A standardized and efficient technique to estimate seed traits in plants with numerous small propagules
This article is part of the special issue “Advances in Plant Imaging across Scales.”
Abstract
Premise
Variation in seed traits is common within and among populations of plant species and often has ecological and evolutionary implications. However, due to the time-consuming nature of manual seed measurements and the level of variability in imaging techniques, quantifying and interpreting the extent of seed variation can be challenging.
Methods
We developed a standardized high-throughput technique to measure seed number, as well as individual seed area and color, using a derived empirical scale to constrain area in Arabidopsis thaliana, Brassica rapa, and Mimulus guttatus. We develop a specific rational model using seed area measured at various spatial scales relative to the pixel count, observing the asymptotic value of the seed area as the modeled number of pixels approaches infinity.
Results
We found that our model has high reliability in estimating seed traits and efficiently processes large numbers of images, facilitating the quantification of seed traits in studies with large sample sizes.
Discussion
This technique facilitates consistency between imaging sessions and standardizes the measurement of seed traits. These novel advances allow researchers to directly and reliably measure seed traits, which will enable tests of the ecological and evolutionary causes of their variation.
In plants, seed number and seed size have important ecological and evolutionary implications. The reproductive fitness of an individual is commonly measured as the number of seeds produced and is therefore of fundamental importance to evolutionary questions. Additionally, seed size can affect fitness, through its impact on dispersal and predation (reviewed in Westoby et al., 1996; Eriksson, 2008; Bogdziewicz et al., 2019), as well as germination, establishment, and survival of the seedling (Krannitz et al., 1991; Eriksson, 1999; Elliott et al., 2007; Martínez-González et al., 2021).
Seed size has been studied extensively, with early emphasis on the extraordinary differences between species, with up to 106-fold variation within a region (Baker, 1972; Westoby et al., 1992; Leishman et al., 1995; Moles et al., 2005). Early research emphasized the relative constancy of seed size within species (Harper et al., 1970), but since then it has been well established that substantial variation exists within and among individuals (Janzen, 1977; Thompson, 1984; reviewed in Michaels et al., 1988; Eriksson, 1999; Gnan et al., 2014; Paczesniak et al., 2022). For example, seed mass varies positively with plant size (Hendrix, 1984; Aarssen and Jordan, 2001), with fruit maturation order (Fuller et al., 1983; Cavers and Steel, 1984; Hendrix, 1984; Torres et al., 2002; de Carvalho et al., 2021), with position within the ovary (Greenway and Harder, 2007), and with paternity (Stanton, 1984; Mazer et al., 1986; Andersson, 1990; Raunsgard et al., 2018). Seed size is thought to evolve as a compromise between producing numerous small seeds containing few resources and producing fewer large seeds containing more resources (Westoby et al., 1992; Leishman and Murray 2001; Gnan et al., 2014). The variation in seed size and number within species can have important ecological and evolutionary consequences; however, quantifying this variation requires accurate and robust measurement of large sample sizes.
The importance of seed traits and standardized, reproducible measurements was emphasized by Saatkamp et al. (2019) in their call for a seed-trait functional ecology program—a research agenda to characterize seed-trait variation that is connected to plant functions and ecological strategies. Such a research agenda requires the development and maintenance of databases and compilation of standardized and useful seed traits at the global scale (Moles et al., 2007). The seed traits that feed into the axes of the seed ecological spectrum include morphological traits (e.g., size, number, shape, color), chemical traits (e.g., toxicity, nutrients), and physiological traits (e.g., light and moisture requirement to break dormancy) (Saatkamp et al., 2019). While seed morphological traits such as size, shape, and color are only one dimension along which seeds can vary, they are arguably the easiest to measure. Despite the relative simplicity of these traits, methods to quantify them are quite limited. Mussadiq et al. (2015) compared the accuracy of various programs (e.g., ImageJ, CellProfiler, P-TRAP, and Smartgrain) to estimate seed number, and found that custom ImageJ macros produced the best estimation of seed count. Their study does not estimate other traits, such as seed size, or include model-based approaches to allow for universal application. The use of high-throughput approaches to characterize seed traits has broad applicability to many agricultural, ecological, and evolutionary studies, particularly those requiring large sample sizes and/or studies of small-seeded species.
The extensive variation in seed morphological traits has ecological and evolutionary significance; however, their characterization is limited by the lack of standardized methods for quantifying seed trait variation. Here, we develop a high-throughput method that has the capacity to measure seed number, size, shape, and color using digital images and theoretical modeling of optimal parameters. Below we describe the protocol for measuring particle number and size; the procedure has the capacity to additionally estimate traits such as shape and relative color (not shown). We use seeds from three model species (Arabidposis thaliana (L.) Heynh. [Brassicaceae], Brassica rapa L. [Brassicaceae], and Mimulus guttatus DC. [synonym Erythranthe guttata (DC.) G. L. Nesom; Phyrmaceae]) to determine the limit in pixel count for reliable measurements and provide context for establishing camera configurations for bulk imaging of seeds (or particles more generally). Importantly, we assess the influence of camera resolution at a millimetric scale on accuracy and robustness of measurement, as well as provide a procedure to perform high-throughput particle size and number measurements using ImageJ and R software tools.
METHODS
Background on identifying objects in digital images
Digital photographs are typically stored as raster data, comprising a two-dimensional grid (x, y) of square pixels, with red, green, and blue (RGB) values for each pixel for color images and a single value for each pixel for grayscale images. The term “resolution” has several meanings, and can refer to the number of pixels making up a given image, the dimension of individual pixels in real-world units (e.g., millimeters, centimeters) within an image, or the scale at which objects of a given size can be resolved in image space. We use the last meaning. To identify objects (such as seeds) in an image, a surface made up of pixels that are distinct from their surroundings is manually or algorithmically delineated. This could be done by relying on differences in pixel value or the change in pixel values as a function of distance within an image. The ability to determine the precise border between an object and the surrounding space is limited by the resolution of the image.
The perimeter of an object identified in an image includes a jagged line of pixels defining the edge, which is not linear in the x or y dimension (Figure 1). The delineation of object borders is imperfect and granular and can impact measurements. The borders produce error in misidentifying pixels associated with the object and in the portion of the pixel covering non-object space alongside actual object surface area. The limits of the square pixel geometry become proportionally more important as the size of the object becomes smaller and the ratio of edge pixels becomes larger. Another source of error is introduced in converting pixels to real-world units of measurement, which requires a pixel-to-area conversion scale. Any error associated with the conversion scale will propagate. To reliably use digital images for object measurement, one needs to account for these sources of error. Resolution and spatial scale are well-studied in geographic information science (reviewed in Atkinson and Tate, 2000), and resolution influences the ability to measure features on a landscape scale such as vegetation changes (e.g., Wu et al., 2002; Dunwoody et al., 2013). Bringing image analysis to propagule measurement enables confident use of high-throughput image processing methods and facilitates measurement of propagules in large quantities or of small sizes.
Propagule imaging
The optimal imaging configuration includes bright, diffuse light from multiple angles to minimize shadows of the imaged object. Later processing can compensate for minor shadowing; however, consistency within and between imaging sessions is vital for meaningful comparison of propagule sizes. We arranged seed propagules on a sheet of white paper, with a camera mounted directly above. We checked that the seeds were within the field of view of the camera, positioned with no overlap. For reproducible imaging, we marked the paper to record the area occupied by the propagules within the field of view, along with the position of the background sheet on the mounting table.
To image seeds, we used manual settings on a digital single-lens reflex (SLR) camera (see Table 1 for camera components and parameters). We set the lens of the camera to either the maximum or minimum magnification, as those limits are reliably reproducible by reaching the physical limit of the lens adjustment. We used an f-stop and shutter speed appropriate for the room lighting, allowing for a sharp and well-lit image within the field of view of the propagules. We used a low ISO as the scene had sufficient illumination to clearly distinguish between the propagules and the white background. The images are captured in a lossless imaging format (e.g., .tif format); if a lossless format is not available, the highest quality, least compressed format should be used, as information will be lost with compressed filetypes such as .jpg files.
| Camera components and parameters | Model or parameter value |
|---|---|
| Camera model | Nikon D3100 |
| Sensor size (mm) | 23.1 × 15.4 |
| Sensor resolution (px) | 4608 × 2304 |
| Lens model | Nikon AF-S DX NIKKOR 18–55 mm f/3.5-5.6 G VR |
| Shutter speed (s) | 1/500 |
| f-stop | f/8 |
| ISO | 100 |
| Focal lengths used (mm) | 18, 55 |
| Camera model | Leica Flexacam C1 |
| Sensor size (mm) | 6.25 × 4.69 |
| Sensor resolution (px) | 4000 × 3000 |
| Microscope model | Leica S8APO |
| Shutter speed (s) | 1/12 |
| f-stop | f/4.9 |
| ISO | 500 |
As a first step, we recommend users identify the camera configuration that results in a single propagule in the image being ≥100 px. This threshold ensures that the scale at which the seeds are resolved can be confidently converted to a physical area, as described below in the Validation section of the Methods. To achieve this, we used a camera height of 30 cm and focal length of 55 mm. With the established camera configuration, we took images with a precise scalebar to produce an empirical conversion ratio between image pixel size and physical size. We used a 12-inch ruler (product number 501-012; Products Engineering Corporation, Torrance, California, USA) placed in the center of the field of view and imaged four times, rotating the ruler by 45° after each image to capture length and diagonal scales. Deriving the empirical conversion ratio from the average of four ruler angles ensures that differences in distortion in x- and y-directions are minimized, to reduce any disparity in converted area for oblong particles. We do not account for distortions associated with un-orthorectified images.
Next, we imaged the propagules. We placed the propagules from a given sample within the designated frame, ensuring no overlap between propagules and the edge of the frame, and removing any detritus of a similar size to the seeds. The images of the seeds and the images of the ruler were transferred into separate folders on the processing computer.
Image processing in ImageJ and R
Using the Rectangle tool, we created a cropping region of interest (ROI) template that contained the propagules in each image. Because we initially standardized the base sheet position on the imaging table and the propagules position within the drawn frame, the ROI template was the same for all images.
We developed an expanded macro (particleSizeID.txt, Appendix S2) from the batch processing instructions developed by Herbert (2011) to streamline image processing. Users select the folder containing the propagule images and provide the case-sensitive file extensions to ensure the appropriate images are analyzed. The ImageJ macro file associated with this procedure (Appendix S2) is opened within the macro window, and the destination folder for the results, in .csv format, is selected.
In processing the images, there are several settings that may be useful in some circumstances. The Subtract Background setting determines whether the objects in the image are emphasized from a background with inconsistent color intensity. This setting is helpful for obtaining propagule counts from noisy images when the propagules are less than 50 px in radius, but ideally is not used for accurate area measures as it alters pixel intensities around the border of objects. Subtract Background was not used here. The Watershed concavity segmenting setting is helpful for obtaining accurate propagule counts at the cost of accuracy in area measurements. The setting segments objects identified in an image at points of concavity, such as the boundary between two roughly circular seeds that are in contact with one another. Because shape does not influence object identification in this macro, the watershed segmenting tool enables the isolation of individual objects from groups that are in contact with one another. However, the boundary derived from this tool does not consistently match the exact propagule edge and so should only be used to improve counts with overlapping image features. The propagules here were not in contact or overlapping and thus the watershed segmentation was not used. Next, the thresholding method is chosen using either the default ImageJ setting or by manually selecting a value between 0–255 to isolate the propagules from the background. With a controlled imaging setup and consistent lighting, the default provided efficient identification using a modified version of the IsoData method described by Ridler and Calvard (1978).
The macro operates by cycling through the input image folder, performing a series of processing steps before saving the propagule measurement in a .csv file. First, each image is cropped to the ROI file boundaries. It then converts the three-channel RGB image to an 8-bit grayscale image using the mean of the three-color values of each pixel. Next, the image is converted from grayscale to a binary image, isolating the propagule features from the background as a mask image. If the watershed segmentation tool is used, it splits concave features in the binary mask into multiple features. Otherwise, each isolated feature in the mask is used to extract the number of pixels, the minimum and maximum caliper distance in pixels, and minimum, maximum, and mean RGB values of each propagule in the original input image, saved to the designated output folder as a .csv file.
The protocol uses the R programming language (R Core Team, 2021) to process the ImageJ results. The particleSizeProcess.R script (Appendix S3) takes the measurement output from ImageJ and performs an initial filter to exclude seed pixel values outside of a given input range; this is to immediately exclude any obvious non-seed particles, such as plant debris or soil. The script then performs additional statistical filtering using the area calculated from each seed (SD filter), excluding values outside an input number of standard deviations above and below the mean. This SD filter can be excluded by setting a null value, or performed on log10-transformed data to manage right-skewness. The output is a list class object with two entries. The first entry is a data frame with a row for each particle identified in the ImageJ macro, containing its source file name, number of pixels counted by ImageJ, converted area calculated by the input scale, the log10-transformed area, the minimum and maximum caliper distance, and the mean, minimum, and maximum of each red, green, and blue channel as captured by the camera sensor. The second list entry contains the same information but is limited to the particles falling within the SD filter included in the function call. If no size filters are applied, the output is the unfiltered data frame containing the above-mentioned data for each particle. With a controlled imaging process, the colors can be used for comparison between samples, although color standardization would be necessary to compare to other propagule data sets.
Validation using a test sample
| Configuration name | Camera components | Camera height above seeds (cm) | Focal length (mm) | Empirical scale (mm/px) | Approximate magnification | Relevant test |
|---|---|---|---|---|---|---|
| Scope (MA) | Flexacam C1, dissection microscope (M. guttatus, A. thaliana) | — | — | 0.000305 | 5.12 | Robustness test |
| Scope (B) | Flexacam C1, dissection microscope (B. rapa) | — | — | 0.00252 | 0.621 | Robustness test |
| H28-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 28 | 55 | 0.0139 | 0.361 | Robustness test |
| H30-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 30 | 55 | 0.0159 | 0.316 | Test sample, robustness test |
| H35-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 35 | 55 | 0.0207 | 0.242 | Test sample, robustness test |
| H40-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 40 | 55 | 0.0255 | 0.197 | Test sample, robustness test |
| H45-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 45 | 55 | 0.0302 | 0.166 | Test sample, robustness test |
| H50-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 50 | 55 | 0.0349 | 0.144 | Test sample, robustness test |
| H55-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 55 | 55 | 0.0396 | 0.127 | Test sample, robustness test |
| H28-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 28 | 18 | 0.0411 | 0.122 | Robustness test |
| H60-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 60 | 55 | 0.0443 | 0.113 | Test sample, robustness test |
| H30-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 30 | 18 | 0.0472 | 0.106 | Robustness test |
| H65-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 65 | 55 | 0.0490 | 0.102 | Test sample, robustness test |
| H69.7-FL55 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 69.7 | 55 | 0.0535 | 0.094 | Test sample, robustness test |
| H35-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 35 | 18 | 0.0614 | 0.082 | Robustness test |
| H40-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 40 | 18 | 0.0744 | 0.067 | Robustness test |
| H45-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 45 | 18 | 0.0883 | 0.057 | Robustness test |
| H50-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 50 | 18 | 0.102 | 0.049 | Robustness test |
| H55-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 55 | 18 | 0.115 | 0.043 | Robustness test |
| H60-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 60 | 18 | 0.129 | 0.039 | Robustness test |
| H65-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 65 | 18 | 0.143 | 0.035 | Robustness test |
| H69.7-FL18 | Nikon D3100, AF-S DX NIKKOR 18–55 mm | 69.7 | 18 | 0.156 | 0.032 | Test sample, robustness test |
Seed size robustness test
To determine the robustness of our protocol and whether imaging height had consistent effects, 10 seeds each of A. thaliana, B. rapa, and M. guttatus were tracked using multiple camera configurations, including those used for the test sample above, and different combinations of camera heights and focal lengths on the Nikon D3100 camera and a Leica S8APO dissection microscope (Leica, Wetzlar, Germany) with a Leica Flexacam C1 (Tables 1 and 2). The 10 seeds for each species were fixed to parafilm on a microscope slide to prevent any displacement that may affect which face of a seed was captured, ensuring the identical layout was captured at each configuration.
RESULTS
Validation using a test sample
To calculate seed area (mm2) from the images, we used the empirical conversion scale outlined above. We applied an initial filter to remove outliers (e.g., soil or organic debris) that were identified as particles two orders of magnitude below the median seed size. We did not apply an SD filter to the data. The distributions of seed area are shown in Appendix S5; we applied a log10-transformation to account for the right-skewness of the data (at all camera configurations; Figure 2).
Where x is the seed area in number of pixels and f(x) is the log10-transformed seed area. This formula best fit the transformed mean seed area represented as a function of pixel count at each configuration and aligned with the expectation that misidentified edge pixels would be of decreasing importance at higher pixel counts. Variable a represents the scalar of the equation, and b is the vertical asymptote. As the number of pixels decreases, the metric size of the particle increases due to edge pixel effects. The horizontal asymptote of the model, c, represents unlimited resolution and the area value that would not include any edge pixel misidentifications, and thus is a theoretical representation of the true area that could be derived from the image. In this case, the area described by the asymptotic value, back-transformed to metric space, was 0.143 mm2 with a 95% confidence interval (CI) of 0.141–0.145 mm2. This represents the best estimate of the true mean size of the sample.
The camera configuration with the greatest resolution was H30-FL55 (where H is the sensor height above the seeds [in centimeters] and FL is the focal length [in millimeters]). This camera configuration had a sample mean seed area of 0.145 mm2 (95% CI: 0.141–0.149; Table 3, Figure 3). We compared all camera configurations using a linear model on the log10-transformed seed area, and none were significantly different from one another, except those involving H69.7-FL18 (the lowest magnification), which was significantly different from all other configurations.
| Model/camera configuration | Log10-transformed area | Back-transformed area (mm2) | 95% CI (mm2) | Mean seed area (px) |
|---|---|---|---|---|
| Inverse model | −0.845 | 0.143 | 0.141–0.145 | — |
| H30-FL55 | −0.839 | 0.145 | 0.141–0.149 | 600 |
| H35-FL55 | −0.830 | 0.148 | 0.144–0.152 | 359 |
| H40-FL55 | −0.837 | 0.146 | 0.141–0.150 | 235 |
| H45-FL55 | −0.842 | 0.144 | 0.140–0.148 | 166 |
| H50-FL55 | −0.839 | 0.145 | 0.141–0.149 | 124 |
| H55-FL55 | −0.846 | 0.143 | 0.138–0.147 | 95 |
| H60-FL55 | −0.848 | 0.142 | 0.138–0.146 | 76 |
| H65-FL55 | −0.841 | 0.144 | 0.140–0.149 | 63 |
| H69.7-FL55 | −0.833 | 0.147 | 0.143–0.151 | 54 |
| H69.7-FL18 | −0.625 | 0.237 | 0.232–0.242 | 10 |
Seed size robustness test
To assess the robustness of our protocol across multiple individual measurements, we tracked the area of 10 seeds at all camera configurations. We used both a Leica Flexacam C1 camera affixed to a Leica S8APO dissection microscope to capture each of the seeds at greater magnifications than possible with the Nikon camera, resulting in images composed of orders of magnitude more pixels. We delineated the area of each individual seed in ImageJ and scaled the sizes according to the empirical conversion value derived in the method above. We used the seed area attained from the microscope images as the reference value against which the camera images would be compared.
We converted each seed pixel count to a metric area using conversion ratios for the appropriate image configuration. Seed areas are shown in Figure 4, with the robustness evident in the relative size and rank order of the replicate seeds at the different imaging configurations. The reliability of the measurements is high for all configurations except for those with the lowest magnification and pixel count. Specifically, measures of B. rapa seeds are robust until the final camera configuration, while measures of the smaller seeds of A. thaliana and M. guttatus become unreliable between the configurations of H28-FL18–H60-FL55, and H40-FL18–H45-FL18, respectively (Figure 5). At these threshold configurations, the mean number of pixels for A. thaliana (H60-FL55) and M. guttatus (H45-FL18) are 44.6 px and 19.0 px, respectively, while the mean number of pixels at the lowest magnification of B. rapa seeds (H69.7-FL18) is 83.0 px. The converted area of A. thaliana and M. guttatus seeds tends to increase at lower pixel values, with particular overestimation of the size of M. guttatus seeds with camera configuration H65-FL18. For raw data of seed pixel counts at each camera configuration, see Appendices S6, S7, and S8 for A. thaliana, B. rapa, and M. guttatus, respectively. The pixel counts across species are compared in Appendix S9.
DISCUSSION
Our study presents an efficient and standardized technique for using digital photographs to estimate seed traits, including seed number, size, and color. We applied the protocol to assess the accuracy of seed size estimates in three model plant species and found broad tolerance for a range of camera configurations. Our findings indicate that increasing the focal distance enhances the accuracy of seed size estimates; however, the improvement plateaus and compromises the ability to image large numbers of seeds at once. We demonstrate the empirical scale necessary to accurately estimate seed area for particles of specific sizes and connect this to the parameters of the imaging system (i.e., camera configuration). We recommend a conservative minimum pixel count of 100 px for imaged particles to ensure accuracy. In the following sections, we elaborate on the significance and future applications of the protocol.
Our results showed that seeds of M. guttatus have a right-skewed distribution, even after log10-transformation. The seed we used came from field-collected seed capsules, so the distribution could be caused by various factors including seed inviability or seed abortion due to inbreeding (e.g., Martin and Lee, 1993) or hybrid incompatibility (e.g., Coughlan et al., 2020). Mimulus guttatus can produce more than 1300 seeds per seed capsule (Waser et al., 1982; Searcy and Macnair 1990), and the relatively small seeds could result in right-skewness in the distribution of size.
Importantly for the accuracy and robustness of the protocol, the distribution and mean seed sizes remained consistent across all camera configurations, with the exception of H69.7-FL18 (i.e., camera height = 69.7 cm, focal length = 18 mm). This exception is most likely a result of poor resolution due to the camera configuration capturing insufficient pixel counts to precisely characterize seed areas. Taking that into consideration and omitting H69.7-FL18 from interpretation, we find that the mean seed area remained consistent across camera heights and focal distances. Therefore, it appears that most camera distances can robustly capture seed size in M. guttatus.
We tested our protocol on three model species—A. thaliana, B. rapa, and M. guttatus—and found the method to accurately and consistently measure seed size at most camera heights and focal lengths. However, error was introduced at species-specific camera configurations that depended on the size of their seeds. The breakdown occurs as camera magnification results in insufficient pixels to precisely resolve differences in imaged propagule area. For A. thaliana, we found that a camera height of 55 cm and a focal length of 55 mm was the threshold distance for accurate seed size measurements, and that any camera heights that were lower and of the same focal length did not differ and therefore were all appropriate for accurate estimation of seed size. The threshold distance for estimating seed size in M. guttatus was higher due to larger seed size relative to A. thaliana and could be estimated accurately with a camera distance up to 60 cm and a focal length of 55 mm. Finally, because B. rapa produces relatively large seeds, all camera distances and focal lengths produced similar seed size estimates. Notably, we also found that for all three species “manual” estimation of seed size (i.e., estimating seed size with a microscope) was not significantly different than seed size estimated with an appropriate camera distance and focal distance ranges described above, thereby validating our technique.
To select an empirical scale for an untested species, we developed a reference heuristic model that facilitates the selection of an appropriate camera configuration to image novel particles of an approximate area at 100 px. The simple model accounts for the empirical scales available according to our camera parameters and configurations and could be produced for any imaging bench.
Although the technique developed here is targeted to seeds, it could be used to estimate the number, size, and color of any small particle. For example, ImageJ has been used to manually estimate pollen number and size in flowering plant species (Costa and Yang, 2009; Kakui et al., 2021), colocalization of organelles within cells (Stauffer et al., 2018), and the abundance of scrub typhus cells multiplying under a microscope (Siritantikorn et al., 2012). The protocol developed here could be applied to a variety of particles and taxa.
CONCLUSIONS
Seed number and seed quality are crucial factors affecting fitness in flowering plants. However, a standardized and efficient technique for estimating these traits is lacking. Our study highlights the importance of validation and robustness when estimating seed size to make accurate comparisons between individual propagules or samples from populations. We developed customized methods for estimating seed size in three model species and found that while certain camera distance thresholds are necessary for capturing accurate seed size measurements, there is considerable flexibility. Our study demonstrates that it is possible to accurately estimate the number, size, and color of most particles efficiently using bulked seed imaging.
AUTHOR CONTRIBUTIONS
C.S., J.L., and J.F. conceived the project; C.S. and J.L. collected the data; C.S., J.L., and J.F. wrote the manuscript; J.L. developed the software and code; J.F. oversaw the methods and writing. All authors approved the final version of the manuscript.
ACKNOWLEDGMENTS
The authors thank colleagues from Queen's University, including I. Lewis and E. Gillette for assistance with photographing seeds, C. Smith and A. Van Natto for insights on statistical methods, and J. Monaghan and A. Rooke for supplying Arabidopsis and Brassica seeds. This research was supported by the Natural Sciences and Engineering Research Council of Canada, through a Canada Graduate Scholarship (C.S.) and a Discovery Grant (J.F.).
Open Research
DATA AVAILABILITY STATEMENT
All supporting data are provided with the article and Supporting Information.
