Helianthus tuberosus L. (Jerusalem artichoke) is native to North America (Swanton et al., 1992). Because it produces large quantities of edible tubers, H. tuberosus was an essential crop for native North Americans before European contact (Kays & Nottingham, 2007). The species has only been weakly domesticated, so high levels of genetic diversity exist among individuals and populations in physiological, morphological, and life‐history traits (Kays & Kultur, 2005; Puttha et al., 2012; Swanton et al., 1992). Also, it has become naturalized and invasive in many regions of the world (Tesio et al., 2012; Weber & Gut, 2004).

We purchased seed tubers of H. tuberosus from three private farms in Tochigi, Chiba, and Gunma prefectures, Japan. Because these farms are at least 80 km apart from each other, we treated the plants from each farm as a distinct population. The seed tubers were divided and planted into individual nursery pots (6 cm in diameter; 0.3 L volume) in a commercial soil mixture (Golden; Iris Ohyama Co.). Total sixty germinated sprouts were transplanted in random order 1 m apart into three‐row plots in a crop field at the Institute for Sustainable Agro‐Ecosystem Services of the University of Tokyo (35°44′03″N, 139°32′22″E) on 28 April 2017. The rows were covered with plastic mulch film (60 cm width). Because this experiment had different research purposes, some plants were paired or grouped, but these plants did not affect the growth of focal individuals grown singly and were omitted from the subsequent analyses. For more details, see our previous study (Fukano et al., 2019). We measured individual height and stem diameter five times during plant development (13 May, 1, 13, 30 June, and 14 July) by using ruler and caliper, respectively.

A low‐cost commercial UAV (DJI Inspire 1, DJI) with a built‐in camera (Zenmuse X5 Pro; 17.3 mm × 13.0 mm CMOS, 4,608 × 3,456 pixels; 16 MB with JPEG format) was flown over the field along a predesigned waypoint mission controlled by a commercial mobile phone application (Litch; VC Technology Ltd.). The waypoint mission plotted a double‐grid at an altitude of 15 m, a cruising speed of 2.5 m/s, camera looking downward, a >90% overlap of photographs to the front and sides, and an average ground sampling distance (GSD) of ~4 mm/pixel. Six ground control points (GCPs) made by acrylic plates were placed evenly (four corners and two nearly central) on the field and measured by Hemisphere RTK differential GNSS devices (Hemisphere GNSS). The point cloud was georeferenced by using a combination of direct georeferencing and the six GCPs. The mean root mean square of each computed GCP was 8.5 mm in the x‐direction, 16.4 mm in the y‐direction, and 20.1 mm in the z‐direction. UAV flight campaigns were conducted nine times during plant development (16, 31 May, 4, 12, 16, 29 June, 3, 7, 10 July).

From a set of multi‐view two‐dimensional images, canopy architectures were reconstructed as point cloud data by Structure from Motion (SfM) and Multi‐View Stereo (MVS) algorithms in Pix4Dmapper software (Pix4D SA). SfM is a photogrammetric technique used to simultaneously estimate the depth of corresponding points and camera position and direction from a set of multi‐view images. First, corresponding points among images are detected based on local features (e.g., SIFT, ORB, AKAZE; see (Tareen & Saleem, 2018) for details). Second, both extrinsic camera parameters (i.e., position and orientation) of a pair of views and the depth of corresponding points are estimated simultaneously by solving linear equations of the relation between camera coordinate frames calibrated with intrinsic camera parameters (e.g., focal length and principal point). A sparse point cloud is obtained as a set of corresponding points in three‐dimensional space by successively applying this estimation procedure to arbitrary pairs of multi‐view images. Third, several camera parameters and the point cloud are refined through bundle adjustment, which is a iterative optimization method to solve nonlinear equation, here, subject to minimization of reprojection error, see (Hartley & Zisserman, 2004) for details. Finally, MVS generates a dense point cloud based on the set of multi‐view images and camera parameters estimated in SfM. In our analysis, we used default intrinsic parameters provided by Pix4Dmapper as initial values. An orthomosaic image of the whole field was then generated from the Digital Surface Model (DSM) based on the dense point cloud (Figure 1, step 1).

We then extract the plot images manually. First, by using the “fishnet” function of ArcGIS 10.5 software (ESRI), a net of adjacent rectangular cells are generated according to user‐input numbers of rows and columns inside the predefined field boundary. Then, the plot ID is semiautomatically recorded in an attribute table by the Field Calculator tool of ArcGIS. Finally, a shapefile that contains all plots information is exported to a self‐developed Matlab (MATLAB v. R2017b, MathWorks Inc.) script to extract corresponded plot images from both the DSM and the orthomosaic.

The image segmentation process is needed to extract individual plants from plot images. In most cases, images are segmented by color on account of a large contrast between plants and bare soil (Fan et al., 2018; Guo et al., 2013, 2017). However, if the background includes objects with similar colors, such as weeds, further processing is needed. Therefore, researchers proposed several methods to distinguish weeds from the plant. For example, the use of a specific camera that can provide more spectral information; the use of complex algorithms such as machine learning with the manual selection of features and models, deep learning with the manual labeled training data (Chandra et al., 2020; Guo et al., 2018; Pérez‐Ortiz et al., 2015; Sa et al., 2017, 2018; Yu et al., 2019). Here, relying on the height of H. tuberosus, we propose the simple WEIPS method to segment plants from weeds. The core conception of WEIPS is segmentation by both color and height (Figure 1, step 2). In parallel, the segmented plot images from orthomosaic are processed by a machine‐learning‐based color pixel segmentation method, and the segmented plot images from DSM are processed by adaptive thresholding of height. Both methods extract a fixed polygon mask that indicates the candidate region. The masks are combined to render an individual plant without weeds.

Several phenotypic traits are extracted from the segmented individual plants. The cover area, major and minor axis lengths, eccentricity, orientation, convex area, filled area, equivalent diameter, solidity, extent, perimeter, and roundness can be easily calculated by MATLAB function “regionprops” (Figure 1, step 3). The height, volume, and outline are computed from the corresponded segmented DSM as following algorithms.

To validate the WEIPS method, we examined the correlation between the canopy coverage rate and the height of individual plants segmented by WEIPS and those manually segmented by author YF, using Pearson's correlation analysis, for each of five measurement dates.

To illustrate the utility of our framework in‐field phenotyping, we examined the phenotypic variations among the three H. tuberosus populations in various traits estimated from UAV imaging (Figure 1) and measured manually (height and stem diameter). The trait values were treated as response variables, and the source populations and planting row were treated as explanatory variables. All trait values were fitted to generalized linear models with a Gaussian distribution. Likelihood ratio tests were used to evaluate the significance of the explanatory variables, excluding the outline. For outline, permutational multivariate analysis of variance (PERMANOVA; Anderson, 2001; Mcardle et al., 2001) was conducted to evaluate the effect among a population by using the pairwise Adonis package (Martinez, 2020) in the R language (R Core Team, 2020) because only dissimilarities could be calculated.

We developed an easy‐to‐use framework for UAV‐based IPP (Figure 1) that requires only a commercial‐level UAV as hardware, and all processes are easy to implement.

We developed a novel technique, WEIPS, to eliminate weeds from UAV images by combining color‐based segmentation and adaptive thresholding‐based segmentation. The performance of WEIPS has been evaluated by Qseg, which is a well‐known measurements of vegetation segmentation method (Guo et al., 2013; Meyer & Neto, 2008). Qseg is defined as below:

where A is the set of the vegetation pixels (v = 255) or background pixels (v = 0) identified by a classification model, B is a reference set of manually segmented vegetation pixels (v = 255) or background pixels (v = 0), m and n are the image row and column sizes, and i, j are the pixel coordinate indices of the images. The more consistent pixels between A and B, the values become the larger ranging from 0 to 1. Namely, the higher the value, the more accurate the segmentation is.

The WEIPS method correctly (Qseg = 0.90) separated the weeds from the images and segmented the individual plants of the focal species, except for five individuals (surrounded by a dotted line; Figure 3a). We found a significant but relatively weak correlation between the canopy cover area of individual plants segmented by WEIPS and those manually segmented (R = 0.7843, Figure 3b). The correlation was weakened by the five exceptions (Figure 3b, blue points). A careful check of the photographs, DSM data, and WEIPS results revealed that the failure to segment those five exceptions were due to the stakes used to support the plants. In three‐dimensional reconstruction, the point cloud data of the five plants included the structure of the stakes, which caused the overestimation of plant height and resulted in an inadequate threshold for the segmentation. When we excluded the five individuals from the analysis, the correlation between WEIPS and manual segmentation became stronger (R = 0.9472, Figure 3c).

There were significant correlations between plant height estimated from the UAV images and that measured by hand on all measurement dates (p < .001 for all dates, Figure 4). The correlations were relatively high (R ² ≈ 0.85) except on 12 June (R ² = 0.47).

The results of all statistical tests are shown in Tables S1 and S2. We detected significant variations among the source populations in several traits estimated by UAV‐based IPP. For example, for the data obtained from 16 May, minor axis length, height, and volume differed significantly among populations (p = .0053, <.001, <.001, respectively). Other traits did not differ among populations. The outline differed between population 1 and 2 and population 1 and 3 (p = .014 and <.001, respectively) but not differed between population 1 and 2 (p = .084).

Supportive data and source code are available at the support page is here as follows: https://doi.org/10.5061/dryad.0cfxpnw0b