A Plug Seedling Growth-Point Detection Method Based on Differential Evolution Extra-Green Algorithm

Abstract

To produce plug seedlings with uniform growth and which are suitable for high-speed transplanting operations, it is essential to sow seeds precisely at the center of each plug-tray hole. For accurately determining the position of the seed covered by the substrate within individual plug-tray holes, a novel method for detecting the growth points of plug seedlings has been proposed. It employs an adaptive grayscale processing algorithm based on the differential evolution extra-green algorithm to extract the contour features of seedlings during the early stages of cotyledon emergence. The pixel overlay curve peak points within the binary image of the plug-tray’s background are utilized to delineate the boundaries of the plug-tray holes. Each plug-tray hole containing a single seedling is identified by analyzing the area and perimeter of the seedling’s contour connectivity domains. The midpoint of the shortest line between these domains is designated as the growth point of the individual seedling. For laboratory-grown plug seedlings of tomato, pepper, and Chinese kale, the highest detection accuracy was achieved on the third-, fourth-, and second-days’ post-cotyledon emergence, respectively. The identification rate of missing seedlings and single seedlings exceeded 97.57% and 99.25%, respectively, with a growth-point detection error of less than 0.98 mm. For tomato and broccoli plug seedlings cultivated in a nursery greenhouse three days after cotyledon emergence, the detection accuracy for missing seedlings and single seedlings was greater than 95.78%, with a growth-point detection error of less than 2.06 mm. These results validated the high detection accuracy and broad applicability of the proposed method for various seedling types at the appropriate growth stages.

1. Introduction

According to the China Statistical Yearbook 2023 [1], the vegetable cultivation area in China covers approximately 23 million hectares, and over 70% of the vegetables are grown using the plug-tray seedling and transplanting method. The position of the growth point in plug seedlings closely correlates with the seed’s position within the corresponding plug-tray hole. Achieving even seeding at the center of each plug-tray hole leads to uniform light exposure and growth of seedlings, which is crucial for ensuring the high speed and quality of seedling transplanting operations. To explore the influence mechanism of seeding position, it is essential to accurately detect the position of the seed sown in the plug-tray hole first. However, directly measuring the position of a seed covered by substrate soil within a plug-tray hole is challenging [2]. Consequently, research on the detection method for plug seedling growth points is valuable for assessing the sown seed’s position within the plug-tray holes.

The conventional approach to detecting plug seedlings typically relies on the identification of features such as shape, color, and texture within two-dimensional images. Yang et al. [3] determined the quality of white palm plug seedlings by calculating the green peak value of their image pixels and using the stem projection area as a criterion. Tong et al. [4] converted images of Salvia militaris plug seedlings to grayscale using the 2G-R-B factor, employed the watershed algorithm to segment the overlapping leaves, and extracted the seedlings from the plug tray to assess their quality based on the area and quantity of leaves. In another study, Tong et al. [5] applied the 3G-R-B grayscale factor to convert seedbed images to grayscale, developed a seedbed segmentation image stitching algorithm utilizing block area, Harris corner detection, and SURF feature point detection, and evaluated the health of individual seedlings by calculating their leaf area. The method of two-dimensional image feature extraction is straightforward, efficient, and versatile, making it suitable for various types of plug seedlings.

In contrast to the relatively basic feature information derived from two-dimensional images, more intricate spatial image features can be extracted from both color and depth images. González-Barbosa et al. [6] developed a 3D scanning prototype utilizing the Kinect 2.0 sensor to monitor the growth of Portos tomato seedlings and achieved a classification accuracy of 99%. Syed, T.N. et al. [7] conducted cluster analysis and three-dimensional model segmentation on point clouds from pepper, tomato, cucumber, and lettuce plug seedlings, extracted characteristic parameters such as leaf area, stem diameter, and plant height, and achieved growth status detection for these seedlings. Buxbaum et al. [8] captured RGB and depth images of lettuce plants and developed an end-to-end deep learning model capable of non-destructively predicting plant biomass with a mean error rate of 7.3%. Otoya et al. [9] studied a leaf area estimation method based on point cloud segmentation and triangulation algorithms to categorize artichoke seedlings into four distinct quality grades. Three-dimensional image processing methods are highly effective in capturing the holistic characteristics of plug seedlings. However, they are data-intensive and are primarily employed for the detection of individual seedlings.

Machine learning techniques have been effectively employed to construct predictive models for seedling quality grading and to differentiate seedlings from the substrate background within plug trays. Samsuzzaman et al. [10] enhanced the boundary contour determination in seedling image segmentation by integrating image feature extraction with a support vector machine (SVM) for overlapping images of seedlings, including peppers, tomatoes, cucumbers, and watermelons. Jin et al. [11] employed the fuzzy C-means (FCM) clustering algorithm for high-throughput detection of 18-day-old cabbage seedling images, achieving individual seedling segmentation with a 97.33% accuracy rate for detecting empty plug-tray holes. Li et al. [12] utilized a multi-layer perceptron (MLP) neural network classifier to distinguish features between seedling leaves and the background substrate, matched the center coordinates of seedling-connected domains with corresponding plug-tray holes, and ascertained the health status by comparing the area of seedling-connected domains against a threshold. Wu et al. [13] employed a template-matching method to extract the height and stem diameter of tomato plug seedlings from a frontal view, calculated the leaf area through a color-based image segmentation approach, and established a predictive model for seedling quality grading by constructing a support vector machine (SVM) classifier. Machine learning methods are typically used for small-scale datasets, which are sensitive to different data types and well-suited for the detection of a single variety of plug seedlings.

Deep learning methodologies have been increasingly adopted for the autonomous grading of plug seedlings. Narisetti et al. [14] used deep learning techniques to automatically detect and segment greenhouse-grown seedlings with an average detection accuracy of over 90%. Li et al. [15] employed mask-RCNN for the segmentation, classification, and counting of watermelon seedling leaves, and enhanced the measurement of leaf area by repairing missing leaf information with CycleGAN. Perugachi-Diaz et al. [16] employed the AlexNet convolutional neural network to predict the growth of cabbage seedlings and achieved a recognition accuracy of 94% in classifying cabbage seedlings. Jin et al. [17] enhanced the U-net model with Resnet to achieve precise segmentation of pepper seedlings and accurate positioning of seedling leaves. Jin et al. [18] developed a grid coordinate-based overlapping image (GCOI) algorithm and trained a prediction model based on faster R-CNN to distinguish between missing, healthy, and unhealthy seedlings. Kolhar and Jagtap [19] introduced a CNN-ConvLSTM-based model for classifying the quality of Arabidopsis thaliana seedlings with a remarkable classification accuracy of 97.97%. Li et al. [20] replaced the backbone part of YOLOv5s with that of EfficientNetv2, modified the PANet structure with Bi-FPN, and introduced the exponential moving average attention mechanism, achieving an average grading accuracy of 95.6% for tomato plug seedlings. Convolutional neural network methods can achieve high seedling detection accuracy without the need for thresholding processing. However, these methods are primarily used for the detection of specific varieties of plug seedlings, as they require the establishment of large training datasets and high-performance computing hardware configurations.

Previous studies on seedling detection have predominantly focused on a single variety of seedlings, utilizing these methods for quality assessment or grading during seedling supplementation or transplanting operations. The aim of this study is to explore an accurate method for detecting the growth points of plug seedlings as a means of indirectly evaluating the seeding position within each plug-tray hole. The specific objectives of this research are twofold: (1) to propose a detection method that encompasses two-dimensional image grayscale processing, boundary division of plug-tray holes, single seedling identification, and growth-point positioning for the early stage of plug seedlings; and (2) to validate the accuracy and adaptability of the proposed method by comparing its detection results with those obtained through manual methods.

2. Materials and Methods

2.1. Materials

The Hikvision MV-CA013-20GC camera (Hangzhou, China), with a 1280 × 1024 pixel resolution and 6 mm focal length, was used to capture clear images of plug seedlings in the laboratory, as shown in Figure 1. As the LED light could provide a stable and uniform illumination and effectively eliminate the impact of shadows and reflections on image acquisition, two 14-Watt LED auxiliary lights were positioned on either side of the camera. The plug seedlings were arranged directly beneath the camera at 440 mm. Hikvision’s MVS (version 3.4.3) software was utilized to record the images in real-time.

The plug seedlings were cultivated during April to June 2024 in Guangzhou, China, with the temperature of 18–25 °C, the relative humidity of 60–80%, and the light intensity of 32,000–33,000 lx. Five trays each of Chinese kale plug seedlings (72 square holes), pepper plug seedlings (72 round holes), and tomato plug seedlings (72 round holes) with relatively uniform growth were randomly selected for observation. Images of the plug seedlings were captured daily within the time frames of 9:00–11:30 AM and 3:00–5:30 PM, and five clear images of each type of plug seedling were selected. A total of 75 images of the plug seedlings were uniformly cropped to a size of 1280 × 640 pixels, as presented in Figure 2.

On the morning of 19 September 2024, five trays of tomato plug seedlings (each with 72 square holes) and broccoli plug seedlings (each with 105 square holes), cultivated in the nursery greenhouse at Langhui Seedling Company in Foshan City, Guangdong Province, China, were photographed and are shown in Figure 3. The Phantom 20 smartphone camera with a resolution of 4080 × 3060 and an equivalent focal length of 24 mm was positioned vertically above the seedlings, maintaining an object distance of 800 mm. These plug seedlings had emerged cotyledons for three days. From the captured images, five clear images of each seedling type were selected and uniformly cropped to dimensions of 3700 × 1850 pixels, as illustrated in Figure 4.

2.2. Detection Protocols

The detection flowchart of plug seedling growth points is illustrated in Figure 5. Initially, the differential evolution algorithm was employed to adaptively optimize the grayscale factors, for enabling efficient removal of the nursery substrate background and extraction of seedling image data. Following this, morphological operations including median filtering, erosion, and dilation were applied to the grayscale image to highlight seedling characteristics. Meanwhile, a masking operation based on a threshold method was conducted on the original image to reveal the background features of the plug tray. A template of a standard-sized plug tray was utilized to divide the boundaries of the plug-tray holes by identifying the peak position on the pixel value overlay curve of the binarized plug-tray image. Finally, connectivity domain analysis was executed to identify a plug-tray hole containing a single seedling and to precisely determine the seedling’s growth point.

For the detection calculations, we utilized an Intel(R) core (TM) i5-9400 CPU with a base clock speed of 2.90 GHz, paired with a GT730 graphics card, 16 GB of RAM, and a 200 GB hard disk. The detection algorithms were developed and executed using Python 3.10 and OpenCV 4.7.0, running on a Windows 11 64-bit professional operating system.

2.3. Plug Seedling Growth-Point Detection Algorithm

2.3.1. Adaptive Optimization of Gray Factor

The grayscale processing of plug seedling images typically employs the 2G-R-B Ultra-green grayscale conversion method [21,22], which has been shown to significantly enhance the green channel and effectively separate green plants from the background image [23,24,25]. To further improve the differentiation between green seedlings and the plug-tray background, particularly for seedlings with varying cultivars and growth conditions, an adaptive optimization of grayscale factors was applied, as illustrated in Equation (1).

$$f\left(x,y\right)=\left\{\begin{array}{c}0,{\omega }_{go}G-{\omega }_{ro}R-{\omega }_{bo}B\le 0\\ {\omega }_{go}G-{\omega }_{ro}R-{\omega }_{bo}B,{\omega }_{go}G-{\omega }_{ro}R-{\omega }_{bo}B\ge 0\end{array}\right..$$

(1)

In Equation (1), $f(x,y)$ is the grayscale processing function for each pixel of the image, where G, R, and B (ranging from 0 to 255) represent the intensity values of the pixel’s red, green, and blue channels, $({\omega }_{go},{\omega }_{ro},{\omega }_{bo})$ is the optimized grayscale factor for the color channels, and ${\omega }_{go}>{\omega }_{ro}>0$, ${\omega }_{go}>{\omega }_{bo}>0$.

The differential evolution algorithm was used to obtain optimized grayscale processing factors for its swift convergence, independence from initial solution parameters, and robustness [26]. Grayscale factors are initialized as an individual vector ${\overrightarrow{x}}_{i,G}$ within the population, as illustrated in Equation (2).

$${\overrightarrow{x}}_{i,G}=({\omega }_{gi,G},{\omega }_{ri,G},{\omega }_{bi,G})$$

(2)

where i denotes the sequence number of an individual in the population, and its range is larger than 4; G represents the generation of evolution; and $({\omega }_{gi,G},{\omega }_{ri,G},{\omega }_{bi,G})$ is the grayscale factor of the ith individual vector, with the conditions of ${\omega }_{gi,G}>{\omega }_{ri,G}>0$, ${\omega }_{gi,G}>{\omega }_{bi,G}>0$.

For individual vectors ${\overrightarrow{x}}_{i,G}$, the mutation vector ${\overrightarrow{v}}_{i,G+1}$ is represented as

$${\overrightarrow{v}}_{i,G+1}={\overrightarrow{x}}_{k,G}+F\times ({\overrightarrow{x}}_{l,G}-{\overrightarrow{x}}_{m,G})$$

(3)

$${\overrightarrow{v}}_{i,G+1}=(v_{gi,G+1},v_{ri,G+1},v_{bi,G+1}).$$

(4)

In Equation (3), k, l, m are three different individuals randomly selected from the population; the mutation factor F ∈ [0,2] is a real constant used to control the scaling of the differential vectors. In Equation (4), $(v_{gi,G+1},v_{ri,G+1},v_{bi,G+1})$ is the corresponding grayscale factor of the mutation vector. The crossover vector ${\overrightarrow{u}}_{i,G+1}$ is obtained through crossover operation on mutation vector ${\overrightarrow{v}}_{i,G+1}$ and the target vector ${\overrightarrow{x}}_{i,G}$.

$${\overrightarrow{u}}_{i,G+1}=\left(u_{gi,G+1},u_{ri,G+1},u_{bi,G+1}\right)$$

(5)

$$\left\{\begin{array}{c}{u}_{gi,G+1}=\left\{\begin{array}{c}{v}_{gi,G+1},j=1 or randb\left(j\right)\le CR\\ w_{gi,G},j\ne 1 and randb\left(j\right)>CR\end{array}\right.\\ u_{ri,G+1}=\left\{\begin{array}{c}{v}_{ri,G+1},j=2or randb\left(j\right)\le CR\\ w_{ri,G},j\ne 2 and randb\left(j\right)>CR\end{array}\right.\\ u_{bi,G+1}=\left\{\begin{array}{c}{v}_{bi,G+1},j=3 or randb\left(j\right)\le CR\\ w_{bi,G},j\ne 3 and randb\left(j\right)>CR\end{array}\right.\end{array}\right.$$

(6)

In Equation (5), $\left(u_{gi,G+1},u_{ri,G+1},u_{bi,G+1}\right)$ is the grayscale factor of the crossover vector ${\overrightarrow{u}}_{i,G+1}$. In Equation (6), randb (j) is the random number within the range of [0,1] to ensure that there is always one factor of the crossover vector ${\overrightarrow{u}}_{i,G+1}$ generated from the mutation vector ${\overrightarrow{v}}_{i,G+1}$; CR is the crossover factor with a value range of [0,1].

Both color and overall texture features of the grayscale image were designed in the objective function of the differential evolution algorithm to better preserve the seedling foreground and eliminate the background of the nursery substrate. The method for calculating the color information in the grayscale image is

$$I\left(x,y\right)=\left\{\begin{array}{c}1,f(x,y)>0\\ 0,f(x,y)>0\end{array}\right.$$

(7)

$$G_P={\sum }_{x=1}^W{\sum }_{y=1}^{H}I(x,y)$$

(8)

In Equation (7), $I\left(x,y\right)$ is a binary function based on the grayscale processing function $f\left(x,y\right)$. In Equation (8), $G_p$ is the color feature value of the grayscale image, and W and H are the width and height of the image, respectively.

Texture information of the grayscale image is extracted with the local binary pattern (LBP) method [27]. The LBP operator is defined within a 3 × 3 pixel neighborhood, with the central pixel serving as the threshold. The gray values of the eight surrounding pixels are compared against the value of the central pixel. If a surrounding pixel’s value exceeds that of the central pixel, the position of that pixel is set as 1, otherwise it is regarded as 0. The resulting 8-bit binary number is sequenced and then converted to a decimal LBP value of the central pixel. The calculation method is

$$m=f_i-f_c$$

(9)

$$s\left(m\right)=\left\{\begin{array}{c}1,m\ge 0\\ 0,m<0\end{array}\right.$$

(10)

$$LB{P}_n(x_c,y_c)={\sum }_{i=0}^7{2}^{i}s(f_i-f_c).$$

(11)

In above Equations, $f_c$ is a local central pixel, $f_i$ is a set of 3 × 3 neighborhood pixels around $f_c$, s(m) is a sign function, and $LB{P}_n\left(x_c,y_c\right)$ represents the texture feature value at $f_c$. The overall texture feature value of the grayscale image ${LBP}_p$ can be represented as

$${LBP}_p=\sum {LBP}_n\left(x_c,y_c\right).$$

(12)

Normalization processing methods for grayscale image’s color and texture feature values are shown in Equations (13) and (14):

$${G_p}^{\prime }={\displaystyle \frac{G_p}{W\times H}}$$

(13)

$$LB{P_p}^{\prime }=\sum {\displaystyle \frac{LB{P}_n(x_c,y_c)-{LBP}_{min}(x_c,y_c)}{LB{P}_{max}(x_c,y_c)-{LBP}_{min}(x_c,y_c)}}.$$

(14)

A weighted function $F\left({\overrightarrow{x}}_{i,G}\right)$ is designed, as shown in Equation (15):

$$F\left({\overrightarrow{x}}_{i,G}\right)=\alpha LB{P_p}^{\prime }\left({\overrightarrow{x}}_{i,G}\right)-\beta {G_p}^{\prime }\left({\overrightarrow{x}}_{i,G}\right)$$

(15)

where α denotes the texture feature weight, β denotes the color feature weight, and $LB{P_p}^{\prime }\left({\overrightarrow{x}}_{i,G}\right)$ and ${G_p}^{\prime }\left({\overrightarrow{x}}_{i,G}\right)$ represent the texture and color feature value of the plug seedling image, respectively.

The new individual ${\overrightarrow{x}}_{i,G+1}$ is generated as

$$\mathsf{\Delta }=F\left({\overrightarrow{x}}_{i,G}\right)-F\left({\overrightarrow{u}}_{i,G+1}\right)$$

(16)

$${\overrightarrow{x}}_{i,G+1}=\left\{\begin{array}{l}{\overrightarrow{x}}_{i,G},\Delta <0\\ {\overrightarrow{u}}_{i,G+1},\Delta >0\end{array}\right.$$

(17)

$${\overrightarrow{x}}_{i,G+1}=({\omega }_{gi,G+1},{\omega }_{ri,G+1},{\omega }_{bi,G+1})$$

(18)

where Δ serves as the selection basis of the individual ${\overrightarrow{x}}_{i,G+1}$, and $({\omega }_{gi,G+1},{\omega }_{ri,G+1},{\omega }_{bi,G+1})$ is the grayscale factor vector of the individual ${\overrightarrow{x}}_{i,G+1}$. In addition, all the grayscale factors are positive real value, and ${\omega }_{gi,G+1}$ is greater than ${\omega }_{ri,G+1}$ and ${\omega }_{bi,G+1}$.

To avoid the iteration stopping at the initial convergence or falling into a local optimal solution, the termination conditions are

$$\mathsf{\Delta }F=F\left({\overrightarrow{x}}_{i,G}\right)-F\left({\overrightarrow{x}}_{i,G+1}\right)\le \epsilon$$

(19)

$$\sigma \le \epsilon +\lambda \left|\mu \right|$$

(20)

where ε is the convergence tolerance, σ is the standard deviation of the fitness function in the current population, μ is the mean of the fitness function, and λ is the relative tolerance.

Based on preliminary experiments, the following parameters were set: a texture feature weight of 0.8, a color feature weight of 0.2, an initial population size of 20, a mutation factor of 0.2, a crossover probability of 0.5, a convergence tolerance of 1 × 10⁻⁵, a relative tolerance of 1 × 10⁻³, and a maximum iteration number of 100. The boundary conditions for the grayscale factors are specified as ${\omega }_{gi,G}\in (1, 5)$, ${\omega }_{ri,G}\in \left(\mathrm{0.5,2.5}\right)$, and ${\omega }_{bi,G}\in \left(\mathrm{0.5,2.5}\right)$. An additional constraint of ${\omega }_{gi,G}\ge {\omega }_{ri,G}+{\omega }_{bi,G}$ is applied to ensure the complete extraction of green seedling features. By executing the differential evolution algorithm, the optimized grayscale factor vector $({\omega }_{go},{\omega }_{ro},{\omega }_{bo})$ can be derived, as shown in Equation (1).

2.3.2. Boundary Division of Plug-Tray Hole

The plug seedling images were transformed into grayscale using the differential evolution algorithm, which facilitated the extraction of the plug-tray background images by masking out the seedling features. The outcome of the OTSU binarization [28] process is depicted in Figure 6.

The binary image of the m-row and n-column plug-tray background was projected in both the horizontal and vertical directions to achieve pixel value superposition [29]. The resulting distribution curves from the pixel value superposition were then smoothed using the Savitsky–Golay filter. Subsequently, (m − 1) or (n − 1) peak positions were identified based on the most significant pixel value differences between adjacent peaks and valleys on these curves, as indicated by the red points in Figure 7.

A standard plug-tray template, consisting of m rows and n columns, is initially segmented by (m − 1) horizontal and (n − 1) vertical boundary lines that encircle the plug-tray holes. This template is then scaled and superimposed onto the plug seedling image, using the center of the plug tray as the reference point. Template matching is employed to ascertain the boundary lines of the plug seedling images. Based on the boundary width size between the two adjacent holes, if the peak positions in Figure 7 are found to be within a 15-pixel margin of the boundaries of the standard plug-tray template, these peak positions are used to replace the template’s boundaries. Figure 8 demonstrates the boundary division results for two seedling trays with 72 square holes and 72 round holes, respectively, thereby validating that the boundary division method is versatile and can accommodate various plug-tray designs and image distortions.

2.3.3. Seedling Identification and Growth-Point Detection

After dividing the boundary lines on the plug seedling image, each plug-tray hole is assigned a number based on its position and boundary information. The grayscale image within each plug-tray hole is then subjected to thresholding segmentation using Otsu’s method. Subsequently, a binary image is produced by removing noise through median filtering [30] and eradicating image voids via morphological processing. Finally, the contours of the connected domains of the seedling leaves within the plug-tray holes are extracted using a contour detection algorithm. The feature extraction process for plug seedlings is illustrated in Figure 9.

Each plug-tray hole may contain a single seedling, multiple seedlings, or be empty. Plug-tray holes in which no seedling-connected domains are detected can be identified as empty. For a seedling that extends across the boundaries of plug-tray holes, it is assigned to the hole that contains more than 50% of the seedling’s connected domain area. The connected domain information of the seedling is then superimposed onto this hole, while the information in the other hole is cleared.

The plug-tray holes containing seedlings are sorted by the area and perimeter of their connected domains, from smallest to largest. The 75th percentile is then used as a threshold. Plug-tray holes with a perimeter or area greater than n times the threshold is regarded as containing multiple seedlings. Based on preliminary experiments, n is determined to be 1.6. Identification results for tomato plug seedlings with cotyledons that emerged for three days are depicted in Figure 10, where white represents empty holes, red indicates holes with multiple seedlings, and green marks holes with a single seedling.

The growth point of a seedling is observed as the point closest to the region where the cotyledons connect. Therefore, it is defined as the midpoint of the shortest distance between the two cotyledon connection domains. We denote arbitrary points on the cotyledon connection domains A and B of the seedling as $P_a(x_a,y_a)$ and $P_b\left(x_b,y_b\right),$ respectively. The Euclidean distance [31] between these two points is calculated:

$$L_{ab}=\sqrt{(x_a-x_b{)}^2+(y_a-y_b{)}^2}.$$

(21)

Once the shortest Euclidean distance between two points, $P_{am}(x_{am},y_{am})$ on connection domain A and $P_{bm}(x_{bm},y_{bm})$ on domain B, is determined, the growth point $P_o(x_o,y_o)$ of the seedling can be defined as

$$(x_O,y_O)=({\displaystyle \frac{x_{am}+x_{bm}}{2}},{\displaystyle \frac{y_{am}+y_{bm}}{2}}).$$

(22)

Figure 11 illustrates the calculation results for the cotyledon connection domains of a pepper plug seedling with cotyledons that have emerged for three days. In the figure, the red points indicate the detected growth points.

2.4. Evaluation Method

For each plug seedling image, the effectiveness of the proposed detection method is evaluated by missing seedling identification rate $P_i$, single seedling identification rate $P_m$, and growth-point position error ${\mathsf{\Delta }}_d$. The three indices are defined by

$$P_i={\displaystyle \frac{n_i}{N_i}}\times 100\%$$

(23)

$$P_m={\displaystyle \frac{m_i}{M_i}}\times 100\%$$

(24)

$${\mathsf{\Delta }}_d=\sqrt{(x_g-x_o{)}^2+(y_g-y_o{)}^2}$$

(25)

where $n_i$ is the number of empty plug-tray holes correctly identified, $N_i$ is represented as the total number of empty plug-tray holes, $m_i$ is the number of plug-tray holes with a single seedling that are correctly identified, $M_i$ is indicated as the total number of plug-tray holes with a single seedling, $(x_g,y_g)$ is the growth-point position coordinate determined artificially, and $(x_o,y_o)$ is the growth-point position coordinate determined by the detection algorithm.

3. Results

3.1. Findings from the Laboratory Plug Seedling Tray Analysis

3.1.1. Results of Adaptive Grayscale Adjustments

The iterative grayscale processing curves along with the optimized grayscale factors for the laboratory plug seedling varieties with cotyledons that emerged over 1 to 5 days are depicted in Figure 12. The fitness functions for all varieties and growth stages achieve convergence within 40 iterations, and the optimized grayscale factors exhibit a certain degree of variation. This indicates that the differential evolution extra-green algorithm can flexibly leverage the green characteristics of the plug seedlings, swiftly determine the globally optimal solution, and maintain strong stability.

The differential evolution and 2G-R-B extra-green algorithms were independently applied to images of tomato, pepper, and Chinese kale plug seedlings with cotyledons that had emerged for three days. The grayscale processing outcomes are presented in Figure 13, Figure 14 and Figure 15. The results indicate that both algorithms are capable of extracting seedling features while simultaneously suppressing the background noise from the plug tray [32]. Notably, the differential evolution extra-green algorithm produced clearer contours of the seedlings, characterized by sharper edges, higher contrast, and better preservation of details, while effectively suppressing background noise. This enhanced clarity is particularly advantageous for the accurate identification and precise positioning of the growth point of the seedlings.

3.1.2. Results of Plug Seedling Identification and Growth-Point Detection

Figure 16 displays the identification results for seedlings with cotyledons that emerged from the first to the fifth day. The missing seedling identification rates for Chinese kale were 100% for the first to fourth days, dropping to 95.16% on the fifth day. For tomato and pepper plug seedlings, these rates were 100% from the first to third days and then decreased to 96.91% and 96.53% on the fourth and fifth days, respectively. The single seedling identification rates on the first day were 97.87% for tomato, 96.22% for pepper, and 99.31% for Chinese kale. As the growth duration increased, the single seedling identification rate reached 100% on the third day for tomato, the fifth day for pepper, and the second day for Chinese kale.

Variance analysis with software of SPSS 27.0.1 demonstrated that the seedling varieties significantly influenced the single seedling identification rates (p = 0.014), and the seedling stage had a significant impact on both the missing seedling identification rates (p = 0.002) and the single seedling identification rates (p = 0.025). This difference may arise from the vary growth rates of the three types of seedlings. Chinese kale seedlings exhibit the most rapid growth, followed by tomato seedlings, while pepper seedlings demonstrate the slowest growth. The optimal time for detecting tomato plug seedlings is on the third day, when both the missing and single seedling identification rates reach a perfect 100%. In the case of pepper plug seedlings, the highest rates of 97.57% for missing seedlings and 99.25% for single seedlings were achieved on the fourth day. Chinese kale seedlings achieve the best identification rates, with both missing and single seedling identification rates hitting 100% on the second day.

The detection results of the growth-point position in plug seedlings is presented in Figure 17. As the growth period extends, the positional errors of the growth points for all three types of seedlings tend to increase. There is a significant correlation between both the variety of seedlings and the duration of their growth period with the growth-point position error. The minimum growth-point position errors of 0.62 mm, 0.72 mm, and 0.69 mm were, respectively, recorded for Chinese kale, tomato, and pepper seedlings with cotyledons on the first day of their initial emergence. The maximum errors of seedlings with cotyledons on the fifth day of emergence were up to 2.50 mm for Chinese kale, 1.61 mm for tomato, and 1.18 mm for pepper.

As the growth period of the three types of plug seedlings extends, the missing seedling identification rate increases, the single seedling identification rate decreases, and the growth-point position errors expand. To attain high precision in both seedling identification and growth-point detection, the optimal detection time should be selected based on the growth characteristics of different varieties of plug seedlings. For laboratory-grown tomatoes, peppers, and Chinese kale, the recommended detection times are on the third, fourth, and second days following the emergence of seedling cotyledons, respectively. These timings yield over 97% missing seedling identification rate, 99% single seedling identification rate, and growth-point position errors of less than 1 mm.

3.2. Results of Plug Seedling Detection in the Greenhouse

Images of tomato and broccoli plug seedlings, cultivated in the nursery greenhouse, were processed with the differential evolution extra-green algorithm. The algorithm yielded optimized grayscale factors of (4.02, 1.54, and 2.47) for tomatoes and (3.89, 1.46, and 2.43) for broccoli, respectively. The resulting grayscale images, as depicted in Figure 18 and Figure 19, demonstrate that the seedling features can be distinctly extracted for plug seedlings in practical production settings. Figure 20 presents the boundary division outcomes for tomato and broccoli seedling trays, which have 72 and 105 square holes, respectively. These results confirm the adaptability of the boundary division method for plug trays of varying specifications.

For the tomato and broccoli plug seedlings with cotyledons that emerged after 3 days, the detection algorithm achieved a missing seedling identification rate of 95.78% and 96.16%, a single seedling identification rate of 98.29% and 100%, and growth-point position errors of 1.51 mm and 2.06 mm, respectively. This proves the adaptability of the detection algorithm to actual seedling production and its capability of achieving high detection accuracy.

The accuracy of seedling identification and growth-point detection for tomatoes grown in the greenhouse was significantly lower than that of those cultivated in the laboratory. This discrepancy is likely due to the more precisely controlled growth environment in the greenhouse, which promotes faster growth, larger leaf expansion, and greater stem height of the seedlings [33]. These factors can lead to increased distortion in the connected domains of the seedlings. Consequently, the optimal detection time for greenhouse-grown seedlings should be approximately one day earlier than for those grown in the laboratory.

4. Discussion

Misidentification of seedlings within a plug-tray hole may result from variations in the growth rates of seedlings within the same tray. During the early growth phase, some seedlings may germinate later and exhibit small cotyledons, which could be mistakenly classified as missing seedlings, as illustrated in Figure 21a. Another instance of misidentification might occur when a seedling’s leaf extends into an adjacent plug-tray hole, as depicted in Figure 21b. In such cases, the extending leaf could be erroneously identified as a seedling in the neighboring hole. Furthermore, as the growth period extends, the stems of larger seedlings are more susceptible to bending, and cotyledons from adjacent holes are more likely to intersect and overlap, as shown in Figure 21c. This can result in distortions within the connectivity domains of the seedling cotyledons, leading to incorrect determination of the growth-point position.

5. Conclusions

This study developed a growth-point detection method for plug seedlings, which serves as an indirect means of assessing sowing positions within plug trays. This research demonstrated that the differential evolution extra-green algorithm could adaptively optimize grayscale factors and extract clearer features for plug seedlings, outperforming the traditional 2G-R-B grayscale factor approach. Utilizing this method, the boundaries of plug trays can be accurately delineated, regardless of the type of plug tray. For tomato and pepper seedlings with cotyledons emerging on the third day, as well as Chinese kale plug seedlings with cotyledons emerging on the second day, laboratory tests achieved a missing seedling identification rate of over 97%, a single seedling identification rate of more than 99%, and a growth-point position error of less than 1 mm. Seedlings of tomato and broccoli cultivated in a nursery greenhouse also demonstrated high detection accuracy, with a recommended detection time that should precede those planted in the laboratory. This detection method has the potential to be integrated into the measurement of seed positions during the plug-tray sowing process. Additionally, it could facilitate seedling supplementation or transplanting operations by providing precise location information for seedlings within the plug-tray holes.

Future study should focus on accurately capturing the edge features of the small seedlings to further enhance the detection accuracy. High-pass filters could be employed to improve the extraction of edge features, thereby precisely differentiating these features from the background. Additionally, regional growth techniques may be utilized to identify the specific plug-tray holes associated with the expanding leaves. To address the challenge of overlapping and intersecting seedling leaves, local edge segmentation of the leaves from different seedlings should be conducted prior to applying the growth-point detection algorithm.