and width. We further updated the model for images that included metric scales taken in field trials to extract metric measurements of panicle traits. Our pipeline enables accurate and cost‐effective phenotyping of quinoa panicles. Using automated phenotyping based on deep learning, optimal panicle ideotypes can be selected in quinoa breeding and improve the competitiveness of this underutilized crop.

Plant phenotyping relevance

キヌア穂の画像取得・セグメンテーション・形状分類・形質抽出を中核とする高スループット表現型解析パイプラインを開発し、手動測定との相関で検証しているため。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.