Imaging and Analysis for Quantifying Maize (Zea mays) Abiotic Stress Phenotypes

Summary

A method is presented for low-cost imaging of plants to measure phenotypes, along with best practices for image capture and an image analysis pipeline for quantifying plant traits. These methods were applied to measure maize (Zea mays) phenotypes under heat, drought, and combined abiotic stress conditions.

Abstract

Measuring quantitative plant phenotypes, or traits, is essential for understanding plant responses to environmental factors, including biotic and abiotic stresses. However, traditional methods of measuring plants can be time-consuming, inaccurate, and destructive, particularly when assessing traits such as biomass and height on an individual basis. Image-based analysis offers an opportunity to reduce labor, improve accuracy and precision, and enable repeated measurements of the same plant over multiple time points. While high-throughput phenotyping facilities provide a solution, they are often prohibitively expensive and limited in availability worldwide. A method is described for capturing plant images using low-cost single-board computers and digital cameras within a photo studio setup, accompanied by an image analysis pipeline based on the free and open-source package PlantCV. This method offers a fast and accurate protocol for image capture, labeling, and transfer, along with best practices for ensuring high-quality image acquisition. Through image analysis, more than 10 phenotypes were quantified for all images in the experiment simultaneously, including traits such as plant size, height, and color. These methods were used to characterize the response of the Zea mays (maize) inbred variety B73 to heat, drought, and combined heat and drought abiotic stress conditions.

Introduction

Measuring plant traits, or phenotypes, is critical to answering fundamental and applied questions in plant science. For example, a researcher seeking to understand the impact of a genetic mutation on plant size will need to quantitatively measure plant height and leaf area¹. A researcher applying a biotic stress (such as bacterial, viral, or fungal infection) or an abiotic stress (such as drought, heat, cold, or nutrient depletion) will need to investigate the effect of the stress on plant color, or "greenness", to measure chlorosis^2,3.

While measuring these traits is essential to basic plant biology as well as applied plant breeding, measurements of these traits by hand can be time-consuming, imprecise, and inaccurate^4,5,6,7. For example, measuring plants for height includes manually using a ruler for each plant, oftentimes in hot or humid environmental conditions, and is subject to error from the viewpoint of the person measuring, as well as what point is considered the "top" of the plant. Furthermore, measurements such as biomass and pigment content are destructive, and the same plant cannot be measured for change over time or used for multiple samplings because it must be destroyed in the process of making these measurements^4,5,6,7. The time-consuming, error-prone nature of these measurements hinders progress in evaluating plants for their phenotypic and physiological stress responses.

Capturing images of plants and measuring quantitative traits from the image provides solutions to these problems^4,5,6,7. Image analysis improves accuracy and precision, while reducing the time a researcher has to make the measurements, and the same plant can be imaged multiple times or used for other samplings due to the non-destructive trait quantification^4,5,6,7. Furthermore, images can be stored and re-analyzed long after an experiment is complete and plants are disposed of^4,5,6,7. Image analysis can be completed individually with low-throughput image editing tools, such as ImageJ (Fiji)⁸, or in higher throughput using computer vision and machine learning with platforms such as PlantCV⁹. PlantCV is a free, open-source image analysis platform that enables researchers to first mask (or separate) a plant or other object from the background of an image, and make measurements of size, morphology, color, and more of that plant or object⁹. Using a modular workflow, a researcher first builds a flexible workflow (or pipeline) of different modules with a sample image, then subjects the remaining images to the high-throughput workflow without interacting with each individual image⁹. Quantitative traits are output and analyzed for statistically significant differences between groups. Deep Learning pipelines can be developed for efficient, automated image analysis after enough images have been captured and labeled with appropriate trait measurements or metadata¹⁰.

"High-throughput" phenotyping can increase the speed and reduce human labor during data collection, and/or during data analysis^4,5,6,7. High-throughput phenotyping facilities have enabled the automation of image collection in controlled environments, where plants move on conveyor belts or via robots to an imaging station, and are automatically labeled with the name (i.e., label with metadata) associated with the plant^{6,11,12,13,14}. While these conveyor belts or robot-enabled facilities enable the capture and labeling of large amounts of images without human intervention, there are few in the world, and they can cost millions of dollars to build, making them inaccessible to many researchers. To apply these concepts of automated imaging at a lower cost, Raspberry Pi (here referred to as single-board computers, SBC) are low-cost computers (less than $35 each) that have been used to capture, name, store, and transfer images of plants grown in controlled environments. These setups capture images in automated, timed intervals using an array of up to 72 Raspberry Pi^12,15,16. While more cost-effective than conveyor belt systems that can cost millions of dollars, previously described SBC growth chambers are most applicable to a top-down view of growing small plants and still require numerous SBC¹⁵. In field environments, gantries, UAS (uncrewed aerial systems) such as drones, and satellites all capture images for analysis of plant traits^17,18. Both field systems and high-throughput controlled environment facilities can be prohibitive in their upfront cost, maintenance, and expert knowledge necessary to build and maintain the systems, as well as analyze images^17,18. Cell phones and digital cameras used alone for plant photography present time-consuming challenges of image transfer and naming with metadata, which is essential for downstream image analysis.

The methods described here present a low-cost solution to plant imaging for the challenges unaddressed by other systems. A photography setup, where plants are transported individually to the studio, allows for the efficient measurement of plants of all sizes in pots. Using a SBC connected to a digital camera, this method first provides standardized capture, labeling, and transfer of images. Using a keyboard or a QR/barcode scanner to name images as they are being captured circumvents the time-consuming and error-prone activity of renaming images with metadata after they have been captured and stored. The composition of the image and recommended camera settings are important for making downstream image analysis possible, and guidelines are detailed here. Finally, an image analysis pipeline in PlantCV is provided to extract quantitative traits from the captured images, such as leaf area, plant height, width, and color.

Relative to conveyor belt systems, this method is low-cost and requires a single SBC. The phenotyping setup described here was constructed to remove the constraints and costs of a conveyor belt or robotic systems - any size plant can be measured because the setup can be separated from the growth space, only a single SBC is needed, and it works with numerous camera types for maximum flexibility and affordability. Similar approaches to the one presented here include the PhenoBox or PhenoRig^12,19. Compared to manual measurements, this method saves time, reduces errors, and does not destroy the plants^4,12,15,19.

Protocol

The details on the reagents and the equipment used in this protocol are listed in the Table of Materials.

1. Preparation of photo studio

1. Hang black photography-grade fabric from a wall or ceiling for side-view imaging. Drape fabric at least 0.5 m onto the floor so that the plant is placed on top of the fabric (Figure 1).
   NOTE: Fabric should be at least 100 cm taller and wider than the plant to be imaged.
2. Place a color card on the backdrop, propped up against an object such as a box so that it is visible in the frame (Figure 1). For plants larger than 30 cm tall or wide, use a color card that is at least 20 cm wide. For plants smaller than 30 cm tall or wide, use a color card that is between 4-8 cm wide.
3. Plug an SD card into a laptop or desktop computer. Download Raspberry Pi Imager from the software section on the website (https://www.raspberrypi.com/software/), and download the provided raspberrypi-image-capture.img.gz file (https://github.com/danforthcenter/photo-maize-paper) to the laptop or desktop computer.
4. Launch the Imager by opening the downloaded imager (.exe) file. Click on the Choose OS button and select the Use Custom button at the bottom. Select the image-capture.img.gz file.
   1. Next, click on the Choose Storage button and select the SD card plugged into the computer. Click on the Write button.
      NOTE: Ensure the correct SD card has been selected, as this process will delete all data stored on the card.
5. Connect the latest version of SBC to a keyboard using a USB cord, to an external monitor using an HDMI cord, and to a power strip using a power cord with a switch. Connect the monitor to the power strip using the power cord provided with the monitor. Place the SD card into the SBC. See Figure 1 for a labeled schematic of the setup.
6. Turn on the power strip by pushing the switch on the strip, turn on the monitor using the button on the monitor, then turn on the SBC by flipping the switch on the power cord to "on".
7. Connect the SBC to the local internet using WIFI or an ethernet cable.
8. If photos are to be transferred directly to a server or cloud storage:
   1. Edit the file "CAPTURE-PHOTO.py" on the SBC to contain the file path to complete the transfer.
   2. Edit the win-credentials file with the path and password for the server or cloud storage using the following command in the terminal: sudo nano /etc/win-credentials.
      NOTE: Contact the internet system administrator to ensure the folder has permission to transfer files.
9. Connect a wireless barcode scanner to the SBC using manufacturer's instructions provided with the barcode scanner, either using a USB or Bluetooth.
10. Turn off the SBC by clicking on the Raspberry Pi button on the upper left of the screen, clicking on Logout, and then clicking on Shutdown to safely shut off the SBC. After shutdown, switch the SBC power switch "on" the power cord to "off" and the power strip to "off".

2. Plant growth and stress treatment

NOTE: Any plant is suitable for this method of phenotyping; these plant growth methods are applicable to the representative results presented here.

1. Obtain seeds of B73 or another maize variety from a stock center.
2. Prepare soil media and growth conditions:
   1. Fill 3.5-inch pots (0.7 L) with a 1:1 by volume mix of Turface MVP and Berger BM7-35 with 3 lb/cubic yard of slow-release fertilizer 14-14-14 (N-P-K). Place pots in a tray with holes and water thoroughly with reverse osmosis (RO) water.
   2. Set a growth chamber to the following environmental conditions: 30 °C day/20 °C night, 16 h day length, 60% humidity, and 500 mmol.m^-2s^-1 of light at the level of the pot during the day. Set a separate growth chamber with the same conditions, except 38 °C day/28 °C night for heat stress.
3. Plant 20 maize seeds, each in a 3.5-inch pot, by placing ~2 cm deep in the soil and covering it with the displaced soil. Water thoroughly with RO water and place in the control growth chamber. Water daily with liquid fertilizer (mixed according to the manufacturer's instructions; see Table of Materials).
   NOTE: Planting extra plants that are needed for the experiment will account for poor germination and lack of uniformity prior to the start of a stress application. Germination rate will vary by variety.
4. After 21 days of growth, transplant 16 of the most uniform plants (n = 4 plants per treatment) to 3.79 L pots filled with the same potting soil mix, pre-wet with RO water. Return plants to the control growth chamber.
5. On day 25 (4 days after transplanting) move half of the plants (n = 8) to the heat stress chamber at the time lights turn on in the chamber. Stop watering the drought-stressed plants in both chambers (n = 4 per chamber).
   NOTE: Drought stress plants may need to be watered so that the plants survive until the end of the experiment; here, drought plants were watered on day 31 and day 37 with fertilizer water.
6. On day 38 (13 days after the start of the stress), remove plants from the growth chambers for imaging.

3. Plant imaging

1. Start up the SBC (see Figure 2):
   1. Turn on the power strip using the switch on the power strip.
   2. Turn on the monitor using the button on the monitor.
      NOTE: If the SBC is turned on before the monitor is turned on, it will not start correctly.
   3. Turn the switch on the SBC power cord to "on".
      NOTE: This will turn on the SBC, and it will boot up on the monitor; there is no "on" switch on the SBC itself, only on the power cord.
2. Place the camera on a tripod or stable table.
   NOTE: The entire plant and color card should be in view on the camera screen. The camera should be perpendicular to the ground and have a direct view (no angle) of the center of the plant. The camera must be compatible with the gphoto2 package, and a list can be found here: http://www.gphoto.org/proj/libgphoto2/support.php. 
3. Provide power to the camera via a power cord with the camera or batteries. Turn the camera on using the button on the camera.
4. Connect the camera to the USB port on the SBC using the data transfer cable provided with the camera.
5. Arrange plant in photography setup:
   1. Place a piece of dark-colored tape on the photography-grade fabric where the pot will be placed to ensure the placement is the same for each plant. Ensure the color card is in line with the piece of tape.
      NOTE: Place the color card at least 10 cm from the edge of the plant. During quality control in the PlantCV workflow (below), the image should be able to be cropped to include only the plant and pot but not the color card to ensure that the color chips are not included in the ultimate trait extraction. See Figure 2 for examples of high and low-quality images.
   2. Place the pot on the piece of tape. For plants that grow in a plane, such as maize (Zea mays) and sorghum (Sorghum bicolor), face the widest angle of the plant towards the camera. For plants that do not grow in a plane, capture 0- and 90-degree angles of the plant, and average the resulting quantitative measurements after image analysis is complete.
6. On the SBC desktop screen shown on the monitor, open the terminal by clicking on the terminal button on the upper left menu bar.
7. If saving directly to a server or cloud storage, mount the storage using the following command by editing to the necessary cloud storage path (cloud-storage-path in the below example) and password, typing the line of code into the open terminal window, and hitting enter on the keyboard. If unchanged from factory settings, the user ID (UID) will be "pi".
   NOTE:
   sudo mount -t cifs -o credentials=/etc/win-credentials,uid=pi //cloud-storage-path /mnt/cloud-storage
8. Type the following code into the terminal and hit enter on the keyboard. A new window will appear.
   ./RUN-PHOTOSTUDIO-SHELL.sh
   NOTE: The RUN-PHOTOSTUDIO-SHELL.sh file calls the CAPTURE-PHOTO.py file, which are pre-loaded onto the pi if it was loaded using the above method. Alternatively, these scripts have been provided on GitHub for putting onto an existing Raspberry Pi SD card. 
9. A graphical user interface window will open. Type in the name of the image, or scan the barcode or QR code of the plant using a barcode scanner.
10. Select an option to save the image locally or to a mounted server/cloud storage.
11. Click Click here to take a picture! button.
12. If local storage was chosen, open the PHOTOS folder on the desktop to examine the image. If a server or cloud storage was chosen, open the image in that location.
13. If necessary, adjust the settings on the camera by unplugging it from the SBC, adjusting using the manufacturer's instructions, and plugging it back into the SBC.
    NOTE: Settings such as exposure, brightness contrast, etc., may need to be adjusted depending on the lighting in the room and the object being imaged. A common mistake is to over-expose the images. To check the exposure of a test image, proceed with the test image through the PlantCV image analysis workflow, which includes a quality control check for exposure. Adjust the camera settings as needed to reduce or increase the exposure until the histograms are appropriately distributed.
14. After all images have been captured, transfer the images from the SBC to a local computer or cloud storage using the preferred method, such as USB storage, internet browser, or SSH transfer.

4. Trait extraction from images

1. Download the following files to a local computer or server: "photo-studio-SV-notebook.ipynb", "workflow.py", and "photo-studio-SV-config.json". All files are available on GitHub (https://github.com/danforthcenter/photo-maize-paper).
2. Install PlantCV on the local computer or server using the steps described (https://plantcv.readthedocs.io/en/latest/installation/). 
3. Open the file photo-studio-SV-notebook.ipynb using the preferred code editor, such as JupyterLab or Visual Studio Code.
4. Run each block of code, editing as necessary given the parameters outlined in the code block to obtain a clean mask of only the plant in the image.
5. When satisfied with the analysis of the sample image, open the file workflow.py in the preferred code editor. Edit this file with any parameters changed in "photo-studio-SV-notebook.ipynb" and save it.
6. Open the file photo-studio-SV-config.json and edit the file paths to direct to a folder of input images, a desired output folder, etc., as denoted in the file.
7. Open a terminal and run the following code to activate PlantCV. To do so, type the code conda activate plantcv into the terminal and hit enter on the keyboard.
8. Change the following line of code to the file path of the "photo-studio-SV-config.json" file (this is the only change to make in the below code).
   1. In the terminal, type the edited code and hit enter on the keyboard. This will run the workflow over all of the images captured in the experiment.
      NOTE:
      plantcv-run-workflow --config /Users/USERNAME/photo-studio-SV-config.json
9. Inspect the output images to determine if the code was appropriate for all images.
   NOTE: The blue and pink outline of the plant should be around the edges of the plant, and there should be no other objects in the image. If necessary, go back to the "photo-studio-SV-analysis.ipynb" and test problematic images and edit the code, specifically the masking steps and noise reduction, to obtain the desired results. See Figure 3 for examples of high, acceptable, and low-quality image segmentation.
10. The output file will be in the form of a .json file. Go back to the terminal (ensure conda activate plantcv has already been run), type the following code, and hit enter on the keyboard to convert the file to .csv for downstream statistical analysis, changing the file path to that of the "results-photo-studio.json" file.
    plantcv-utils json2csv -j /Users/USERNAME/results-photo-studio.json -c /Users/USERNAME/results-photo-studio.csv
    NOTE:  To convert the measurements from pixels to a standard unit of measure, such as cm, utilize the conversion provided from the find_color_card function in the PlantCV workflow.
11. Perform statistical analysis of the output quantitative traits to determine differences in plant phenotypes due to different plant species, varieties, treatments, and timepoints.
    NOTE: The pcvr package in R Studio has been formulated to analyze the outputs from PlantCV for statistical significance and plotting of results²⁰. See a summary and visual representation of traits collected in Figure 4.

Results

To develop this method, B73 maize plants were evaluated for quantitative phenotypes under control temperatures, heat temperatures, well-watered, and drought conditions, as well as their combinations, for a total of four treatments. B73 is a well-known inbred maize variety with a wealth of genetic data²¹.

Image analysis in PlantCV successfully color-corrected the images, labeled the quantitative results with the plant metadata, and segmented (in other words, separated) the plant from the background, as is demonstrated in Figure 3. An unsuccessful analysis would provide output images that do not show the blue outline or pink shape around the outside of the plant, but rather include additional noise or missing pieces of the plant (Figure 3, low-quality example). A dual-channel threshold was necessary to include as much of the plant as possible in the mask (see alternative single-channel strategies in the "photo-studio-SV-notebook.ipynb"). However, some plants, particularly those with purple-colored stems, did not have a perfect segmentation, but were sufficient for downstream analysis (Figure 3).

The PlantCV analysis provided 16 quantitative, single-value traits, summarized in Figure 4. A single-value-trait provides one value per plant - for example, the leaf area, height, width, and hue circular mean. Leaf area was considered using the "analyze.shape" function (Figure 4) and considered the entire plant. Height and width, however, were considered using the "analyze.bound_horizontal function", with a bound drawn at the plant soil line. This is because plants that droop over the side of the pot will have a greater height considering the plant as a whole object, and instead should only be measuring from the plant base where it meets the soil. In addition to the single-value traits, PlantCV outputs multi-value traits, which are histograms, particularly of the color values, representing each pixel of the plant. The hue circular mean is a single-value-trait that is the average hue value for the whole plant.

To determine which traits are the most interesting, the variance explained by treatment for each trait was calculated, as well as the correlation between the different traits using the "frem" function in the pcvr R Studio package²⁰ (Figure 5A). Leaf area, height, width, and hue circular mean were selected for downstream analysis because they explained greater than 50% of the variance due to treatment, and are relevant measures of plant growth (area, height, and width) and health (hue). Representative images of the plant with the largest and smallest leaf area demonstrate the variation in phenotypes collected in this experiment (Figure 5B,C, respectively). The plant with the largest leaf area also had the highest mean hue and was a B73, well-watered, heat-stressed plant (Figure 5B). The smallest leaf area had the smallest mean hue (Figure 5C), and was a B73, drought-stressed, heat-stressed plant.

Water treatment had a significant effect in determining leaf area (F(1,13) = 226.5, p = 1.32 x 10^-9), height (F(1,13) = 21.1, p = 0.0005), width (F(1,13) = 75.5, p = 8.92 x 10^-7), and hue circular mean (F(1,13) = 27.8, p = 0.0002), determined using a linear regression model for treatment effect and a two-way ANOVA in R Studio (complete results in Table 1). Temperature treatment had a significant effect on height (F(1,13) = 5.94, p = 0.03), but not leaf area, width, or hue circular mean (p > 0.05, Table 1). The interaction between temperature and water status was not a significant factor in any of the traits measured (p > 0.05, Table 1).

Next, individual treatments were compared to the control and each other using linear regression and post-hoc test (estimated marginal means) with a Sidak correction for multiple comparisons (Figure 6). Leaf area, plant height, plant width, and hue circular mean were significantly reduced in drought-stressed plants compared to well-watered under both temperature conditions (p < 0.05, Figure 6). Heat stress while well-watered only reduced plant height (p < 0.05, Figure 6), but did not significantly reduce leaf area, plant width, or hue circular mean. This difference in hue can be further investigated by assessing the multi-value color trait hue, plotted as a histogram in Figure 7, rather than just an average. The reduction in mean hue due to drought was due to a shift from green to yellow pixels, known as chlorosis (Figure 7). While heat did not show a significant difference in average hue, it did exhibit a reduction and increase due to both yellowing (chlorosis) and darkening of the green color (Figure 7).

After imaging, the plants were cut at the root-shoot junction, and the weight of the above-ground tissue was recorded. The roots were gently washed and dried with a towel, then weighed, to calculate the total plant biomass (also known as weight) for comparison to the image-based phenotyping method described. Leaf area measured via image analysis strongly correlated to plant biomass (R² = 0.84, Figure 8).

Figure 1: Image of the photo studio setup with labeled components for collecting plant images. (A) Photography-grade fabric, (B) color card, (C) line of tape for placing pots uniformly, (D) digital camera, (E) single-board computer, (F) monitor displaying image capture process. Please click here to view a larger version of this figure.

Figure 2: Overview of the imaging workflow. A representative "high quality" and "low quality" image is provided. Please click here to view a larger version of this figure.

Figure 3: Overview of the image analysis workflow in PlantCV. Representative images of "high", "acceptable", and "low" quality segmentation are presented; PlantCV traits are represented using a pink outline of the whole-plant shape, and a blue outline of the plant area. Please click here to view a larger version of this figure.

Figure 4: Summary and visual representation of the shape traits captured from analysis. Shape function in PlantCV. Purple boxes represent traits that do not utilize another trait in their calculation. Orange boxes represent traits that utilize other traits in their calculation; arrows indicate which traits are used in these calculations. Please click here to view a larger version of this figure.

Figure 5: Image-based phenotyping captures phenotypic variance and correlated traits. (A) Variance of traits explained by treatment (right), and correlation of traits to each other (left). Each trait is labeled, followed by its unit of measurement; "NA" means it is a unitless measurement. (B) Image of the plant that had the highest leaf area and highest hue circular mean (degrees). (C) Image of the plant that had the smallest leaf area and lowest hue circular mean (degrees). Please click here to view a larger version of this figure.

Figure 6: Boxplot comparison of treatment effects on plant phenotypes. Leaf area (A), height (B), width (C), and hue circular mean (D) are compared for the effect of drought, heat, and the combined stresses. n = 4. Letters represent statistically significant differences between treatments (p < 0.05); treatments were compared using linear regression and post-hoc test (estimated marginal means) with a Sidak correction for multiple comparisons. Please click here to view a larger version of this figure.

Figure 7: Histograms of color (hue) summed of all pixels of each treatment combination, normalized by the total number of pixels in the treatment combination. Histograms are colored by the color value for the hue degree. Please click here to view a larger version of this figure.

Figure 8: Correlation of leaf area measured from images to whole-plant biomass. Each black point represents one plant. The pink line represents the linear correlation between leaf area (X) and biomass (Y). Please click here to view a larger version of this figure.

Trait	Factor	degrees of freedom	F-value	p-value
leaf area (cm^2)	Temperature	1	0.5372	0.4766
leaf area (cm^2)	Water	1	226.5318	1.32E-09
leaf area (cm^2)	Temperature:Water	1	2.1526	0.1661
leaf area (cm^2)	Residuals	13
height (cm)	Temperature	1	5.9426	0.0298935
height (cm)	Water	1	21.107	0.0005029
height (cm)	Temperature:Water	1	2.1752	0.1640545
height (cm)	Residuals	13
width(cm)	Temperature	1	2.2495	0.1575
width(cm)	Water	1	75.5474	8.92E-07
width(cm)	Temperature:Water	1	0.0839	0.7766
width(cm)	Residuals	13
hue circular mean (degrees)	Temperature	1	0.9154	0.3561432
hue circular mean (degrees)	Water	1	27.7988	0.0001509
hue circular mean (degrees)	Temperature:Water	1	3.4792	0.0848724
hue circular mean (degrees)	Residuals	13

Table 1: Results of the linear regression model and two-way ANOVA analyses examining the effects of water treatment, temperature treatment, and their interaction on plant traits.

Discussion

Image-based phenotyping provides accurate, fast, quantitative measurements of plant phenotypes^4,5,6,7. This method is significantly faster, more precise, more accurate, and provides a greater number of measurements than manually using a ruler to record plant height, or qualitative measures of plant color features that may relate to health^4,5,6,7. Leaf area was seen here, and in previous research¹⁰, to strongly correlate to plant biomass (Figure 8), suggesting this measure could reliably replace the time-consuming, destructive, and error-prone process of manually weighing whole-plant biomass. While a camera alone could be used for capturing the images, the addition of the low-cost SBC increased throughput through automation of image collection and image labeling with necessary metadata¹⁵. This method is suitable for one to ~300 plants that are imaged once daily (or less frequently). For larger numbers of plants in an experiment, a conveyer-based, overhead gantry, or robot-enabled imaging system is more suitable since the time to move the plants individually for imaging may be longer than a working day^11,13. For more frequent timepoints (for example, an image every hour rather than once per day or less frequently), camera setups inside of a growth chamber are more suitable because they capture images automatically, but are limited by the size of the plants, and need additional SBC¹⁵.

While PlantCV is presented as the image analysis option here, other software exists. ImageJ (Fiji) is a free software package for image analysis that can also be used for images of plants through a graphical user interface (GUI)⁸; however, it is not particularly developed for whole-plant image analysis, is less adaptable to high-throughput computing, and is less easily repeatable than already-working pipelines like those in PlantCV that can be easily applied to new images⁹. Deep Learning (such as BioDock, see Table of Materials) and Segment Anything²²) methods are attractive for their ability to easily segment plants from the background, and could be used for the segmentation tasks described here. However, these Deep Learning models do not provide the plant phenotype measurements described here¹⁰, and to make these more complex measurements of quantitative phenotypes may require a large amount of training images (oftentimes greater than 10,000) and computing power beyond a basic laptop or desktop computer¹⁰. These models also require significant computer memory and processing power, which are unavailable to many researchers¹⁰.

The most critical step of the protocol is arranging the plant imaging setup and ensuring there are no objects (such as other plants, a color card, or plant debris) close to the plant of interest¹⁴. If images are not set appropriately, or a color card is not used, they will be impossible to analyze later because there is no mechanism for standardization. While no calibration steps are necessary in this protocol, color correction and size standardization during processing are essential. Always ensure the entire color card is in frame, but not overlapping with the plant (Figure 2). Another common pitfall is uneven lighting, which can cause shadows or bright shots; we recommend using bright, diffuse overhead light and covering windows in the room to prevent uneven lighting (Figure 2). Additional lamps can be added at a face-on angle to the plant if overhead lights are causing shadows. The type of lightbulb is not a concern, given that color will be corrected in the image using a color card. Blurred images are not acceptable for analysis - use auto-focus on the camera, or the manufacturer's instructions for manual focus to ensure the images are clear.

Similarly, ensuring images are not over-exposed (a common mistake) allows for appropriate analysis (Figure 2); over-exposed images cannot be fixed in a color correction during image analysis because the pixel values are beyond the camera's detection range with the current settings¹⁴. To prevent over-exposure, evaluate the color histograms or RGB values for a test image using the PlantCV workflow described here (Figure 3), and reduce the brightness setting on the camera if the exposure test determines the image is overexposed. The RGB values (and thus the histograms) should not have values at 0 or at 255, which is the value range because it is possible those values are above or below the range of detection and can only record the minimum or maximum rather than the true value. A test image should always be evaluated using these methods for over-exposure (Figure 2 and Figure 3); we suggest utilizing the PlantCV exposure quality control provided in this publication as a check, which determines a quality image to have less than 5 percent of pixels at 0 or 255 for each color. As a rule of thumb, cameras and analysis software provide histograms, which should have equal and complete histogram tails; values cut off below 0 (underexposed) or above 255 (overexposed) should be re-taken with adjusted camera settings. We suggest choosing a location with bright, even, artificial lighting with no windows for illumination, as windows can be too bright, uneven, and frequently change light exposure to the space. Images should appear evenly lit, and the gray squares of a color card should not appear white, which suggests over-exposure.

Troubleshooting may be necessary for the image capture code depending on the IP and security settings of the user's institution. To troubleshoot, contact the IT department for internet connectivity and file transfer issues. A common mistake is using the wrong size SD card for the SBC, or turning on the SBC before the connected monitor is turned on. To troubleshoot, follow the above protocol precisely, and re-load the SBC software. Online forums are an excellent resource for troubleshooting and finding solutions for the specific camera of choice when using gphoto2. To troubleshoot PlantCV image analysis and use the most up-to-date versioning, utilize the documentation page (https://plantcv.readthedocs.io/en/latest/), or post an issue on GitHub (https://github.com/danforthcenter/plantcv/issues). In particular, if the thresholding method in this protocol does not sufficiently separate the plant from the background, utilize alternative thresholding methods provided in “photo-studio-SV-notebook.ipynb” or on the documentation page.

This method can be applied to any plant species and treatment that can be transported to the assembled phototography setup described. For example, it is not applicable to plants growing in a field, since they cannot be transported to the photography setup without digging up the plant. In this situation, plants could be cut down and subsequently imaged using the protocol, although the measurement would be destructive. Multiple plants can be captured in a single image and analyzed, but a user should beware of keeping plants from touching each other (Figure 2), and plants are placed in the same location on the backdrop each time to ensure the workflow can be run automatically over the entire set of images. If a top-down view is necessary for plants such as Arabidopsis thaliana, use a tripod to move the camera to the top of the plants placed on the photography-grade fabric; the remaining methods for image capture and analysis remain the same.

As demonstrated here with maize, this method is important for both plant breeding and fundamental insights into how plants tolerate abiotic stresses, such as heat and drought. Heat and drought stress combined contribute to 15%-20% yield loss in maize annually worldwide²³. Yield losses are only expected to increase with global climate change, as temperatures are expected to increase 1.4-4.4°C by the year 2100²⁴. Evaluating maize diversity for stress tolerance, as well as determining the phenotypic, physiological, and molecular mechanisms of abiotic stress tolerance, are critical research aims for global food security²⁵. To address these questions, researchers in all adjacent fields - physiologists, geneticists, breeders, molecular biologists, etc. - must be able to quickly, effectively, and often non-destructively evaluate the phenotypes of maize plants under abiotic stress²⁵. The method and example results shown here address this need, and are an important tool for researchers to address real-world agricultural challenges.

The phenotypes observed here were hypothesized, given a wealth of previous research - drought stress reduces plant size and greenness (measured by hue circular mean)²⁶ (Figure 6, Figure 7, and Table 1). The phenotype of heat stress with well-watering unexpectedly did not significantly reduce leaf area, although with more replicates, this relationship could be better evaluated statistically, and height was reduced as expected (Figure 6). While the hue circular mean was not changed due to heat, the color profile was indeed altered to include both more yellow and darker green pixels, showing that evaluation of just the mean hue is insufficient, and the multi-value-trait histogram of color should be considered (Figure 7). These phenotypic changes describe a plant's response to the environment, and additional techniques such as measurements of gas exchange, microscopy imaging, and more could be employed to assess the physiological changes that occur in response to stress. This result demonstrates the power of precise, quantitative measurements that would otherwise be missed using manual or qualitative measures of plant health. These insights are critical for breeding and engineering crops resistant to climate change, which significantly reduces global crop yields and is predicted to worsen in future years.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
15W USB-C Power Supply	CanaKit	DCAR-RSP-3A5-C
25604 UC-E6 USB Cable	Nikon	18208256044
3.5 inch pots	Hummert International	13006000
Berger BM7 35% Bark HP	Hummert International	10121500
Biodock 2024		AI Software Platform	www.biodock.ai.
Code editing software	Jupyter		https://jupyter.org/install
Code editing software	Visual Studio Code		https://code.visualstudio.com/download
Color Card ColorChecker Classic	Calibrite	CCC
Color Card ColorChecker Passport	Calibrite	CCPP2
Conviron Growth Chamber	Conviron	BDW80 model
Dell 22 Monitor	Dell	SE2222H
EH-67 AC Adapter	Nikon	AZ185
Image analysis software	PlantCV		https://plantcv.readthedocs.io/en/latest/installation/
Jack's Water Soluble Fertilizer 15-5-15	Hummert International	7590200	liquid fertilizer
Keyboard and Mouse	Logitech	1440638	MK270 wireless keyboard and mouse
Maize seed	USDA-ARS Germplasm Resources Information Network (GRIN)	B73	https://www.ars-grin.gov/
Micro HDMI to HDMI cable	Amazon	B07KSDB25X
Nikon Coolpix B500 camera	Nikon
Nikon D7200 camera	Nikon
Osmocote 14-14-14	Hummert International	7630000
Photography-grade fabric	Yayoya		Polyester Fabric Black Photo Backdrop for Photo Video Studio Televison
Raspberry Pi 4 Case with heatsink	iUniker	‎B07ZVKN262
Raspberry Pi 4 Model B Rev 1.1	Raspberry Pi Foundation	RP4B4GB_BP
Raspberry Pi 4 PiSwitch (USB-C)	CanaKit	RSP-PISWITCH-USBC
Raspberry Pi software	Raspberry Pi Imager		https://www.raspberrypi.com/software/
Statistical analysis software	Rstudio		https://posit.co/downloads/
Tripod	K&F Concept	T255A3+BH-28L
Turface MVP	Hummert International	10240000
Wireless Barcode Scanner	Netumscan	B08X4NMX2M	1D and 2D barcode scanner