Analyzing architectural diversity in maize plants using the skeletonimage-based method

LIU Min-guo , Thomas CAMPBELL, LI Wei , WANG Xi-qing #

1 College of Biological Sciences, China Agricultural University, Beijing 100193, P.R.China

2 Frontier Technology Research Institute of China Agricultural University in Shenzhen, Shenzhen 518000, P.R.China

3 Center for Crop Functional Genomics and Molecular Breeding, China Agricultural University, Beijing 100193, P.R.China

4 Department of Biology, Northeastern Illinois University, Chicago 60625, USA

Abstract Shoot architecture in maize is critical since it determines resource use, impacts wind and rain damage tolerance, and affects yield stability.Quantifying the diversity among inbred lines in heterosis breeding is essential, especially when describing germplasm resources.However, traditional geometric description methods oversimplify shoot architecture and ignore the plant’s overall architecture, making it difficult to reflect and illustrate diversity.This study presents a new method to describe maize shoot architecture and quantifies its diversity by combining computer vision algorithms and persistent homology.Our results reveal that persistent homology can capture key characteristics of shoot architecture in maize and other details often overlooked by traditional geometric analysis.Based on this method, the morphological diversity of shoot architecture can be mined (quantified), and the main shoot architecture types can be obtained.Consequently, this method can easily describe the diversity of shoot architecture in many maize materials.

Keywords: maize, shoot architecture, persistent homology, phenotyping technology, morphological diversity

1.Introduction

Maize (ZeamaysL.), the most extensively farmed staple grain crop, has a diverse morphological and genetic structure, and plays a crucial role in ensuring worldwide food security (Buckleret al.2006; Li Pet al.2019;Gaoet al.2023).Recently, maize shoot architecture has received extensive attention because it is involved in resource efficiency, yield stability, and resistance to damages caused by wind and rain (Fasoula and Fasoula 2010; Strableet al.2017; Xueet al.2017; Tianet al.2019).The diversity of inbred lines is of utmost importance for heterosis breeding, particularly when describing germplasm resources (Smith and Smith 1989;Shrestha 2016).

Structure-related geometric and quantitative characteristics of maize, such as plant height, ear height,leaf number, leaf length, leaf angle, and leaf orientation,have been traditionally used to describe shoot architecture(Duncan 1971; Zheng and Liu 2013), thus far contributing to considerable advances in shoot architecture research (Lambert and Johnson 1978; Stewartet al.2003; Bouchetet al.2017).Traditional methods for describing maize architecture tend to use only a few key characteristics to reflect the overall architecture, possibly due to the complexity and laborious measurements and records required.Persistent homology (PH) is a new topological model interpretation method that can extract morphological traits from 2D or 3D representations and can be used to compare highly different forms (Feiet al.2022).PH is particularly well suited for quantifying branch topology because it can quantitatively summarize complicated changes with a single distance index (Deloryet al.2018).Liet al.(2017) introduced the PH method into plant research and demonstrated its potential in describing grape inflorescence structure and soybean branching characteristics with X-ray imaging and manual labeling separately (Li Met al.2019; Dhakalet al.2021).However, their method is limited by expensive equipment and manual labeling and ignores its advantages in diversity analysis.

Here, we show a new and novel maize architecture description pipeline based on persistent homology to quantify the differences between complex architectures and explore plant architectural diversity among maize populations.

2.Materials and methods

2.1.Plant materials and plant image description

A maize inbred line population, including 508 lines, was used to analyze the shoot architecture (Appendix A).The maize inbred lines of the population include those from tropical, subtropical, and temperate backgrounds,representing the global diversity of maize (Yanget al.2010, 2011).These lines were planted under the greenhouse conditions with two replicates, and each plant was scanned from the seedling to the heading stage using the RAP platform (Zhanget al.2017).All maize inbred lines were photographed every 3 days; a total of 16 photos (T1-T16) were taken.In this study, only the last plant image (T16) was selected for analysis.Most of the materials used in the images are at the heading stage.Images of abnormal growth (death at the initial stage of growth) were filtered, and 1 006 pictures were retained(Appendix A).

2.2.Acquisition of traditional geometric traits for shoot architecture

Traditional geometric traits, including tips, convex hull area, solidity, perimeter, and width and height, were evaluated with PlantCV packages (Gehanet al.2017).The measurement process of traditional geometric traits is shown in Appendix B.

2.3.Extracting topological characteristics of maize shoot architecture

The original image was transformed into a mask image through color space transformation and binarization with PlantCV in Python (Fig.1-B).A skeleton image with a pixel width was created based on the mask image(Fig.1-C).The cleaned skeleton image was read with R,and black pixels were located (Fig.1-D).The distance between pixels was used to calculate an interconnectivity relationship.An undirected network was built with the igraph package in R Software (Csardi and Nepusz 2006).Using the lowest point in the skeleton image as the base point, the geodesic distance from all skeleton points to the base point was calculated (Fig.1-E).Persistent homology was calculated in MATLAB using the JavaPlex package (Tauszet al.2014) and output to a result matrix data based on the geodesic distance and connection of all points (Fig.1-F and G).The bottleneck distance was estimated using the bottleneck function of the TDA package in R to represent the similarity between architectures (Fasyet al.2021).A pairwise distance matrix was used in this study to analyze the variation in the shoot architecture of maize as well as its diversity.The code can be found at https://github.com/liumiguo/plantTypeWithPH

Fig.1 Analysis workflow for shoot architecture of maize.A, original image.B, binary image.C, pruned skeleton images.D, point coordinate image.E, an undirected network was constructed based on the position of black pixels, and the geodesic distance from all pixels to the based point was calculated with the lowest position of the image as the based point.F, the persistent homology matrix.G, the persistence barcode.

2.4.Statistical analysis

Multidimensional scaling (MDS) was performed on the pairwise bottleneck distance matrix with the cmdscale function in R to project the data into lower-dimensional Euclidean space (R Core Team 2021).The variance rate of each MDS axis was calculated based on the ratio of the eigenvalues of each axis to the total eigenvalues.The first five MDS axes accounted for 81% of the total variance (Fig.2-A).The relationship between the first five MDS axes and traditional geometric traits of shoot architecture was then explored and shown using heat maps with the ggcorrplot package (Kassambara 2019)in R (Fig.2-C).We calculated the five quantiles of the first MDS axis and obtained the image ID corresponding to each quantile to examine information about the main variation trend in maize shoot architecture as reflected in the first MDS axis (Fig.2-D).

Fig.2 Shoot architecture characteristics in maize captured by persistent homology.A, proportion and accumulated proportion of multidimensional scaling (MDS) axis.B, probability density distribution of MDS axis.C, correlation analysis between traditional geometric traits and MDS axes.D, shoot architecture of maize corresponding to the five quantiles of the first MDS axis (see Materials and methods section).

KMEDOIDS analysis was used to determine the clustering relationship between maize plants using the pam function in the clustering package (Maechleret al.2021).Changes in gap statistics associated with various classification numbers were used to identify the best classification number (Tibshiraniet al.2001).In a two-dimensional scatter plot in which different types of plants are given different colors, the first two MDS axes (which account for 69% of the variation)were chosen to map the relationship between the relative architecture differences between various plants.The KMEDOIDS results contain image IDs for representative architectures of each type.Images of various types of architecture were displayed to illustrate their characteristics (Fig.3-D).

Fig.3 Diversity analysis of shoot architecture in maize.A, gap statistics for various cluster numbers (K), with K=12 having the highest value.B, the number of varying types of maize plants.Based on KMEDOIDS, all plants were divided into 12 types, and the number of each type was computed.C, the scatterplot of the first two MDS scores, which account for 69% of the total variance,in which different colors correspond to shoot architectures of 12 types, and black dots indicate the location of the 12 central architectures.D, the architecture of the central plant shoot of the 12 types corresponding to the K=12 in A.

Images and data can be obtained from www.plantphenomics.hzau.edu.cn.

3.Results

3.1.New method captures multiple architecture characteristics of maize shoots

Based on MDS analysis, several characteristic indices reflecting shoot architecture were constructed by projecting the distance matrix into low-dimensional Euclidean space.Of these, the first two MDS axes account for 69% of the variance, and the first five MDS axes account for 81% (Fig.2-A).These MDS axes reflect the majority of the information in the architecture.The first MDS axis reflects the primary variation trend, and the five quantiles in the MDS axis demonstrate that the crown size of the corresponding shoot increased as the MDS1 value increased (Fig.2-D).

The traditional geometric traits, such as height, width,and perimeter, were assessed based on the maize shoot image (Appendix B).The relationships between traditional geometric traits and MDS axes were investigated(Fig.2-C).MDS1 had a significant positive relationship with most of the traits, particularly convex hull area and plant height, but a significant negative correlation with solidity.MDS2, MDS3, and MDS4 correlated negatively with most traditional geometric traits, particularly MDS4 with tips and perimeter.MDS5 had a positive association with most traditional geometric traits.However, these relationships were weaker than MDS1.

3.2.This new technique reveals diversity in maize shoot architecture

KMEDOIDS was used here to group all inbred lines of maize based on their shoot architectures.Gap statistics were used to determine the optimal number of clusters,and the results showed that K=12 had the highest value,meaning 12 was the optimal number of clusters (Fig.3-A).Based on this result, shoot architectures were divided into 12 types, the central architecture was displayed (Fig.3-D),and the number of each type was computed (Fig.3-B).Two MDS axes were used to draw the scatter diagram of shoot architectures after dimension reduction and to give different colors to different architecture types according to KMEDOIDS results, in which the black point in the center represented the architecture (Fig.3-D).The shoot architecture can be divided into 12 types.Mainly groups of all architectures according to the changes in MDS1 and MDS2 axes, where G1, G3, G9, G10, and G12 represent the types with smaller crown size, and G4, G6, and G7 represent the types with larger crown size were shown in Fig.3-C.However, G11 and G5 were different in the MDS2 direction, which was significantly negatively related to both the aspect ratio and height but not to the width.Moreover, hierarchical clustering can also be used to explore relationships between all materials and examine their main types and differences (Appendix C).

4.Discussion

In this study, maize shoot branching characteristics were characterized using the persistent homology method,which condensed complex morphological traits into a distance metric (bottleneck distance) (Liet al.2017;Deloryet al.2018).This method first requires a complete and clear shoot image, which can be obtained using a professional phenotyping platform such as RAP (Zhanget al.2017).Other commercial cameras can also be used.Because extraction of the skeleton is the basis for describing a structure, it is important to unify the background in order to extract the plants more easily,and topological features can only be rendered from the extraction of a complete skeleton.Bottleneck distances are used to describe the similarity between two topological features, so a distance matrix is constructed between the materials, representing diversity in plant architecture within the population.The dimensionality reduction and clustering methods are used to explore the structural diversity of maize populations based on the distance matrix.Multidimensional scaling (MDS) is a multivariate data analysis technique that displays a “distance” data structure in low-dimensional space (Zhanget al.2017).KMEDOIDS is based on the most centrally located object in a cluster, which is less sensitive to outliers (Park and Jun 2009).

This new method captures diverse aspects of shoot architecture, uncovers main trends in variation, and can accurately analyze a wide variety of shoot architectures and characterize representative shoot architectures efficiently.A possible application of this method is to describe the morphological diversity of plant architecture across a large number of maize germplasm resources and explore the main plant types.It can also be used to compare the degree of similarity in plant architecture between different populations.Additionally, we can select a plant architecture either similar to or significantly different from a target plant architecture from a population,which would be helpful in selecting materials from many populations.However, even though this method can capture the overall characteristics of maize plant architectures, it ignores some information about the leaves and stems, such as the diameter of the stems and the surface features of the leaves.

5.Conclusion

Topological features are more effective than traditional geometric description methods in capturing architectural traits and enabling the exploration of morphological diversities of shoot architecture across a wide range of plant materials.The work presented here provides a new approach to analyzing shoot architecture in maize and offers a new perspective on studying the morphological diversity of plant architecture.

Acknowledgements

The study work was supported by the National Key Research and Development Program of China(2022ZD0401801) and the Chinese Universities Scientific Funds (2023TC107).

Declaration of competing interest

The authors declare that they have no conflict of interest.

Appendicesassociated with this paper are available on https://doi.org/10.1016/j.jia.2023.05.017