Transformer vs CNN: A Comparative Study of Wheat Spike Detection Under Field Conditions

Abstract

Wheat spike detection plays an important role in phenotyping, as it provides an approach for direct yield estimation and serves as an indicator of yield potential. Traditional method of phenotyping involved manual counting of wheat heads which is a time-consuming, and error-prone process. However, the advent of Deep Learning (DL) techniques has revolutionized this process by allowing automated detection and counting of wheat heads using high-resolution imagery, hence facilitating large-scale, High Throughput Phenotyping (HTP) of wheat. Despite technological advancements, issues related to environmental variability, differences among cultivars, and overlapping heads continue to make automated detection task difficult and error-prone. To address these issues, researchers have focused on enhancing the robustness of DL models and increasing the diversity of wheat head datasets to improve detection accuracy and reliability of this approach. In this study, we have compared the performance of three state-of-the-art DL models, YOLOv10x, RetinaNet, and MM-Grounding DINO, for wheat head detection. To ensure diversity of the dataset, we have integrated two datasets, the Global Wheat Head Detection (GWHD) 2021 and the SPIKE dataset representing various wheat genotypes. This study aims to advance wheat head detection methods and offer a comparative evaluation of these three DL models.

Plant phenotyping relevance

小麦穂の画像検出・カウントによる表現型取得を中心に、複数の深層学習モデルとデータセットを比較評価しており、方法比較・ベンチマークとして適格。

Code and data availability

The paper uses the public GWHD_2021 and SPIKE datasets, but these are cited prior-work datasets, not paper-specific deposits, and no authors' code, models, or data availability statements with URLs are provided. No qualifying assets exist.

No evidence-backed public reproduction asset is currently recorded.