Anther size quantification in wheat using deep learning under normal and heat stress conditions

Abstract

Key message We present an imaging-based deep learning phenotyping pipeline that classifies heat-stressed wheat anthers and quantifies size traits using YOLO, enabling fast, precise, scalable measurements to support breeding heat-resilient wheat varieties. Abstract Terminal heat stress is a major abiotic stress causing significant yield loss in wheat. Anther size, a key trait of terminal heat stress tolerance, is least studied in wheat due to complex and tedious scaling trait, and short live span. To provide user friendly approach to plant breeders, integration of the modern digital imaging and deep learning techniques together is the current need of high-throughput phenomic era. In this study, we introduced a hybrid approach that amalgamates the strengths of deep learning models for both binary classification and precise morphological analysis of anther images of 177 wheat accessions under normal and heat stress environments. ResNet18 with 94% accuracy, outperformed the traditional models like CNN and MobileNetV2, achieving high classification performance. For morphological trait extraction, we employed YOLOv8, a cutting-edge object detection model known for its high speed, accuracy, and computational efficiency. YOLOv8 successfully localized anthers and measured width and length with strong agreement to experimental measurements, as validated by Bland–Altman analysis. Its precise detection capability and lightweight architecture makes it ideal for high-throughput phenotyping using digital imaging approach. To further boost the interpretability of our deep learning models, we utilized Grad-CAM, a powerful technique for visualizing class-specific features in the network’s decision-making process. This facilitated in categorizing the key visual features within the anther images that had the greatest influence on the model’s decision-making process. This cohesive workflow not only sets a new benchmark in image-based classification and morphological measurement but also proposes an accessible tool for rapid, real-time phenotyping, supporting data-driven breeding strategies aimed at improving wheat resilience under terminal heat stress.

Plant phenotyping relevance

コムギ葯の画像取得、深層学習による分類・形態形質抽出、実測値との技術検証を中心とする明確な植物フェノタイピング手法研究。

Code and data availability

The paper describes a deep learning phenotyping pipeline (CNN/ResNet18/MobileNetV2, YOLOv8, Grad-CAM) on anther images of 177 wheat accessions, but no public dataset, image repository, code, or trained model is deposited. The only availability statement is 'Data is provided within the manuscript or supplementary info,'

No evidence-backed public reproduction asset is currently recorded.