Introduction

Rice (Oryza sativa) is one of the key staple foods, feeding > 50% of the global population (Muthayya et al., 2014). Breeding for rice improvements in yield potential, grain quality and resistance to stresses is vital for its climate-change adaptation and, thus, food security in many rice-consuming nations (Nakashima et al., 2007; Jagadish et al., 2012). This relies on selecting favourable phenotypes of agronomic traits from thousands of varieties in the field, which in turn heavily relies on specialists' visual assessment (Bevan et al., 2017; Roitsch et al., 2019). To help accelerate the selection procedure, many field-based phenotyping approaches have been introduced (Zhao et al., 2019; Yang et al., 2020).

Additionally, as agronomically important traits are controlled by the expression of multiple genes and modulated by the environment, phenotyping can assist researchers to understand underlying biological mechanisms that contribute to genetic gain (Hartung & Schiemann, 2014; Furbank et al., 2019). Through genome-wide association studies (GWAS), the genetic architecture of some agronomic traits in rice has been dissected (Huang et al., 2010; Yang et al., 2014; Tang et al., 2019), laying the foundation of identifying functional diversity of alleles to discover valuable genes (Xing & Zhang, 2010). These contributions have led to advances in rice genetics and the development of new varieties with desired qualities, including high yield potential, resistance to stresses and increased resource-use efficiency (Barabaschi et al., 2015; Du et al., 2018; Li et al., 2018).

Certain traits such as plant height can be phenotyped at a specific time point; however, for growth- and yield-related traits that are genetically complex and influenced heavily by environmental factors, their phenotypes need to be examined dynamically (Naito et al., 2017; Mu et al., 2022). Nevertheless, to achieve this target, consistent data collection and trait analysis are required, which has posed significant challenges in developing reliable solutions for practical breeding programmes and field-based plant research (Shakoor et al., 2017; Pieruschka & Schurr, 2019). In essence, several problems need to be addressed, including: (1) scalability, trials are normally large-scale and at multiple sites; (2) affordability, resources are usually limited and solutions need to be cost-effective; (3) accuracy and repeatability, analysis results should be consistent and reproducible in other trials; (4) processing cycle, the duration between breeding cycles or multiseason experiments is often brief, requiring data to be processed, analyzed and fed-back promptly to enable timely decisions (Großkinsky et al., 2015; Atkinson et al., 2018). Recently, several advances have been adopted by breeders and plant researchers, but many attempts remain at early stages (White et al., 2012; Juliana et al., 2019). New tools derived from some academic research have often worked at relatively small scale and with limited accessibility as a result of bespoke hardware, proprietary software and specialized packages, preventing them from being employed easily (Yang et al., 2020, 2021). Furthermore, to exploit genomic resources, traits of interest and genetic diversity need to be assessed across sites and seasons, demanding accessible data collection and analysis toolkits (Naito et al., 2017; Atkinson et al., 2018). Hence, methodological advances shall intend to address the above challenges, which is the emphasis of this study.

One of the most exciting advances recently was the rapid development of unmanned aerial vehicles (UAVs, also known as unmanned aircraft systems) and their applications in crop monitoring resulting from their mobility, throughput and affordability (Shi et al., 2016; Maimaitijiang et al., 2017; Jang et al., 2020). There are numerous examples in the literature reporting UAV-based phenotyping using sensors such as red-green-blue (RGB) cameras, multi- and hyperspectral devices, Light Detection and Ranging (LiDAR), and thermal and infrared sensors (Kachamba et al., 2016; Gracia-Romero et al., 2017; Harkel et al., 2020; Hyyppä et al., 2020). Some work also reported quantitative trait loci (QTL) mapping of traits including plant height and vegetation fraction (Hassan et al., 2019; Wang et al., 2019; Ogawa et al., 2021). Nevertheless, most of these studies focused on estimating static traits collected at specific time points (Shakoor et al., 2017; Rodene et al., 2022), which often missed the dynamic nature of plant growth and development. Key agronomic traits (e.g. senescence and stem elongation) vary in time and space, which require new approaches to collect and analyse (Xu et al., 2018; Anderson et al., 2020). In fact, in a trial containing diverse genotypes, each line grows at a different pace, and thus dynamic analysis can provide meaningful comparisons between the genotypes (Hartung & Schiemann, 2014). Finally, changing behaviours of target traits, within or across seasons, can characterize the plant's complex responses to external stimuli, which are direct evidence to reveal spatial and temporal changes in the expression of genes and their regulators (Roitsch et al., 2019; Mu et al., 2022).

To extract meaningful information from UAV-collected imagery, many analytic solutions have been developed to measure traits related to yield, stress tolerance and growth patterns, using morphological, spectral and textural properties (Perez-Sanz et al., 2017; Jiang et al., 2021), most of which have focused on dryland crops. For example, Easy MPE (Tresch et al., 2019) combined excess green (ExG) and automatic thresholding to study soybean; AirSurf (Bauer et al., 2019) employed deep learning to count and classify lettuces; Grid (Chen & Zhang, 2020; Tang et al., 2021) utilized pixel-wise K-means clustering to delineate irregular (e.g. zigzag) or regular (e.g. grid-based) trial layouts for wheat trials; R/UAS::plotshpcreate (Anderson & Murray, 2020) created polygon shapefiles using parameters (e.g. field direction and plot size) to study maize; FIELDimageR (Matias et al., 2020) incorporates manual inputs (e.g. row and column numbers) into the extraction of plot-based traits for potato.

Still, limited tools are available for nonexperts to examine multigenic traits and develop markers for paddy field crops (e.g. rice), which are complex as a consequence of changing water levels (e.g. resulting from rainfall and draining) and many voluntary plants (e.g. duckweed) compared with dryland crops (Ogawa et al., 2021). Moreover, few research groups have the resources to process large-scale aerial images, or to develop complex algorithms to address problems in automated trait analysis (Roitsch et al., 2019; Zhu et al., 2021). Hence, along with the development of open-source computer vision, machine learning and data science libraries (Howse, 2013; Virtanen et al., 2020), open solutions will be valuable to equip plant researchers with new toolkits to study complicated crops.

In order to address some of the challenges, we have developed AirMeasurer, an open-source platform that automates trait analysis for rice trials using 2D orthomosaics and 3D point clouds acquired by low-cost UAVs. First, we established tailored protocols for regular flight missions and data pre-processing. Secondly, varied 2D/3D analysis algorithms were integrated into the platform to quantify static traits such as seedling number, plant height, canopy coverage and vegetative indices, using morphological, spectral and textural signals. Thirdly, we developed an original algorithm to compute dynamic traits based on static traits, including growth rates of the target traits and their rapid growth phases, which were time-consuming or impossible to score previously. To ensure that our work could reach the broader research community, we created a graphical user interface (GUI) for nonexperts to use. Finally, to validate the platform and its utility in research, we applied the AirMeasurer-derived traits collected from hundreds of rice landraces and recombinant inbred lines (RILs) in a multiseason case study (2019–2021) to genetic mapping studies (i.e. GWAS and QTL mapping) and identified reliable loci.

中文翻译：

---

AirMeasurer：开源软件，用于量化多季节空中表型的静态和动态特征，以增强水稻的遗传图谱研究

介绍

水稻 ( Oryza sativa ) 是重要的主食之一，养活了全球 50% 以上的人口（Muthayya等人，  2014 年）。提高水稻产量潜力、谷物质量和抗逆性的育种对于适应气候变化至关重要，因此对于许多水稻消费国的粮食安全至关重要（Nakashima 等人，2007 年；Jagadish等 人，2012年 ）。这依赖于从田间的数千个品种中选择有利的农艺性状表型，而这又在很大程度上依赖于专家的视觉评估（Bevan 等人，2017年 ； Roitsch等人，  2019 年）). 为了帮助加快选择过程，引入了许多基于田间的表型分析方法（Zhao等人，2019 年；Yang等人，  2020 年）。

此外，由于农艺学上重要的性状受多个基因的表达控制并受环境调节，表型分析可以帮助研究人员了解有助于遗传增益的潜在生物学机制（Hartung 和 Schiemann，2014 年；Furbank 等 人，2019年 ）。通过全基因组关联研究（GWAS），解析了水稻部分农艺性状的遗传结构（Huang et al .,  2010 ; Yang et al .,  2014 ; Tang et al .,  2019），为鉴定水稻农艺性状奠定了基础。等位基因的功能多样性以发现有价值的基因 (Xing & Zhang,  2010). 这些贡献推动了水稻遗传学的进步和具有所需品质的新品种的开发，包括高产潜力、抗逆性和提高资源利用效率（Barabaschi 等人，2015 年；Du等人 ， 2018年；  Li等人等人，  2018 年）。

某些性状如株高可以在特定时间点进行表型分析；然而，对于基因复杂且受环境因素影响很大的生长和产量相关性状，需要对其表型进行动态检查（Naito 等人，2017年 ； Mu等人，  2022 年）。然而，要实现这一目标，需要一致的数据收集和性状分析，这对为实用育种计划和田间植物研究开发可靠的解决方案提出了重大挑战（Shakoor 等人，2017 年；Pieruschka 和Schurr，  2019年 ）。本质上，需要解决几个问题，包括：(1)可扩展性，试验通常是大规模的并且在多个地点进行；(2)可负担性，资源通常是有限的，解决方案需要具有成本效益；(3)准确性和重复性，分析结果在其他试验中应具有一致性和可重复性；(4)处理周期，育种周期或多季节实验之间的持续时间通常很短，需要及时处理、分析和反馈数据，以便及时做出决策（Großkinsky 等，2015；  Atkinson等， 2018  ）。最近，育种者和植物研究人员采用了几项进展，但许多尝试仍处于早期阶段（White等人.,  2012 ; 朱莉安娜等人，  2019 年）。由于定制硬件、专有软件和专用软件包，从一些学术研究中衍生出来的新工具通常在相对较小的规模上工作，并且可访问性有限，从而阻碍了它们被轻易使用（Yang 等人，2020年，  2021年）。此外，为了开发基因组资源，需要跨地点和季节评估感兴趣的特征和遗传多样性，需要可访问的数据收集和分析工具包（Naito 等人，2017年 ； Atkinson等人，  2018 年）). 因此，方法学的进步应旨在应对上述挑战，这是本研究的重点。

最近最令人兴奋的进展之一是无人驾驶飞行器（UAV，也称为无人驾驶飞机系统）的快速发展及其在作物监测中的应用，这是由于其机动性、吞吐量和可承受性（Shi et al., 2016 ;  Maimaitijiang et al .,  2017 年；Jang等人，  2020 年）。文献中有许多例子报告了使用红-绿-蓝 (RGB) 相机、多光谱和高光谱设备、光探测和测距 (LiDAR) 以及热和红外传感器等传感器进行基于无人机的表型分析（Kachamba 等人，2017 年）。 ,  2016 年；Gracia-Romero等人，  2017 年; 哈克尔等人，  2020 年；Hyyppä等人，  2020 年）。一些工作还报告了植物高度和植被比例等性状的数量性状位点 (QTL) 作图（Hassan等人，  2019 年；Wang等人，  2019 年；Ogawa等人，  2021 年）。尽管如此，这些研究中的大多数都侧重于估计在特定时间点收集的静态特征（Shakoor等人，  2017 年；Rodene等人，  2022 年）), 这往往错过了植物生长发育的动态性质。关键农艺性状（例如衰老和茎伸长）随时间和空间变化，需要新的方法来收集和分析（Xu等人，  2018 年；Anderson等人，  2020 年）。事实上，在包含不同基因型的试验中，每条线都以不同的速度生长，因此动态分析可以提供基因型之间有意义的比较 (Hartung & Schiemann, 2014  )。最后，在季节内或跨季节改变目标性状的行为可以表征植物对外部刺激的复杂反应，这是揭示基因及其调节因子表达的空间和时间变化的直接证据（Roitsch等人，  2019 年；穆等人，  2022）。

为了从无人机收集的图像中提取有意义的信息，已经开发了许多分析解决方案来测量与产量、胁迫耐受性和生长模式相关的特征，使用形态学、光谱和纹理特性（Perez-Sanz 等人，2017 年；Jiang 等人，  2017年）。 ,  2021 ), 其中大部分都集中在旱地作物上。例如，Easy MPE (Tresch et al .,  2019 ) 结合过量绿色 (ExG) 和自动阈值来研究大豆；A ir S urf (Bauer et al .,  2019 ) 采用深度学习对生菜进行计数和分类；网格(Chen & Zhang,  2020 ; Tang et al .,  2021 ) 利用像素级 K 均值聚类来描绘小麦试验的不规则（例如之字形）或规则（例如基于网格）试验布局；R/UAS:: plotshpcreate (Anderson & Murray,  2020 ) 使用参数（例如田地方向和地块大小）创建多边形形状文件来研究玉米；FIELD image R（Matias等人，  2020 年）将手动输入（例如行号和列号）纳入马铃薯基于地块的性状提取中。

尽管如此，非专家可用于检查多基因性状和开发水田作物（例如水稻）标记的工具仍然有限，水田作物由于水位变化（例如降雨和排水造成）和许多自愿植物（例如浮萍）而变得复杂与旱地作物相比（Ogawa等人，  2021 年）。此外，很少有研究小组有资源来处理大型航拍图像，或开发复杂的算法来解决自动特征分析中的问题（Roitsch等人，  2019 年；Zhu等人，  2021 年）。因此，随着开源计算机视觉、机器学习和数据科学图书馆的发展（Howse，  2013; Virtanen等人，  2020 年），开放式解决方案对于为植物研究人员提供新工具包来研究复杂作物非常有价值。

为了应对其中的一些挑战，我们开发了 AirMeasurer，这是一个开源平台，可以使用低成本无人机获取的 2D 正射镶嵌和 3D 点云自动对水稻试验进行性状分析。首先，我们为定期飞行任务和数据预处理建立了量身定制的协议。其次，平台集成了多种 2D/3D 分析算法，利用形态、光谱和纹理信号量化静态性状，例如幼苗数、株高、冠层覆盖度和植物指数。第三，我们开发了一种原始算法来计算基于静态特征的动态特征，包括目标性状的增长率及其快速生长阶段，这在以前是耗时的或无法评分的。为了确保我们的工作能够影响到更广泛的研究社区，我们创建了一个供非专家使用的图形用户界面 (GUI)。最后，为了验证该平台及其在研究中的实用性，我们应用了 A在多季节案例研究（2019-2021 年）中从数百个水稻地方品种和重组自交系 (RIL) 收集的 ir Measurer 衍生性状进行遗传作图研究（即 GWAS 和 QTL 作图）并确定了可靠的位点。