Analysis of key factors affecting maize tassels detection and construction of shared dataset based on UAV

Abstract

Background: Rapid and accurate detection of tassels is of great significance for maize breeding, seed production and the acquisition of key growth stage. To liberate manpower and improve the efficiency of production management, many automatic detection methods with acceptable accuracy have been proposed. However, images acquisition parameters of these methods were quite different, so they cannot provide an operable standard for practical applications. In this study, based on multi-temporal unmanned aerial vehicle (UAV) RGB images with maize flowering stage, we created UAV Maize Tassel Detection (UAVMTD) dataset, and used Faster R-CNN to answer what are the key factors affecting detection accuracy from two aspects of efficient use of samples and data acquisition standards. Based on the detection results, we estimated tasseling date of different plots and analyzed varieties’ differences. Results: The results show that model performance would not be greatly affected before the amount of training data changed by orders of magnitude, but it can be improved effectively by adjusting sub-images’ sizes, and the final model was selected with AP@0.5IOU was 0.916; images obtained at 12 pm were more suitable for tassels detection, AP@0.5IOU, recall and precision were 3%, 2% and 6% higher than that at 8 am; optimal spatial resolution was around 1cm for tassels detection by considering the recognition effect and data acquisition efficiency. Conclusions: This study analyzed key factors affecting maize tassels detection and provided a reasonable reference for future applications, which is helpful to screen out varieties from large-scale breeding materials.

Plant phenotyping relevance

UAV画像によるトウモロコシ雄穂検出のデータセット構築、検出精度の検証、撮影条件の最適化を中心とする植物フェノタイピング研究である。

Code and data availability

The paper's UAV maize tassel detection dataset (UAVMTD: 142 UAV RGB images, 28,182 labeled tassels) is explicitly stated to be publicly available on the authors' GitHub repository, which is an allowed URL. No separate analysis code or trained model checkpoints are explicitly deposited.