G-ROBOT: An Intelligent Greenhouse Seedling Height Inspection Robot

,


Introduction
Seedling cultivation is a key step in vegetable production and an important bridge between the source of seeds and vegetable products [1]. Although the vegetable seedling industry has developed rapidly in recent years, but most of the enterprises' seedling production management is mainly based on experience, yielding a low seedling growth rate and labor productivity; this is still far from agricultural modernization [2][3][4]. With the increase in labor costs, and the rapid development of information technologies such as the Internet of Tings and big data, the intelligent and industrialized seedling production has become an inevitable trend in the development of the modern vegetable industry. Continuous monitoring and control during seedling cultivation is an efective way to improve seedling quality and survival rate, and data collection and analysis enable us to efectively perceive crop growth conditions and then take corresponding prevention measures [5,6]. A seedling factory deploys a large-scale sensor network for intelligent monitoring purposes, which is considerably more efcient, accurate, and real-time than the manual method; however, such networks are highly complex in wiring and are costly. An intelligent mobile robot is a more fexible and cost-effective solution.
Currently, research is being conducted on inspection robots for crop growth monitoring. For example, Guo et al. designed a multi-degree-of-freedom robot system for greenhouse facility image acquisition and environment monitoring, which can accurately acquire images and data through the perception layer, and send the data for analysis and storage via a wireless network, thus enabling fne collection of data perception environment monitoring data for greenhouse production [7]. Liu et al. designed a monitoring device based on a mobile robot that enables remote real-time monitoring of the greenhouse environment [8]. Li et al. developed a mobile and suspended-rail crop growth and environmental information monitoring system, which enables integrated monitoring of crop growth and environmental information in horticulture facilities through multisensor information fusion [9]. Han et al. developed an indoor inspection robot that can conduct autonomous inspections based on pre-laid electromagnetic guide wires and transmit the collected temperature, humidity, CO 2 concentration, and other information to a cloud server for growers to view from WeChat [10]. Lu et al. designed and manufactured a small wheeled tobacco plant protection machine for the high-ridge-furrow environments, which has improved the performance of the transmission system and steering system through a clever mechanical structure to meet the stability requirements for feld operations [11]. Barker's team designed a multisource detection sensor system based on a vehicle-mounted platform to collect images of seedlings from multiple angles [12,13]. Te Sunti team's agricultural robot platform, built with corrosionresistant aluminum profles in the main body, using an Arduino development board as the control chip, and an integrated image processing algorithm in the host computer, can be used to identify and pick small fruits [14]. Bai et al. used a robotic phenotype collection platform to carry a highthroughput multisensor system that consists of fve sensor modules for measuring crop canopy traits from feld plots and geo-referencing sensor measurements using GPS as well as incorporates two environment monitoring sensors [15]. Atef et al. used a robot to detect leaf traits in greenhousegrown maize and sorghum. Te robot automated the measurement of plant leaf characteristics using a four-degree-of-freedom manipulator with a portable spectrometer and a thermistor for leaf temperature measurement [16]. In the progress both at home and abroad, the existing robots for crop growth monitoring are not well adaptable to greenhouse environments. Because they perform poorly in seedling growth diagnosis and real-time system control and data management, there is still large room for development and improvement.
Among the many phenotypic parameters that refect the growth status, the seedling height is an extremely critical factor. Te measurement of seedling height can provide an important basis for the quantitative analysis of the sound seedling index and also help in the cultivation of seedlings [17]. However, there are very few studies on intelligent seedling height inspection robots for greenhouse environments, and directly introduced inspection robots for other purposes cannot adapt to the special environments of seedling greenhouses and perform poorly in seedling height measurement. Terefore, we developed an intelligent greenhouse seedling height inspection robot for seedling growth monitoring. Te robot can walk around agilely and intelligently in the seedling greenhouse, collect seedling growth information and environmental information in a comprehensive and stable way, and analyze and store seedling height data in real time using cloud-based image processing algorithms and interactive software. Te robot is expected to become a powerful assistant for greenhouse crop research and cultivation personnel by improving their labor efciency and reducing their labor intensity, thereby promoting the intelligent process of greenhouse crop research and cultivation.

Performance Requirement Analysis.
Seedling production is mainly carried out in multispan greenhouses or glass greenhouses. A seedbed is generally 1.7 m wide, 18-20 m long, and 0.65 m high. A large number of seedbeds are evenly arranged and distributed in a huge space of several hundred to several thousand square meters, with an average spacing of approximately 0.6 m between seedbeds. Te seedbeds can be moved around slightly to facilitate workers' operations. Te robot walks around the seedling greenhouse autonomously, collecting environmental information and taking photos of seedlings and uploading them to the cloud server, wherein the images are processed and analyzed to obtain seedling growth data and stored for users to view in real time through the interactive software on the client computer. Based on task requirements, the robot should provide the following features: (1) the robot should have good trafcability when moving around and adapt to diferent types of greenhouse road surfaces; (2) when facing obstacles or potholes on the road, the robot should be able to overcome obstacles and remain stable when being disturbed in its movement; (3) the shooting height of the image acquisition device can be adjusted to adapt to diferent greenhouses and seedlings' growth periods; (4) the robot can move autonomously in a complex environment; (5) an accurate and efcient image processing algorithm should be developed for in-situ detection of seedling heights; and (6) greenhouse environmental data, seedling images, and seedling height data shall be transmitted, stored, and managed via a wireless network.

Robot Body Design
2.2.1. Overall Structural Design. Te robot has a modular structure that consists of a multi-terrain replacement chassis, an electronic control lift image acquisition bracket, and a quick release mechanism. Te multi-terrain replacement chassis consists of components such as the drive part, motion part, suspension part, and frame; the electronic control lift image acquisition bracket comprises components such as the motor, telescopic structure, and camera mount; the quick release mechanism comprises components such as the split connection part and reset pin. As shown in Figure 1, the mechanical structure of the robot is drawn and machined using SolidWorks software.
Te height of the seedbeds in the greenhouse is approximately 600-700 mm above the ground; the growth height of the seedlings is approximately 50-150 mm; the minimum imaging distance of the camera is 250 mm; and the spacing between the bottoms of seedbeds is approximately 600 mm. To meet the actual crop inspection and image acquisition requirements and facilitate the handling and movement of the robot, the length, width, and height of the designed robot were 570, 420, and 900-1300 mm, respectively, and the weight was 23 kg. Te main structural parameters of the robot are listed in Table 1.

Shock-Absorbing Wheel Replacement Chassis.
Te chassis is responsible for the movement of the robot and overcoming obstacles. Due to complex terrains in diferent greenhouses, to adapt to diferent road surfaces, the robot chassis is powered by four 3.1 Nm 57 stepper motors that are coupled with diferent tires through a designed 8 mm shaft hole coupler for wheel replacement. For a greenhouse on paved roads with large spacings between seedbeds, low-cost ordinary round rubber tires can be used. In the case of narrow and complex roads in a greenhouse, a McNamee wheel chassis can be used for agile multidirectional movement based on its high mobility. A track chassis is suitable for a greenhouse on nonpaved roads and can maintain a tight grip on uneven roads by adjusting the belt tensioner. Te design of the replacement chassis structure signifcantly improves the robot's adaptability and trafcability. Te suspension mechanism is responsible for maintaining the overall stability of the robot by ensuring that the violent shaking of the image acquisition camera caused by robot movement is minimized when the tire chassis is in contact with the ground [18]. Te chassis features a fourwheel independent suspension mechanism with the tires mounted on the crank guide-bar suspension mechanism composed of rocker links, bearings, and springs, and swinging around the suspension pivot point with rocker links when passing over potholes, thus enhancing the grip of the chassis on the ground and absorbing shocks. Te structure of the chassis suspension mechanism is shown in Figure 3.

Electronic Control Lift Image Acquisition Bracket.
A multisource camera is mounted on this bracket for image acquisition. A 24 V electric push rod with a stroke of 400 mm is used to constitute an electrically controlled image acquisition device. Te two poles of the DC motor of the push rod are, respectively, connected to two sets of relays and then connected to the positive and negative poles  Journal of Robotics 3 of the power supply. By controlling the conduction and disconnection of the two sets of relays, the switching of the positive and negative poles of the DC motor is completed, so that the speed of the electric push rod is 12 mm/s forward or reverse height adjustment. Te RealSense camera and Kinect camera are fxed onto the image acquisition device through diferent brackets to collect RGB-D information of seedling images. As shown in Figure 4, the electronic control lift image acquisition bracket allows the camera to acquire the images of seedlings in the height range of 900-1300 mm.

Quick Release Mechanism.
Te quick release mechanism is mainly responsible for the separation/combination of the robot chassis from/with diferent end actuators. It is mainly divided into upper and lower parts. Te upper part is a boss structure with pin holes on the side, and its upper surface is connected with the connecting seat of the end efector; the lower part cooperates with the upper connecting body and is connected to the chassis through the base. After the two are inserted, the locking is completed by the spring return pin. Owing to the aluminum alloy, it features a lightweight design with a high carrying capacity. As shown in Figure 5, the quick release mechanism achieves a split design for the robot, which enables the end actuator to be combined with or separated from the chassis within seconds, thus greatly facilitating the handling and storage of the robot.

Robot Motion Simulation.
Because the robot needs to be as stable as possible when it is moving around to reduce the jitter of the camera module to acquire higher quality images of the seedlings, we used Adams to perform kinetics  simulations and verify the reliability of the suspension mechanism and the entire system as a whole. We constructed a map with uneven surfaces to simulate actual road surface undulations, modeled the robot in equal proportion using SolidWorks, and then imported the model into Adams. Te kinetic parameters were set according to the actual parts mating and material characteristics. Te multiple-degree-of-freedom spring force was decomposed to the vertical direction as its equivalent combined force. Te driving force parameters were set according to the mechanical characteristics of the motor. For kinetic analysis of the chassis, the robot mass m was set to 23 kg; the global gravitational acceleration g was set to 9.81 kg/s 2 ; the friction coefcient μ was set to 0.3, which is the static friction coefcient between rubber and concrete road; and the direction perpendicular to the upward direction of the base plate was chosen as the main reference direction to refect the undulation degree of its center of gravity. Te simulation results are shown in Figure 6. Te vertical coordinate of the curves is the displacement of the center of mass of the bottom plate in the vertical direction, and the horizontal coordinate is the time used by the robot to move forward. Te red curve is the result without the chassis suspension mechanism, while the blue curve is the result with the chassis suspension mechanism. According to the simulation results, the chassis suspension mechanism has improved the overall stability of the robot, especially on undulating roads where it can efectively isolate the shocks of the vehicle from the ground. Figure 7, during the greenhouse production, it is usually necessary to inspect the seedbeds in specifed areas and design an intelligent navigation mode based on the operational requirements in the greenhouse. In this mode, the robot moves to the specifed point on the radar map along the planned route to complete the task.

STM32-Based Motion Control Module. As shown in
Te control system for the intelligent navigation mode features a distributed design for motion control through intermodular communications and collaborations. Te main control unit employs a STM32F407 series microcontroller with low power consumption, high stability, and rich interfaces, as well as 114 programmable I/O ports, 17 timers, and 17 communication interfaces, enough to meet the control and communication requirements. Te processor is from the Raspberry Pi 4B series and integrates 2 HDMI ports, 4 USB ports, and wired and wireless network interface cards that can transmit HD video streams as well as send and receive various data simultaneously. A LIDAR is used to sense the robot's surrounding environment. A Silan A1 LIDAR with a sampling frequency of 8000 times per second and a scanning frequency of 5.5 Hz is used to collect the road data within 12 m. Te inertial sensor (IMU) is responsible for sensing the attitude of the robot during its motion. A nine-axis inertial sensor is used as the IMU to measure the angular velocity, acceleration, attitude angle, etc., during robot motion.
Te intelligent navigation feature is based on the simultaneous localization and mapping (SLAM) technique. Te robot calculates its own position while building a map of the environment based on the information from the LiDAR and inertial sensors and odometer, and then it navigates to the specifed points based on the planned route. A ROS environment is built based on Raspberry Pi Ubuntu 20.04 system to provide the mapping and navigation features of the robot. Te GMapping algorithm [19] is used by the robot for mapping, synchronous localization, and map saving. A GMapping node is built to import the data of the LiDAR and motor odometer in real time, and after converting the radar coordinate system to the chassis coordinate system, a map is built using the IMU's attitude information. Te control system loads the built map, which provides the initial position, direction, and target point of the robot, planned a route using the A * algorithm [20], obtained the chassis motor speed control instruction, and sent the instruction to the STM32 controller via a serial port.
After receiving the motor control command, the STM32 controller generates PWM signals through the four channels of the advanced timer and then drives the chassis motor after power amplifcation by the drivers of the four stepper motors, thus enabling the robot to navigate around.

Raspberry
Pi-Based Image Acquisition Module. As shown in Figure 8, the image acquisition module system of the robot consists of a camera system, an information processing unit, and a data transceiver module. Te camera system consists of a RealSense and a Kinect camera mounted at the end of the abovementioned electronic control lift image acquisition bracket and a surveillance camera at the front of the chassis. Among them, Intel's RealSense D415 camera can acquire RGB-D information at 1280 × 720 Journal of Robotics 5 resolution and the Azure Kinect camera can acquire image color information at 4096 × 3072 resolution and image depth information at 1024 × 1024 resolution. Te Kinect camera can provide better-quality depth images, and RealSense is used to obtain better-quality color images with diferent imaging principles. Te purpose of using these two RGB-D cameras at the same time is for subsequent image fusion to obtain better RGB-D image quality to meet the needs of crop image acquisition. Te surveillance camera features a LeSports 3-in-1 camera, which is used to transmit the picture in front of the robot in real time and provide the remote monitoring function of the robot. With a limited space inside the robot, the information processing unit uses a Raspberry Pi 4B with dimensions (L × W × H) of 88 mm × 58 mm × 19.5 mm. Its 1.5 GHz 64bit quad-core processor and TCP/IP protocol-based wireless network communication support can meet the needs for image acquisition and transmission.
Te data transceiver module is a 5-mode and 13-frequency 4G network transmission module, which features ultralow latency, data encryption, and stable signals, and can meet the networking requirements of the robot. Te image acquisition system runs in the Raspberry Pi Ubuntu 20.04 environment. Te image acquisition code is compiled with Python 3.7 and linked to the camera video stream for image acquisition by calling OpenCV and NumPy interface library functions. During image acquisition, the RealSense camera or Kinect camera is connected to Raspberry Pi through a USB3.0 interface, and images of seedling crops are acquired and then uploaded to the cloud server for storage and processing through the 4G transmission module. Te surveillance camera is connected to the Raspberry Pi via an USB interface and a video stream is established with the cloud server via TCP/IP protocol to transmit the robot surveillance images in real time for retrieval and viewing.

Environmental Information Collection Module.
Te environmental information collection mainly focuses on light intensity, temperature, humidity, and carbon dioxide (CO 2 ) concentration in the greenhouse environment, which have a great impact on crop growth. Te environmental information collection system consists of a sensor, STM32 processor, and 4G transmission module, as shown in Figure 9. A light intensity sensor based on the BH1750 chip is used to measure the light intensity within the range of 0-65535 lx; a temperature and humidity sensor SHT30 is used to measure the temperature and humidity within the ranges of −40-125°C and 0%-100%, respectively, and with the accuracies of ±0.3°C and ±2%, respectively. A CCS811 sensor is used to measure the CO 2 concentration within the range of 400-5000 mg/m³.   Journal of Robotics When collecting environmental information, the sensor converts the received environmental information analog signals to digital signals through the AD conversion chip and sends the data to STM32 through TTL serial communication or I2C communication for them to upload the environmental data to the cloud server for storage through the 4G transmission module.

Power Management Module.
Te mobile robot is powered by a 24 V vehicle lithium battery. Due to diferent voltage requirements of various electronic control systems and sensors, the PW2902, PW2183, and PW2052 chips are connected in a series-parallel combination to build a power management module that supplies 24 V, 12 V, and 5 V highcurrent outputs, and provides rectifcation, overvoltage, and overcurrent protection and reverse polarity protection features. As shown in Figure 10, the step-down regulator feature is simulated electrically using the Simulink tool in MATLAB software to simplify the analog chip circuitry. Te simulated waveform results are shown in Figure 11. It can be seen that when the power input is 24 V 0.6 A, the voltage output by the simplifed circuit is stabilized at 12 V in a very short period of time, and its waveform is in line with expectations. Tis design meets the functional needs of each module for power supply. Figure 12, the robot can be controlled and managed through the PC-based host software, which greatly enhances the realtime performance and convenience of system. A mobile laptop serves as the host computer (equipped with a Core i7-3632 CPU, 2.20 GHz, 12 GB RAM, 64-bit OS). Te robot's host control software for Windows is developed in PyQt5 + Qt Designer environment using Python language. Te software is designed with an easy-to-understand and simple GUI, which makes it easy for the user to get started, and integrates various features such as robot control and operation status monitoring, image shooting control, crop image data viewing and management, etc.

Host Computer Control Software. As shown in
Te Raspberry Pi processor functions as an intermediary for wireless control of the robot motion by the host PC. When the host PC and the Raspberry Pi are at the same hotspot, the host PC establishes a connection with the Raspberry Pi by accessing its IP address. To perform a control operation, the host computer sends a "control instruction + check code" to the Raspberry Pi IP in the form of a character string. Upon reception of the character string, the Raspberry Pi decodes it into control instruction characters and then sends it to the STM32 via a USART serial port to control the motor operation, thereby controlling the robot operation. After the "Start" button in the host software is clicked, the robot moves along the route of destinations predefned in the ROS system. After the "Stop" button is clicked, the robot stops moving. In addition, buttons such as "forward/backward" and "turn around" in the host software can be used for manual and remote control of the robot.
Features of the host computer such as robot operation monitoring, seedling growth status diagnosis and analysis, crop information, and environmental data management are provided by the server-side program deployed on the cloud server. Te server-side program is developed using Node.js language, which uses a socket interface for communication. It features a Nginx architecture and integrates a MySQL database to build a cloud data storage and processing system that serves as a bridge between the robot and the host software.
Te surveillance camera at the front of the chassis uploads the captured images to the server and establishes a video stream, which is accessed by the host computer to monitor the robot operation. Te environmental information data collected by the robot sensors are uploaded to the server every hour. After the "Shoot" button in the host software is clicked, the collected seedling images would be uploaded to the server and saved; after clicking the "Data Analysis" button, the service program would call the image processing algorithm to detect and analyze the uploaded images and obtain the biomass data of the plant such as seedling height. After the "Save Results" button is clicked, the data in the server database can be exported to an Excel table for easy viewing and management.

Image Processing Algorithm for Seedling Height
Detection. Seedling height is a critical biomass parameter in the determination of the seedling growth quality. Seedling height is the distance from the base of the plant to the top of the main stem, i.e., the main stem growth point [21]. Te identifcation and localization of growth points is the key to seedling height measurement. In this project, we tested fve deep learning networks and selected the EfcientNet network [22] to identify seedling growth points. EfcientNet is a weighted bidirectional feature pyramid network (BiFPN) as shown in Figure 13. Te network allows simple and fast multiscale feature fusion; second, the network incorporates a compound scale dilation method that uniformly scales the resolution, depth, and width of all backbone, feature, and prediction networks. With this new idea, the detection accuracy of the EfcientNet network has been greatly improved compared with other networks.
We used the training method of migration learning and the weights of the ResNet and VGG network models trained with the Pascal VOC dataset [23] as the initialized weights of the network models. Labellmg software was used to label the  To solve the overftting issue caused by a small amount of data and to improve the training efect of the model and the accuracy of the results, the dataset was enriched by data augmentation. Te complexity of the samples was increased using traditional dataset augmentation methods, e.g., image rotation (45°, 60°, 90°rotation), brightness adjustment (0.8x and 1.3x), contrast enhancement (0.8x), addition of Gaussian noise (standard deviation 0.1), and mirroring (horizontal rotation), and the dataset was expanded to 8 times of its original size. Te preprocessed images were manually labeled using Labellmg software. Mark the growth points of multiple seedling images in a single image with rectangular boxes and name them growpoints, and save the results as .xml fles. After the annotation is completed, each image corresponds to a .xml fle with the same name. A total of 1600 near-growth point color images were produced. To ensure the dissimilarity of the data, 90% of it were used for the training set and 10% for the test set, which were put into a deep learning network model for training.
A schematic diagram of seedling height measurement is shown in Figure 14. Te center of the growth point prediction box identifed by the EfcientNet network serves as the pixel coordinates of the growth point. Te spatial coordinates of each pixel point in the depth image captured by the Kinect camera can be calculated using equation (1).
By mapping the pixel coordinates of the growth point to the depth image, the spatial coordinates of the growth point can be extracted [24], and then, the depth between the growth point and the camera plane (h2) can be calculated. If the measurement environment remains unchanged, the distance h1 from the camera plane to the top of the seedling pot can be measured manually, and the seedling height h3 can be calculated using equation (2).

Robot Operational Stability Test.
Te operational stability test of the robot prototype was conducted in the intelligent glass greenhouse of Huazhong Agricultural University during May 24-26, 2022, as shown in Figure 15.  Figure 13: Te framework of EfcientNet.
Journal of Robotics average relative humidity values were, respectively, 76%, 78%, and 85%; the weather was sunny and the light condition was good; the foor of the greenhouse was a cement pavement foor; the height of the seedbed was 750 mm; and the spacing was 800 mm. Te test started with the robot being fully charged. It moved around the seedbeds in the intelligent navigation mode to collect images and its operation was monitored from the host software. Te robot was operated from the fully charged voltage of 24.4 V till 22.8 V until it ran unstable.
Te test results showed that the robot featured an average battery life of 5.2 h in the greenhouse environment and high trafcability and stability during its operation. It was highly reliable and able to perform specifed operations for crop phenotype detection and environmental data collection. Te robot, server, and host computer communicated with one another stably and properly even under high temperature and humidity conditions, indicating that it could adapt to complex environments.

Environmental Data Validity
Testing. Te environmental data collected during the robot prototype test were collated, and the values from an environmental monitor from Changzhou Ekos Electronic Technology Co., Ltd. were used as comparison values to verify the accuracy of the environmental sensors, including temperature and humidity sensors, lightness sensors, and CO 2 sensors. Te resolution and range of the temperature measurement were 0.01°C and −40-60°C, respectively; the resolution and range of the relative humidity measurement were 0.01% and 0%-100% RH, respectively; the resolution and range of the illuminance sensor measurement were 10 lx and 0-100000 lx, respectively; the range of the CO 2 sensor was 0-5000 mg/m³. Te environmental data collected by the robot prototype were used as the measured values, and the above data were updated every 1 h. By comparing the measured values of the greenhouse environmental data with the comparison values ( Figure 16), we concluded that the following: the maximum diference between the measured value of temperature and the comparison value was 1.96°C, with a maximum relative error of 5.8%; the maximum diference between the measured value of humidity and comparison values was 1.79%, with a maximum relative error of 3.0%; the maximum diference between the measured value of light intensity and comparison values was 333 lx, with a maximum relative error of 2.6%; and the maximum diference between the measured value of CO 2 concentration and comparison values was 66 mg/m³, with a maximum relative error of 8.6%.

Seedling Height Detection Validity Test.
Phenotypic parameters of the seedling crop canopy were measured in the greenhouse for early Jia 8424 watermelon seedlings at the one-true leaf and one-apical bud stage, and Fengle Golden A pumpkin seedlings at the one-true leaf and one-apical bud stage as well as two-true leaf and one-apical bud stage. While moving around, the robot used the Kinect camera to acquire RGB color images and depth images; then it used the image processing algorithm to identify the growth points of seedlings at diferent growth stages and fnally measured seedling heights.
To verify the accuracy of the growth point detection algorithm, 49 watermelon seedlings at the young seedling stage, 45 watermelon seedlings at the one-true leaf and oneapical bud stage, 44 pumpkin seedlings at the one-true leaf and one-apical bud stage, and 47 pumpkin seedlings at the two-true leaf and one-apical bud stage were randomly selected for the test, as shown in Figure 17.
It is difcult to intuitively draw the pros and cons of each model algorithm by comparing the detection images of the fve models for watermelon seedlings [25]. Using 160 images of watermelon seedlings as the test set, the fve networks were quantitatively evaluated using AP (average precision) and F1 parameter (an evaluation index that comprehensively considers precision and accuracy). Te test results are shown in Table 2. It shows that the AP and F1 values of the Ef-cientNet network for the detection of watermelon seedling growth points are higher than those of the other four target detection models, and the detection time of the fve models is not much diferent. Terefore, the EfcientNet network is determined as the detection model for the growth point of fruit and vegetable seedlings in this paper.
To verify the accuracy of the seedling height calculation algorithm, the heights of the abovementioned tested seedlings were measured. Te height from the growing point of the seedling to the surface of the plug was measured using a vernier caliper with an accuracy of 0.1 mm, which served as the actual seedling height value. To make the test results more intuitive, the actual values of seedling heights measured manually were used as the x-axis and the seedling heights calculated using the proposed algorithm were used as the y-axis to draw scatter plots as shown in Figures 18(a)-18(d). Tese fgures are the scatter plots for watermelon seedlings at the young seedling stage, watermelon seedlings at the one-true leaf and one-apical bud stage, pumpkin seedlings at the one-true leaf and one-apical bud stage, and pumpkin seedlings at the two-true leaf and one-apical bud stage, respectively.
To compare the agreement between the seedling height calculated using the proposed algorithm and the manually measured seedling height values, a least squares regression analysis was conducted to linearly ft the scatter plots of the    two datasets, and the corresponding goodness of ft R 2 and root mean square error (RMSE) between the predicted and true values were calculated. Both R 2 and RMSE are indicators used to evaluate and describe the degree of agreement between the two datasets. A higher value of R 2 indicates a better ft between the predicted and true values.
Equations (3) and (4) are used to calculate R 2 and RMSE, respectively.  Figure 18: Scatter plot of actual seedling heights and measured seedling heights. (a) Watermelon seedling data at the young seedling stage. (b) Watermelon seedling data at the one-true leaf and one-apical bud stage. (c) Pumpkin seedling data at one-true leaf and one-apical bud stage. (d) Pumpkin seedling data at the two-true leaf and one-apical bud stage.
Te test results of seedling height calculations are shown in Table 3, which indicate that the R 2 values of the seedling heights measured using the proposed algorithm and the seedling heights measured manually were greater than 0.9 for the four seedling stages of fruit and vegetable seedlings, and the values of RMSE for these stages were 2.81, 3.69, 3.43, and 4.83, respectively. Te ftted equations obtained were near the direct proportion straight line with a slope of 1. Te results confrm that the seedling heights calculated by the algorithm of this paper are accurate. In Table 3, "W-YOUNG" represents watermelon seedlings in the seedling stage, "W-ONE-TRUE" represents watermelon seedlings in the one leaf and one heart stage, "P-ONE-TURE" represents pumpkin seedlings in the one leaf and one heart stage, and "P-TWO-TRUE" represents pumpkin seedlings of the two leaves and one heart stage.

Conclusion
We designed an intelligent and modular greenhouse seedling height inspection robot to acquire images of seedlings and environmental data during seedling cultivation in a greenhouse. Te test results confrmed the following. First, the robot is highly versatile and stable. Its multi-terrain replacement chassis can adapt to diferent types of road surfaces in the greenhouse and its independent suspension structure design enhances the stability of the robot in motion, so that the robot can complete inspection tasks in various common greenhouses. Te designed electronic control lift image acquisition bracket can capture highquality images of seedling crops, which meets the shooting requirements of greenhouse seedlings at diferent heights. Second, the robot can collect data in a stable way. Trough the deep learning algorithm based on the EfcientNet network to identify the growing point, the plant height data of the seedling can be accurately measured and the efcient measurement of the in-situ crop is realized. Moreover, its environmental data collection module can accurately obtain light intensity, temperature and humidity, and CO 2 concentration data in the greenhouse. At last, the robot system features a high integration level and high real time performance. Te host computer and cloud server can connect the system modules in real time, thus making it easier for the user to monitor and control the robot as well as analyze and manage data. In the follow-up, image fusion technology can be used to improve the quality of the collected images to further improve the accuracy of algorithm recognition; by further integrating each module and upgrading the humancomputer interaction software, the integration of the robot system can be higher, and the operation efciency can be improved while being convenient to use. Te robot plays a signifcant role in assisting greenhouse seedling cultivation research and promoting mechanization and intelligence of seedling cultivation.

Data Availability
Te data used to support the fndings of this study are available from the corresponding author upon request.

Conflicts of Interest
Te authors declare that they have no conficts of interest regarding the publication of this article.