A high-throughput and low-cost maize ear traits scorer

Abstract

In this study, based on automatic control and image processing, a high-throughput and low-cost maize ear traits scorer (METS) was developed for the automatic measurement of 34 maize ear traits. In total, 813 maize ears were measured using METS, and the results showed that the square of the correlation coefficient (R²) of the manual measurements versus the automatic measurements for ear length, ear diameter, and kernel thickness were 0.96, 0.79, and 0.85, respectively. These maize ear traits could be used to classify the type, and the results showed that the classification accuracy of the support vector machine (SVM) model for the test set was better than that of the random forest (RF) model. In addition, the general applicability of the image analysis pipeline was also demonstrated on other independent maize ear phenotyping platforms. In conclusion, equipped with image processing and automatic control technologies, we have developed a high-throughput method for maize ear scoring, which could be popularized in maize functional genetics, genomics, and breeding applications.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11032-021-01205-4.

Introduction

Maize (Zea mays L.) is one of the most important food and feed crops in the world (Chen et al. 2014). Maize ear traits, which include seed distribution, seed quantity, seed size, ear size, and ear shape, are important traits for genetic breeding research, as well as for physiological and ecological cultivation research (Yang et al. 2010; Liu et al. 2020). Currently, maize ear traits are mainly measured manually, which is time-consuming, tedious, and vulnerable to human error (Li et al. 2014). Thus, it is necessary to develop a high-throughput phenotyping machine to extract ear traits automatically.

In recent years, the morphological characteristics of agricultural products have been investigated and extracted by image analysis, which is more objective, automatic, and higher throughput than previous methods (Wu et al. 2020). Dubey et al. applied 45 morphological characteristics (such as shape, size, and color) to recognize wheat seed varieties (Dubey et al. 2006). Mebatsion et al. (2013) classified 5-grain varieties using morphological characteristics and color features. They extracted the Fourier coefficients, length-to-width ratios, maximum diameters, roundness, and RGB color features of individual grain images and built classification models that achieved a recognition accuracy for 5-grain varieties of greater than 98.5%. Igathinathane et al. (2009) developed an ImageJ plugin to detect a single disjoint particle shape and determine its size distribution. Panigrahi et al. (1998) developed computer vision-based techniques to classify different shapes of maize ears, and a maximum correspondence of 80% could be obtained for classifying cylindrical maize ears based on 18 samples. Wang et al. (2010) used computer vision to grade the appearance quality of fresh maize ears based on the HSI color model, and the results showed that compared to manual measurements, the average errors of the bare tip position, ear length, and ear maximum diameter were 2.27 mm, 1.96 mm, and 0.54 mm, respectively. Tanabata et al. (2012) developed SmartGrain software for high-throughput measurement of seed shapes.

Several studies have also contributed to automatically extracting image-based maize ear traits. Ma et al. (2012) measured the 3D breeding features of multiple maize ears based on the HSV color space of images, and the measurement accuracy of ear volume could reach 94%. Zhao et al. (2009) assessed the suitability of image processing techniques for measuring and quantifying maize ear traits under stationary conditions. The results showed that the relative measurement errors were 6.2% and 1.6% for ear length and ear diameter, respectively. A single maize ear was placed on a turntable, and side view images of each ear were collected and merged (Wang et al. 2013). However, this method took a long time (an average time of 102 s/ear) to measure limited ear traits of a single ear.

In our previous study, our group developed a high-throughput yield trait scorer for rice grain (Duan et al. 2011) and maize kernels (Liang et al. 2016). In this study, to phenotype more comprehensive maize ear and kernel traits, we developed a high-throughput and integrated maize ear traits scorer (METS) based on automatic control and image analysis techniques to measure 34 maize ear and kernel traits. These image-based traits were tested to classify three maize groups with different maize kernel shapes. In addition, the general applicability of the image analysis pipeline was also demonstrated on other independent maize ear phenotyping platforms.

Materials and methods

Materials

In total, 1894 maize inbred lines were used in this study; 813 of them were maize hybrid combinations from a complete diallel cross design, 296 lines were recombinant inbred lines derived from superhybrid maize, and 785 lines were selected from a maize association mapping population that has been widely used for numerous traits. First, maize ear and kernel traits (including the ear length, ear diameter, kernel thickness, ear number per row, ear row number, and kernel type) were measured manually. Then, the maize ears were put in the METS for automatic measurements. The automatic measurements of the maize ear traits are shown in Table 1.

Table 1.

Automatically extracted results of 34 traits per ear by METS

Traits		Abbreviation
Color features	The average of 0.5*red minus 0.5*blue	Ave_Gray
	The average of red plane	Ave_R
	The average of green plane	Ave_G
	The average of blue plane	Ave_B
	The average of hue plane	Ave_H
	The average of saturation plane	Ave_S
	The average of intensity plane	Ave_I
	The average of value plane	Ave_V
	The average of lightness plane	Ave_L
Texture features	The mean value	M
	The standard error	SE
	The smoothness	S
	The third moment	MU3
	The uniformity	U
	The entropy	E
Ear morphological traits	Ear length	EL
	Ear length without ear barren tip	ELWEBT
	Ear diameter	ED
	Ear barren tip length	EBTL
	Ear type	ET
	Total kernel number	TKN
	Ear projected area	EPA
	Ear projected perimeter	EPP
	Ear projected perimeter-area ratio	EPAR
	Ear length-width ratio	ELWR
	Kernel number per row	KNPR
	Ear row number	ERN
Kernel morphological traits	Kernel thickness	KT
	Kernel width	KW
	Kernel thickness-width ratio	KTWR
	Kernel color	KC
	Kernel area	KR
	Kernel perimeter	KP
	Kernel perimeter-area ratio	KPAR

System description

The system schematic diagram and prototype of the METS are shown in Fig. 1. The system includes four main parts: a maize ear transmission part, a maize ear detection and control part, an image acquisition part, and an image processing part. (1) The maize ear transmission part was designed to deliver maize ears through the imaging area automatically, which consisted of a frame, a conveyor belt, and T-baffles of equal space (Fig. 1a). Equally spaced T-baffles were designed to prevent the maize ears from rolling when they were delivered via the conveyor. A maize ear is placed between two T-baffles, and up to three ears can be placed on the belt at one time for acquired imaging (Fig. 1b). A black conveyor belt was used to enhance the contrast between the maize ears and background. (2) The maize ear detection and control parts were designed to detect maize ear positions and control image acquisition and conveyor belts. The maize ear detection and control part consists of a stepper motor, a programmable logic controller (PLC), inspection chamber, and an eddy current sensor (Fig. 1c). When an ear passes through the image area and is detected by the eddy current sensor, the eddy current sensor sends a signal to the PLC to stop the conveyor belt and simultaneously triggers the CCD camera (UI-1490LE-C-HQ, IDS Imaging Development Systems GmbH, Germany) to acquire images of the ear. After acquiring ear images, the PLC starts the conveyor belt driven by the stepper motor and delivers the ears to the collection unit. (3) The image acquisition part mainly consists of a CCD camera and LED (IDS-FSH-400 × 400 W, IDS Imaging Development Systems GmbH, Germany). The LED is fixed on the top of the inspection chamber and provides uniform illumination for imaging. The configuration of image acquisition is shown in Fig. 1d. (4) The image processing part was designed to process ear images and extract ear traits. In addition, the ear weight is obtained via an electronic balance. The PLC was an S7-200 (Siemens, Germany), and image acquisition and communication with the PLC were developed in LabVIEW 11 (National Instruments, USA).

Fig. 1.

The maize ear traits scorer. a The inspection chamber; b the maize ear samples; c the appearance of the maize ear traits scorer; d the configuration of inspection unit

Operation procedure and system control

The system operation procedure (shown in Fig. 2a) includes the following steps: (1) start the system and scan or input the barcode; (2) acquire the total ear weight via the electronic balance; (3) put maize ears on the conveyor belt with T-baffles; (4) detect the maize ears with the eddy current sensors and stop the conveyor belt; (5) acquire and deliver each image with three ears to the computer for image processing; (6) store the results, including the original color image and the maize ear traits, in the user-predefined folder; and (7) start the conveyor belt again to process the next maize ears. The METS operation is shown in Supplemental file 1, Movie 1.

Fig. 2.

The system operation procedure and the flowchart of the image processing procedure for measuring maize ear traits. a The system operation procedure; b the flow chart of the image processing and trait extraction procedure; c the maize ear traits

ESM 1.

Download video file ^{(9.7MB, mp4)}

(MP4 9919 kb)

Image processing

The flowchart of image processing for measuring maize ear traits is outlined in Fig. 2b. The multiple-ear (N ears) image (Fig. 3a) was captured by the CCD camera and sent to a computer queue. The image processing step was synchronous with the image acquisition step, which includes the following steps: (1) fetch one image from the queue and segment the binary image of whole ears from the background, such as with the color threshold operator for image segmentation and several morphological operators (open, filling holes, etc.), and remove small particles (Fig. 3b); (2) extract a single ear binary image of the ith ear (Fig. 3c) with the region of interest (ROI); (3) extract the original color image of the ith ear and combine the original color image with the mask of the binary image of the ith ear (Fig. 3d); (4) subtract the blue plane extracted under the RGB color space from the saturation plane extracted under the HSL color space, which turns the ith ear original color image into a grayscale image (Fig. 3e); (5) segment the grayscale image into a binary image with a fixed threshold (Fig. 3f); (6) obtain an image of a single ear without a barren tip (Fig. 3g); (7) convert the color image without a barren tip into a grayscale image using 0.5 times the red plane minus 0.5 times the blue plane, and then segment the maize kernels from background (Fig. 3h) using Otsu’s thresholding method; (8) identify and remove small impurities; (9) extract the touching kernels whose kernel area was more than 1.3 times the median area, isolate the touching kernels with watershed segmentation (Fig. 4b), and merge them with the untouching kernels (Fig. 4c).

Fig. 3.

The key image processing procedures. a The original image; b the binary image of the original image after small particles have been removed; c single binary image of an ear extracted using the region of interest; d single RGB image of an ear extracted by a single binary image multiplied by the original image (a); e single gray image obtained by subtracting the blue image from the saturation image; f single binary image without an ear barren tip; g single original image without an ear barren tip; h single binary image with every kernel segmented

Fig. 4.

The treatment of touching kernels. a Image with touching kernels and non-touching kernels; b image of non-touching kernels (top) and image with touching kernels (bottom); c image merging non-touching kernels with touching kernels after the touching kernels were processed with the watershed segmentation and oversegmentation methods

Trait extraction

Extraction of the ear length, ear diameter, and ear barren tip length

With the ear binary image, the EL was defined as the maximum distance between the two bases of the ear (Fig. 5a). The ED was defined as the average width of the ear’s middle part (approximately three kernel lengths) (Fig. 5b). The ear barren tip length (EBTL) was calculated as the ear (Fig. 5a) minus the ear without the ear barren tip (Fig. 5b).

Fig. 5.

Sketch map for extracting the ear length, ear diameter, and ear barren tip length. a Sketch map for the ear length (EL); b sketch map for the ear diameter (ED) and ear barren tip length (EBTL)

Extraction of the kernel number per row and kernel thickness

Every kernel position was calculated, and three kernels, whose positions were closest to the center of the ear, were defined as the center kernels (the light blue kernels shown in Fig. 6b). Based on the center kernels, the kernels in the same row were traced with the criterion that the closest kernel within 45 degrees was recognized as the next (left or right) traced kernel.

Fig. 6.

Sketch map for extracting the kernel number per row (KNPR) and kernel thickness (KT). a The whole kernel binary image after segmentation; b the three-row kernels facing the camera and the 20 kernels of three rows for calculating the kernel thickness (KT); c enlarged image for calculating the missing kernels

After one row of kernels was traced, the number in one row was counted and defined as N₁. Then, the distances between one kernel and the kernel next to it (defined as RR_i, which is shown in Fig. 6d) were calculated, and the median of all the RR_i values (defined as RR_MEAN) was obtained. The distances between the point to the left of one kernel and the point to the right of the kernel adjacent to the original kernel were also calculated (defined as LR_i, which is shown in Fig. 6d). Then, the number of kernels N₂ that were possibly missed was calculated using Eq. (1).

$$N_2=\sum \left(\text{round}\left(\frac{{\mathrm{LR}}_i}{{\mathrm{RR}}_{\text{mean}}}\right)\right)$$ 1

Then, the kernel number per row (KNPR) is num_i = N₁+N₂.

According to the above method, the KNPR of three rows in the middle part was calculated, and the final KNPR of this ear was defined as the maximum.

We extracted the total length L_i of 20 kernels in the middle part of three rows and then used Eq. (2) to calculate the kernel thickness (KT, Fig. 6b).

$$\mathit{KT}=\frac{\left(\frac{L_1}{20}+\frac{L_2}{20}+\frac{L_3}{20}\right)}{3}=\frac{\left(L_1+L_2+L_3\right)}{60}$$ 2

where L_i (i = 1, 2, 3) is the whole length of 20 kernels in the ith row.

Extraction of the ear type and ear row number

The ear edge angle, which was the angle between line 1 and line 2 (shown in Fig. 7a), was used to depict the ear type (ET). With the ear binary image, we calculated the widths of positions A and B, which were 1/4 of the distance between the two bases of the ear. Then, we used Eq. (3) to calculate the ear edge angle. If the ear edge angle was less than 4 degrees, the ET was cylindrical; otherwise, it was conical (Zhao et al. 2009).

$$\alpha =\text{atan}\left(\frac{\mathrm{abs}\left(\frac{H_2-H_1}{2}\right)}{L}\right)=\text{atan}\left(\frac{\mathrm{abs}\left(H_2-H_1\right)}{2L}\right)$$ 3

where α is the ear edge angle, H₁ is the width of position A, H₂ is the width of position B, and L is the distance between A and B.

Fig. 7.

Sketch map for determining the ear type (ET) and calculating the ear row number (ERN). a Sketch map for determining the ET; b sketch map for calculating the ERN

To calculate the ERN based on one ear side, the kernel width, ear diameter, and distance between two adjacent kernels in the middle part of each ear were obtained. Figure 8 b shows a sketch map for calculating the ear row number. Suppose the middle maize ear section was round, and there were n (n could be 1, 2, or 3) kernels between points M and N. Then, the angles θ₁, θ₂, and θ₃ could be calculated by Eq. (4), and the ERN could be calculated by Eq. (5).

$${\theta }_1=\mathrm{ar}cos\left(\frac{\mathit{AO}}{r}\right),{\theta }_3=\mathrm{ar}cos\left(\frac{\mathit{BO}}{r}\right),{\theta }_2=\pi -{\theta }_1-{\theta }_3$$ 4

Fig. 8.

Sketch map for determining the kernel type. a The original image; b the preprocessed image for determining the kernel type

Then,

$$\mathrm{E}\mathit{RN}=2\times n\times \mathit{\text{round}}\left(\frac{\pi }{{\theta }_2}\right)$$ 5

where r is the radius of the cross-section of the ear, n, AO, and BO.

Extraction of the kernel type

The kernel type includes dent, semi-dent, and flint. First, the original image was preprocessed, and the ear barren tip was removed. Second, the color features and texture features of the ROI, which was defined as the center part of the ear (shown as the red rectangle in Fig. 8b, H was 1/2 of the ear diameter and W was 1/2 of the ear length after removing the ear barren tip). The color features included the averages of 0.5 times the red plane minus 0.5 times of the blue plane (gray), red plane, green plane, blue plane, hue plane, saturation plane, intensity plane, value plane, and lightness plane of the ROI. The texture features were the six histogram features, namely, the mean value (M), the standard error (SE), the smoothness (S), the third moment (MU3), the uniformity (U), and the entropy (E), which were calculated using the following equations.

$$\mathrm{M}=\sum _{i=0}^{L-1}{\mathrm{G}}_{\mathrm{i}}p\left(G_i\right)$$ 6

$$\mathit{SE}=\sqrt{\sum _{i=0}^{L-1}{\left({\mathrm{G}}_{\mathrm{i}}-M\right)}^{2}p\left(G_i\right)}$$ 7

$$S=1-\frac{1}{1+{\mathit{SE}}^2}$$ 8

$$\mathit{MU}3=\sum _{i=0}^{L-1}{\left({\mathrm{G}}_{\mathrm{i}}-M\right)}^{3}p\left(G_i\right)$$ 9

$$U=\sum _{i=0}^{L-1}{p}^2\left(G_i\right)$$ 10

$$E=\sum _{i=0}^{L-1}p\left(G_i\right){log}_{2}p\left(G_i\right)$$ 11

where G_i is the ith gray level and p(G_i) is the probability of G_i. L is the maximum gray level.

Finally, the color features and the texture features of the ROI combined with the morphological features of the ear and kernel (Table 1) were used to classify the three types of kernels (dent, semi-dent, and flint) using a support vector machine (SVM, Byun and Lee 2002) and random forest (RF, Hu et al. 2020) models.

Software interface

The software interface is illustrated in Fig. 9. Three maize ear images can be obtained and processed simultaneously. The original color image, the segmented binary image, the maize ear and kernel traits, and the indexed barcode are displayed on the interface. The barcode can be scanned automatically or manually input. The 34 extracted traits are listed in Table 1. The software operation procedure of the Maize Ear Traits Scorer (V 1.0) is shown in Supplemental file 2, Movie 2. The main instructions of the image analysis pipeline are shown in Supplemental file 3, Note s1.

Fig. 9.

The software interface. The lower left part of the interface is the path and file controls. The top left of the interface is the original image display control. The middle part of the interface is the processed image display controls, and the right side of the interface is the maize ear traits and kernel traits display controls

ESM 2.

Download video file ^{(8.9MB, mp4)}

(MP4 9136 kb)

Results and discussion

Results

Measuring the accuracy of the ear length, ear diameter, kernel thickness, kernel number per row, and ear row number

To evaluate the measurement accuracy of the METS, 813 maize ears were automatically measured using the system and manually measured. The mean absolute percentage error (MAPE, defined by Eq. (12)) for the EL, ED, and KT were 1.99%, 2.13%, and 4.07%, respectively. The root mean square errors (RMSEs, defined by Eq. (13)) for the EL, ED, and KT were 0.33 cm, 0.11 cm, and 0.26 mm, respectively. The square of the correlation coefficient (R²) values for the EL, ED, and KT were 0.96, 0.79, and 0.85, respectively. The scatter plots of the automatic measurements versus the manual measurements of the EL, ED, and KT are shown in Fig. 10a, 10c, and 10e, respectively. The frequency number plots of the EL, ED, and KT are shown in Fig. 10b, 10d, and 10f, respectively. The absolute percentage errors of 97.66% of the ears for the EL, 92.74% of the ears for the ED, and 73.06% of the ears for the KT were less than 5%. The frequency number plots of the KNPR and ERN are shown in Fig. 10g and 10h, respectively. The mean absolute percentage errors (MAPEs) of 0, ± 1, ± 2, ± 3, and ± 4 for the KNPR were 32.68%, 46.73%, 17.76%, 2.71%, and 0.12%, respectively. The mean absolute percentage errors (MAPEs) of the 0, ± 1, ± 2, ± 3, and ± 4 rows for the ENR were 68.63%, 12.27%, 12.27%, 2.09%, and 0.74%, respectively. The mean absolute error (MAE, defined by Eq. (14)) was 0.83 rows.

$$\mathit{\text{MAPE}}=\frac{1}{n}\sum _{i=1}^n\frac{\left|x_{\mathit{ai}}‐x_{\mathit{mi}}\right|}{x_{\mathit{mi}}}\times 100\%$$ 12

$$\text{RMSE}=\sqrt{\frac{\sum _{i=1}^n{\left(x_{\mathit{ai}}-x_{\mathit{mi}}\right)}^2}{n}}$$ 13

$$\mathrm{MAE}=\frac{1}{n}{\sum }_{i=1}^n\left|N_{\mathit{ai}}-N_{\mathit{mi}}\right|$$ 14

where x_mi is the manually measured value, x_ai is the automatically measured value, N_mi is the manually measured value of the ERN, N_ai is the automatically measured value of the ERN, and n is the number of maize ear samples.

Fig. 10.

Frequency number plots and scatter plots of the automatic measurements versus the manual measurements of some of the ear traits. a Scatter plots of the automatic measurements versus the manual measurements of the ear length; b frequency number plots for the ear length; c scatter plots of the automatic measurements versus the manual measurements of the ear diameter; d frequency number plots for the ear diameter; e scatter plots of the automatic measurements versus the manual measurements of the kernel thickness; f frequency number plots for the kernel thickness; g frequency number plots for the kernel number per row (KNPR); h frequency number plots for the ear row number (ERN)

System measuring accuracy of kernel type

In total, 1081 ears, which included 209 ears with dent kernels, 347 ears with semi-dent kernels, and 525 ears with flint kernels, were tested to evaluate the classification accuracy of kernel type. Thirty-two variables, namely, 9 color features, 6 texture features, and 17 morphological features, were normalized to 0–1. Then, the eigenvalues of 1081 × 32 covariance matrix were obtained using principal component analysis (PCA, Rahman et al. 2020). The eigenvalues and their contribution rates are shown in Table 2, and the cumulative contribution rate of the principal components is shown in Fig. 11. From Table 2 and Fig. 11, the cumulative contribution rate of the first 15 principal components was 99.04%. Therefore, the first 15 principal components were selected for further classification analysis.

Table 2.

The eigenvalues and their contribution rates of the principal components

PCs	PC1	PC2	PC3	PC4	PC5	PC6	PC7	PC8	PC9	PC10	PC11	PC12	PC13	PC14	PC15	PC16
Eigenvalue	14.93	5.33	1.98	1.82	1.63	1.50	1.07	0.91	0.75	0.61	0.39	0.27	0.21	0.15	0.13	0.08
Contribution rate (%)	46.67	16.65	6.20	5.68	5.08	4.69	3.35	2.83	2.35	1.92	1.23	0.84	0.66	0.48	0.41	0.24

Fig. 11.

The cumulative contribution rate of the principal components for selecting the number of variables

In this study, 4/5 of the total samples were selected as the training set and 1/5 were selected as the test set based on the Kennard-Stone algorithm (Emerson 2015; Huang et al. 2019), and two different models (i.e., support vector machine (SVM) and random forest (RF) models) were established to classify the kernel type. The particle swarm optimization (PSO, Venter and Sobieszczanski-Sobieski 2003) algorithm was used to select the intercept (c) and the degree of the polynomial (g) of the polynomial kernel for the SVM. The optimal values of c and g for this study were 67.3483 and 0.01, respectively. A comparison of the classification results using the two models is shown in Table 3. In this study, the SVM model was better than the RF model with an 86.11% classification accuracy for the test set.

Table 3.

Comparison results of the different model for kernel type classification (%)

Model	Dent type		Flint type		Semi-dent type		Average		Average 10-fold cross-validation accuracy
	Train accuracy	Test accuracy	Train accuracy	Test accuracy	Train accuracy	Test accuracy	Train accuracy	Test accuracy
SVM	85.03	78.57	90.24	86.67	87.77	89.86	88.44	86.11	78.17
RF	100	71.43	100	80.00	100	92.75	100	82.41	76.87

Heat mapping of all the extracted maize and kernel traits

In this study, a correlation analysis between the 32 color features, texture features, and morphological features of maize ears, and maize kernels was carried out. The correlations between all the studied feature pairs are presented in Fig. 12. A darker blue on the color scale represents a positive correlation, and a darker red represents a negative correlation.

Fig. 12.

Heat map analysis of the color features, texture features, and morphological features

In this study, under a P value of 0.01, the total kernel number (TKN) was positively correlated with the ear length without the ear barren tip (ELWEBT) (r = 0.76), the ear length (EL) (r = 0.66), the kernel number per row (KNPR) (r = 0.85), the ear area (EA) (r = 0.70), and the ear row number (ERN) (r = 0.54), and was negatively correlated with the kernel thickness (KT) (r = − 0.45), implying that a longer ear with more kernels per row and smaller kernel thickness could yield more kernels. The ear length (EL) was positively correlated with the ELWEBT (r = 0.92), the KNPR (r = 0.78), the EA (r = 0.89), the EP (r = 0.79), and the TKN (r = 0.66), implying that longer ears could yield more kernels per row, larger area, larger perimeter, and more kernels total. The ear length without the ear barren tip (ELWEBT) had a stronger correlation with the KNPR (r = 0.88) and the TKN (0.75) than the ear length, implying that the ELWEBT would strongly affect these traits. The kernel thickness (KT) was negatively correlated with the KNPR (r = − 0.54) and the TKN (r = − 0.45), implying that smaller kernel thickness could result in more kernels per row and more kernels total.

Discussion

A high-throughput maize ear traits scorer (METS) was developed, by which 3 ears can be measured at one time. The time cost of image acquisition was approximately 2 s, and the time cost of image processing for 3 ears was approximately 8–10 s. In total, 34 ear features, namely, 9 color features, 6 texture features, and 19 morphological features of one ear, were extracted. We used 32 of these features to classify the kernel type. In the future, we will establish a kernel type classification model based on deep learning, which needs a robust training set. With a portable power source, the METS could also be used to obtain maize ear traits in the field.

We also compared the measurement accuracy of the METS with that of a commercial maize ear scorer (YTS-MET, Wuhan GreenPheno Science and Technology Co., Ltd., China). We randomly selected 112 maize ear images acquired with YTS-MET to calculate the maize ear length and width. Then, the 112 images of maize ears were also processed using Maize Ear Traits Scorer V 1.0. The correlation coefficients of the ear length and ear diameter between the YTS-MET and METS were 0.86 and 0.94, respectively (Fig. 13). We demonstrated that the image analysis pipeline (Maize Ear Traits Scorer V 1.0) could be extended to other independent maize ear phenotyping platforms.

Fig. 13.

The comparison of the measurement results from this study versus those from a commercial maize ear scorer. a Scatter plots of the measurements with the commercial maize ear scorer versus the measurements of the ear length from the developed maize ear scorer; b scatter plots of the measurements with the commercial maize ear scorer versus the measurements of the ear diameter from the developed maize ear scorer

Conclusion

In this study, we developed a high-throughput maize ear traits scorer (METS) to obtain 9 color features, 6 texture features, and 19 morphological features of maize ears. We evaluated the measurement accuracy and demonstrated the adaptability of the system to different varieties of maize ears. Compared with traditional phenotyping methods of maize ear traits, METS provides an automated, high-throughput, and low-cost (~ 5000 dollars) method for measuring maize ear traits.

Supplemental information

ESM 3 ^{(20.6MB, partial)}

(PARTIAL 21068 kb)

ESM 4 ^{(743KB, xls)}

(XLS 743 kb)

ESM 5 ^{(18.7KB, docx)}

(DOCX 16 kb)

ESM 6 ^{(460.3KB, rar)}

(RAR 460 kb)

Author contributions

Liang X. processed the images, analyzed the data, and wrote the paper. Ye J. designed the hardware of the system and performed parts of the experiments. Li X. conceived the experiment. Tang Z. analyzed parts of the data. Zhang X., Li W., and Yan J. provided the materials. Yang W. supervised this project and revised the paper.

Funding

This work was supported by the grant from the National Key Research and Development Program (2020YFD1000904-1-3), the National Natural Science Foundation of China (31770397), and the Fundamental Research Funds for the Central Universities (2662020GXPY010, 2662020ZKPY017).

Data availability

The source code and data for this study can be accessed via http://plantphenomics.hzau.edu.cn/download_checkiflogin_en.action.

Declarations

Consent for publication

Consent for publication was obtained from all authors.

Competing interest

The authors declare no competing interest.