-
Plant height, tiller number, and crown diameter were measured five times and shoot dry weight was measured four times during the experiment. Across sampling times, substantial changes of these traits were observed for individual pots, ranging from 21.0 to 157.0 cm for plant height, 3 to 61 for tiller number, 1.8 to 13.0 cm for crown diameter, and 1.1 to 183.1 g for shoot dry weight (Fig. 1). Moreover, all traits were significantly correlated with each other, with the highest correlation found between tiller number and shoot dry weight (r = 0.96***), followed by tiller number and plant height (r = 0.84***), shoot dry weight and plant height (r = 0.82***), shoot dry weight and crown diameter (r = 0.80***), and tiller number with crown diameter (r = 0.75***) (Supplemental Table S1).
Figure 1.
Boxplots of the manually collected traits of switchgrass at different times of plant growth under controlled environment conditions.
RGB image-based parameters
-
RGB images were collected on the day prior to manual trait collection during the experiment. The image-extracted measurements also varied considerably across sampling times and pots, ranging from 22.4 to 152.8 cm for side-view height (SHT), 443.5 to 13,590.6 cm2 for side-view convex hull (SCH), 97.8 to 5,193.9 cm2 for side-view projected area (SPA), 34.1 to 13.7 cm for side-view maximum width (SMD), 56.7 to 4,723.8 cm2 for top-view convex hull (TCH), and 22.1 to 1,328.4 cm2 for top-view projected area (TPA) (Fig. 2). These measurements were all significantly correlated with each other, including those derived from side-view, top-view, and between side-and top-view images (Supplemental Table S2). Specifically, the highest correlation was found between SPA and TPA (r = 0.97***), followed by SCH and SHT (r = 0.96***), SCH and SPA (r = 0.95***), SCH and TPA (r = 0.95***), SHT and SPA (r = 0.94***), SCH and TCH (r = 0.93***), TCH and TPA (r = 0.91***), and TPA and SMD (r = 0.91***).
Figure 2.
Boxplots of the red-green-blue image-extracted measurements of switchgrass at different times of plant growth.
Correlation and model between RGB image-based and manually collected measurements
-
The correlation between image-extracted and manually collected measurements is of importance in determining prediction accuracy made by HTP. In this study, we found that the manually collected traits were all significantly associated with the image-extracted measurements. Plant height was highly correlated with side-view measurements SHT (r = 0.996), SCH (r = 0.97), and SPA (r = 0.94) (Table 1). Plant height was well predicted by SHT and SCH, with R2 of 0.99 for SHT and 0.93 for SCH (Fig. 3). The root mean square error (RMSE) of the model was 3.56 for SHT and 10.6 for SCH (Fig. 3). Plant height was also well correlated with the top-view-image measurements such as TPA (r = 0.91) and TCH (r = 0.85) (Table 1). Although significant, plant height was less correlated with SMD (r = 0.74). Tiller number was strongly correlated with SPA (r = 0.93) (Table 1), followed by TPA (r = 0.91), SCH (r = 0.85), SHT (r = 0.84), and TCH (r = 0.77) (Table 1). Tiller number was relatively well predicted by SPA with R2 of 0.86 (Fig. 3) and RMSE of 4.45 (Fig. 3). A relatively lower correlation was observed between tiller number and SMD (r = 0.67). Crown diameter was significantly correlated with SCH (r = 0.85), SMD (r = 0.83), TCH (r = 0.82), SHT (r = 0.79), TPA (r = 0.79), and SPA (r = 0.76) (Table 1).
Table 1. Pearson correlation coefficients among manually collected traits and red-green-blue image-extracted measurements in switchgrass across sampling times.
Plant height Tiller number Crown diameter Shoot dry weight SCH 0.97*** 0.85*** 0.85*** 0.79*** SHT 0.996*** 0.84*** 0.79*** 0.82*** SPA 0.94*** 0.93*** 0.76*** 0.94*** SMD 0.74*** 0.67*** 0.83*** 0.37* TCH 0.85*** 0.77*** 0.82*** 0.55** TPA 0.91*** 0.91*** 0.79*** 0.86*** N = 97 for correlation analysis between plant height, tiller number and crown diameter with all images based on measurements; N = 31 for correlation analysis between shoot dry weight and image-based measurements. *, **, *** represent significance at P < 0.05, 0.01, and 0.001, respectively. Figure 3.
The selected models to predict the manually collected traits using red-green-blue image-extracted measurements across sampling times. RMSE, the root mean square error. N = 97 for plant height and tiller number analysis. N = 31 for shoot dry weight.
Shoot dry weight was correlated with SPA (r = 0.94) (Table 1). However, it was better predicted by SPA with a polynomial with R2 of 0.95, compared to the linear model with R2 ≈ 0.88 (Fig. 3). The RMSEs of the linear and polynomial models were 19.2 and 12.8 (Fig. 3). Shoot dry weight was also correlated with TPA (r = 0.85), SHT (r = 0.82), and SCH (r = 0.79), but SMD and TCH were less correlated with shoot dry weight (Table 1). We also analyzed shoot dry weight correlation by excluding the data from the final harvest date. As a result, correlations between shoot dry weight and RGB image-based measurements were improved by 4 to 26% (except for TPA) (Supplemental Table S3), increasing from 0.79 to 0.93 for SCH, 0.80 to 0.90 for SHT, 0.94 to 0.98 for SPA, 0.37 to 0.63 for SMD, and 0.55 to 0.77 for TCH.
-
The switchgrass cultivar Cave-in-Rock was used in the experiment, as it is commonly cultivated and highly adaptable to a wide range of soils and growing conditions. Seeds were sown in 32 bar-coded pots (23 cm diameter, 21 cm deep) filled with propagation potting mix (Sungro, Agawam, MA, USA). The pots were placed in a well-controlled growth chamber room in the AAPF at Purdue University (IN, USA) (Fig. 4a). Seven days after germination, one healthy plant was kept in each pot. Plants were grown under temperatures of 29 °C/26 °C (day/night) and photosynthetically active radiation of 800 μmol m−2 s−1 with a 16-h photoperiod. Plants were automatically watered as needed and fertilized with a soluble fertilizer (N-P-K, 15-5-15 Cal-Mag) (Scotts Inc., Marysville, OH, USA). After 38 d of establishment, manual trait collection started along with initiation of high-throughput imaging.
Figure 4.
(a) Growth chamber and (b) red-green-blue imaging tower in the Ag Alumni Seed Phenotyping Facility at Purdue University (IN, USA)
Manual trait measurements
-
Plant growth and biomass component traits were manually measured at 38, 45, 57, 69, and 78 d after planting (DAP), except that shoot dry weight measurement was omitted at 38 d. This ensured a range of plant growth stages to be characterized during the experiment. Plant height was determined from the soil surface to the top of the uppermost leaf blade. Tiller number was counted for each pot. Crown diameter was taken by measuring the widest point across plants approximately 2.0 cm above the soil surface. At each harvest, all tillers from the randomly selected pots were collected for determining shoot dry weight after drying at 80°C in an oven for 3 d. The remaining pots were kept for continuous measuring and imaging. There were 32, 32, 20, 13, and 7 pots for measurements of plant height, tiller number, and crown diameter at their respective DAPs. For shoot dry weight, there were 12, 7, 5, and 7 pots harvested at 45, 57, 69, and 78 DAPs, respectively.
High-throughput phenotyping
-
AAPF is an automated HTP facility that enables imaging-based, high-throughput, nondestructive measurements of crop traits. The plants were imaged each time prior to manual trait collection. For imaging, bar-coded pots in growth chambers at the AAPF were automatically transferred to the AAPF’s RGB imaging tower (Fig. 4b). A custom-made ARIS RGB imager (ARIS, Eindhoven, The Netherlands) equipped with a standard 5 Megapixel RGB camera (Basler Ace, Germany) and a monochrome camera (acA2440-20gm) for chlorophyll fluorescence image acquisition were used to automatically acquire side- and top-view plant images, which are illustrated in Fig. 5. The detailed description of image acquisition processes through multiple imaging booths at AAPF were described previously[10]. Briefly, the fluorescence images were leveraged in the early step of image processing, namely the image segmentation step. Using the fluorescence image of a plant, the black and white template of an image was established using the Otsu algorithm. This template was used to isolate the corresponding RGB plant from its background, leaving only the plant’s pixels in the cut out image. After the plant in the RGB image was segmented, the traits such as the width, height, and area were measured. We used the script developed in Python for computations.
Figure 5.
Depiction of the extraction of red-green-blue image-based traits. For side-view, the length of the green box is an estimate of height (SHT), the yellow dash line is the maximum width of the plants, the area of the blue polygon is convex hull (SCH), and the red outline of the plant is the estimated side projected area (SPA). For top-view, the area inside the blue polygon is convex hull (TCH) and the red outline of the plant is the estimated side projected area (TPA).
In each RGB imaging event, 12 predefined side-view images from 360 degree angles (0, 30, 60, 90, 120, 150, 180, 210, 240, 270, 300, and 330) and one top-view image were acquired for each pot. Several RGB image-based traits were extracted including SHT, SCH, SPA, SMW, TCH, and TPA. The descriptions of these parameters are provided in Table 2. All images were analyzed by using a proprietary image analysis pipeline provided by ARIS, which conducts image segmentation using the chlorophyll fluorescence image of a plant[10].
Table 2. Definition and abbreviation of each image-based high-throughput trait.
Trait Abbreviation Definition Side-view height SHT Length from the lowest to the highest point of the plant Side-view convex hull SCH Smallest polygon that contains all plant materials from the side Side-view projected area SPA Total side-view projected area obtained by outlining all plant materials Side-view maximum width SMD Maximum width of the plants from side to side Top-view convex hull TCH Smallest polygon that contains all plant materials from the top Top-view projected area TPA Total top-view projected area obtained by outlining all plant materials Statistical data analysis
-
The values from the image-extracted measurements obtained through 12 side-views from 0 to 330 degree angles were averaged for each pot at each sampling time for correlation analysis with manually collected traits. For the top-view, one image was collected for each pot at each sampling time and was used for correlation analysis. The data from different sampling times across all individual pots was pooled for determining Pearson correlation coefficients using the SAS program (version 9.4; SAS Institute, Cary, NC, USA). Seven data points were excluded from all correlation analysis due to a failure in obtaining a top-view image at a particular time during the experiment.
-
We thank Purdue Institute for Plant Sciences for facilitating controlled environment phenotyping research and thank Chris Hoagland for technical help and assistance in high-throughput phenotyping. We also thank Yunfei Gao, Mengxin Xu, Jia Tang, and Yankai Wang for their assistance in planting and harvesting plants.
-
About this article
Cite this article
Jiang Y, Yang Y. 2022. High-throughput phenotyping for plant growth and biomass yield of switchgrass under a controlled environment. Grass Research 2:4 doi: 10.48130/GR-2022-0004
High-throughput phenotyping for plant growth and biomass yield of switchgrass under a controlled environment
- Received: 23 February 2022
- Accepted: 29 June 2022
- Published online: 28 July 2022
Abstract: Switchgrass (Panicum virgatum L.) is a native and prominent perennial grass species used for feedstocks. High-throughput phenotyping of biomass component traits is desirable for switchgrass improvement and production. The objective of this study was to establish correlations between the manually measured traits and image-extracted measurements in switchgrass grown in a controlled environment. Red-green-blue (RGB) images from side- and top-views were automatically collected from the plants varying in growth stages for assessing their relationships with manually measured traits. Plant height, tiller number, crown diameter, and shoot dry weight were all significantly correlated with RGB image-based measurements including side-view height (SHT), side convex hull (SCH), side projected area (SPA), top convex hull (TCH), and top projected area (TPA). For a particular plant trait, a good prediction was observed based on an image-based measurement, including plant height and SHT (R2 = 0.992), tiller number and SPA (R2 = 0.86), crown diameter and SCH (R2 = 0.72), and shoot dry weight and SPA (R2 = 0.88). Plant height was also well predicted by SCH (R2 = 0.94) and SPA (R2 = 0.88). Overall, SHT, SCH, and SPA extracted from RGB images well predicted plant height, tiller number and shoot dry weight. The results demonstrated that the image-based parameters could be leveraged in quantifying the growth and development of switchgrass.
-
Key words:
- Biomass yield components /
- High-throughput phenotyping /
- RGB images /
- Switchgrass