Plant phenomics has become a crucial part of agricultural science and life science research [4,8,9]. Using high-throughput collection platform facilities for plant phenotypes, a system can obtain high time sequence plant phenotype data, and a plant phenome extensive data management system can be constructed to explore the relationship between the plant genotype phenotype and environment [9]. In recent years, variety recognition, digital breeding, and intelligent cultivation management have become the priority of plant resource analysis. Since the late 20th century, the acquisition and analysis of high-throughput, high-quality in situ plant phenotypic data have been important research directions in the field of modern agriculture and, thus, have been widely investigated. However, most of the current measurement methods of plant phenotypes still require performing plant sampling and obtaining the phenotypic parameters of plants through multi-angle vision or 3D reconstruction with depth information in a laboratory environment. Still, these methods cannot dynamically and genuinely reflect the growth of plants in a natural environment [21]. Therefore, the acquisition and analysis of high-throughput plant phenotype data based on field in situ have been highly significant. In recent research, numerous IoT technologies have been used to collect plant phenotype data to promote throughput and automation performances.

With the continuous cooperative research and development in multidisciplinary fields, including sensor technology [22], communication technology [23], plant science [24], computer science [25], and engineering science, agricultural IoT technology has made rapid progress by applying environmental sensor networks, nondestructive imaging, spectral analysis, robot technology, machine vision, and lidar technology. Agricultural IoT uses different sensor equipment and sensing technologies, as well as wireless sensor networks, mobile communication wireless networks, and the Internet, to collect data on agricultural production, product circulation, and plant body. The collected agricultural data are preprocessed, integrated, and analyzed [26]. With the rapid development of IoT technology and its introduction into agriculture, modern imaging technology, and automated industrial equipment, high-throughput techniques have been widely employed to obtain plants’ phenotypic characteristics in various growing environments using integrated imaging sensors. Numerous studies have been conducted to address the problem of low phenotypic throughout and spatial–temporal limitation of crop populations in the field, and different solutions have been proposed [16,27–29]. For instance, the PhenoWatch (https://www.zeal-quest.com/) field phenotypic analysis system is a mobile phenotypic imaging system, which was designed on the basis of the sensor-to-plant concept. In addition, 2 rack systems are used in the PhenoWatch system for measurement purposes: the field gantry version and the removable field version. The system uses a laser radar system to obtain 3D high-precision point cloud data on plants, and a variety of phenotypic parameters and information on plants are obtained through special software and algorithms. The 3D central imaging unit of the PhenoWatch for plant phenotype includes multiple measurement sensors, such as lidar, RGB, and multispectral cameras. In addition, the FieldScan is Netherlands’ 3D high-throughput plant laser scanning and measurement system. This system represents a high-throughput field plant phenotyping platform specially designed for complex field environments, which can perform the measurement under the all-weather condition in any environment. Furthermore, this system can perform full-automatic, high-throughput scale-type measurements on plants planted in natural farmland within an area of 20 m × 100 m. Moreover, it can acquire 3D multispectral point cloud data on plants and can automatically calculate phenotypic parameters, such as plant height, 3D leaf area, projected leaf area, digital biomass, leaf angle, leaf area index, light penetration depth, and leaf coverage. In addition to the phenotypic parameters, the multispectral reflection at each point of a plant can also be measured. Information, such as greenness, chlorophyll, and aging, can be obtained by analyzing a single band, whereas the normalized difference vegetation index, enhanced vegetation index, and photochemical reflectance index can be obtained by analyzing several bands. The bands measured using a multispectral camera can be customized from a list of wavelengths to meet the requirements of different application areas. The FieldScan system can be used for the measurement of a study area and individual plants. The matching environmental sensor collects ecological data of the measurement area and combines the plant growth data and environmental parameters for conducting a comprehensive analysis. The University of Nebraska–Lincoln proposed a suspended cable phenotypic platform, which covers an area of 0.4 ha and has a payload capacity of a sensor box of 30 kg. This platform integrates a 4-band multispectral camera, a thermal infrared camera, a 3D scanning lidar, and a visible near-infrared spectrometer to realize the plant parameter measurement. The sensor platform can move at a maximum speed of 2 m/s and move between any 2 positions in the imaging area within 45 s, achieving a positioning accuracy of 5 cm in the XYZ coordinate system. The core imaging area is divided into 128 zones, each of which has a size of 4.6 × 6.1 m and contains 6 crop rows with a row spacing of 0.76 m. The entire room is equipped with a drip irrigation system and weather stations. The remaining significant components supporting the facility operation include the subsurface drip irrigation system, electrical building, and 4 winch houses for winch operation. The platform runs every other day, collecting an average of 10 GB of data.

Among many phenotypic parameters, in situ measurement data on plant height have always been a hot research topic in crop science. The height of maize plants is an important agronomic trait related to the lodging resistance of a maize stalk and an essential index of maize phenotypic parameters. The phenotypic measurement of plant height can provide necessary data for quantitative analysis of the genotype-environment interaction effect and contribute to maize breeding [29,30]. Most of the traditional measurement methods of crop height are based on manual measurement. The vertical distance between the highest point of a plant and the soil surface is measured by placing the ruler near the plant or the infrared distance meter at the highest point of the natural growth of maize. This measurement method is substantially affected by the subjective factors of experimenters and is prone to errors. To address this problem, recent studies used a laser scanner and performed image-based fixed-point plant height extraction and RGB-D camera plant height measurement [31–33].

The LQ-FieldPheno facility, which is presented in Fig. 1A, is located at the National Agricultural Information Engineering Technology Research Centre, Beijing, China (39°56′36.0″N, 116°17′11.2″E, 60 m above sea level). The portable phenotypic high-throughput automatic acquisition system based on the field track uses a crop phenotypic multisensor central imaging unit as a core and integrates different systems, including an automatic control system, an ascending field track, an intelligent navigation vehicle, and an environmental sensor. It can acquire phenotypic data of a field plant population canopy in the whole growth period in a high-throughput, full-automatic, and high time sequence manner. Moreover, this system has the characteristics of multiple acquired data, vital timeliness, robust expandability, high cost, and flexible customization. The overall efficiency of the system can reach 15 s per plot. Furthermore, the system can operate 24 h a day and continuously detect up to 5,760 zones; it processes only RGB images and can effectively measure by lidar up to 5,760 plots/day for different types of plants, including maize, wheat, rice, soybeans, vegetables, flowers, and fruit trees.

The field track platform is used for data acquisition; a depth camera and 3D lidar are employed as leading collection equipment for the maize population point cloud data. In addition, the phenotypic platform includes an industrial camera, thermal infrared camera, multispectral camera, and integrated control equipment, as shown in Fig. 1C. The running track of the phenotypic platform is 14 m long and 4 m wide and has a running speed of 5 cm/s. The platform height is 1.8 m above the maize canopy, as shown in Fig. 1B. As shown in Fig. 1D, in the control system, the height of the collection equipment can be adjusted continuously according to the current growth state of maize plants. The collection angle is perpendicular to the maize growth direction.

A powerful microcomputer (Intel NUC 11 Essential) would be installed in the data acquisition platform. The microcomputer would be responsible for data acquisition sensor calls, platform operation path planning, and data upload from other environmental sensors. In addition, a routed base station with a bandwidth of 1 GB has been set up in a nearby laboratory. The platform’s microcomputer is always in network coverage, which both reduces the latency of the control system and ensures that data are uploaded in real time. All data processing algorithms are developed in advance and stored in the cloud. Once the data have been uploaded, it can be parsed in real time via the data processing interface.

Fourteen maize varieties, including m751, zd958, and xy335, were selected as test objects. The selected seeds were treated and sowed on 2020 June 21. One row of each variety was planted with a planting density of 60,000 seeds/ha. The data collection was conducted in the jointing stage, and the daily collection periods included 8:00 to 8:30, 13:00 to 13:30, and 18:00 to 18:30. Furthermore, the test period was from 2020 August 5 to 2020 September 30, during which the shallow temperature was 19 °C, while the extremely high temperature was 35 °C.

Various plant sensors were mounted on the LQ-FieldPheno platform to capture multimodal plant traits, and they included:

An NUC minicomputer with a Windows system controlled all sensor modules, collected sensors data, and stored the measurement data into the personal computer platform. Detailed information on the sensors is given in Table 1.

The set route overall efficiency could reach 45 plants/min. According to the measurement system size, it could measure hundreds, thousands or even tens of thousands of plants, or hundreds of communities daily. Suitable plants for measurement included maize, wheat, rice, soybean, vegetables, flowers, fruit trees, and other plant groups. The images acquired by the sensors are shown in Fig. 2.

The control diagrams of the phenotyping platform and the remote control system of the phenotyping platform are shown in Fig. 3A and B, respectively. The operating system included a host computer and a local server. The operational interface of the host computer could be logged in a short-distance range via Wi-Fi (within 50 m without shielding) or remotely via LAN. It was used for pre-operation preparation and data acquisition control, which included the processes of experiment registration, platform status monitoring information, control of each sensor and fill light, track path planning, real-time visualization of the data acquisition by a bridge process, data storage, and data transmission. The local server was connected to the host remotely, which could perform real-time job control, status monitoring, data return, data management, and statistical analysis of the job process.

The PLC platform controlled the XYZ 3-axis motor to move in the entire study area. The platform PLC control box was set next to the track-type platform. The PLC host was connected to the network with the RJ45 interface and could be used by the NUC industrial control host equipped with professional software to communicate with the host computer. The values of real-time XY coordinates and speed were displayed in the LCD number box, and the path was displayed on the right side. The green point denotes the current coordinate point and the grey line represents the running route; the green line indicates the running course. The software could modify the parameters of each of the sensors, such as exposure time and data storage format, according to the specific environmental conditions.

The system main interface is shown in Fig. 3C, where it can be seen that it included the main overview information bar, sensor status bar, deep camera operation bar, multispectral operation bar, and lidar operation bar. The multispectral operation bar included information on the sensor connection status, SD card status, and settings, such as photographing mode, which was used to preview and set the multispectral sensor. The overview bar showed the GPS status, remaining space in the directory tree, and an optional storage folder to store the collected data. The deep camera operation bar included the sensor acquisition speed and the preview of the RGB, depth, and infrared images.

In this study, lidar data were used to estimate the phenotyping of maize (includes plant height, leaf length, leaf width, etc.). The specific steps of the data acquisition and processing process were as follows:

The plant height data extracted from the point cloud data were compared with the manual measurement results. In this study, the manual measurement results were obtained by the experimenters in the field. Namely, the daily plant height measurement data of all plants in the study area were averaged, and the average value was used as a representative value of the plant height for that day. The correlation coefficient (R²) and the root mean square error (RMSE) were selected to assess the results, and they were calculated as follows, respectively:where m denotes the number of comparison objects, v_l is the manual measurement result, v_l^{\prime } is the phenotypic parameter extracted from the automatic segmentation result using the algorithm, and {\overline{v}}_l is the mean value of the manual measurement result.

Before collecting the data, the best field angle, and height of the lidar, multispectral camera and RGB camera were evaluated, and the results are shown in Fig. 10. For lidar, from top to bottom and from left to right, the optimal height values were as follows: 100, 110, 120, 130, 140, 160, 170, 180, 190, 200, 210, and 230 cm. On the basis of the texture data of the point cloud, the sensor needed to be far away from the canopy to obtain a large angle of view. The lower the distance from the canopy was, the clearer the texture details of a plant were, as shown in Fig. 10A. Multispectral images were acquired at different heights, and 5 sensor-matching parameters were calculated for the multispectral camera. First, the translation parameters of the 5 sensor images were calculated at 5 height values (30, 55, 105, 155, and 205 cm), and the x and y translation parameters were fitted to obtain the translation parameters at any given height. Then, an image pixel resolution at different image heights was calculated, and the pixel size of the imaged object in the image at any height was fitted, as shown in Fig. 10B. For the RGB camera, 4 indexes of distortion coefficient, definition, time efficiency, and data redundancy were considered the compatibility of the field of view of each sensor, and the current requirement of the extracted parameter information were analyzed; the optimal shooting height of the image was obtained to be 1.7 to 1,9 m, as shown in Fig. 10C.

As shown in Fig. 10D, the central imaging unit had the most significant utility ratio at a distance of 1.8 m from the crop canopy and could obtain accurate RGB, multispectral, and lidar data in terms of plant details and acquisition efficiency. The size of the positive mosaic RGB and multispectral images was 10.7 m \times 4.81 m, and the lidar data area was calculated considering the optimal field of view angle and operating speed to be 6.6 m \times 2.96 m. The overlapping area of the 3 sensors was regarded as a region of interest, and the region of interest was constantly determined to calculate and analyze the plant phenotypic indicators.

The upper left corner of the field crop phenotyping high-throughput platform was set as the origin of the coordinate system. The sequential orthophoto images acquired by the field phenotyping platform and the coordinates of each photo were processed in batches. Automatic stitching of the global orthophotos was performed by a homonymous image point matching algorithm. An international orthophoto image was obtained for the area covered by the device, as shown in Fig. 11. This image was used for the automated analysis of field phenotypes conducted by the image processing algorithms and for a quick check of the completeness of the collected data.

Next, each planting plot was segmented and labeled according to the initial planting map of the field and the centroid of each maize plant; the level set algorithm was used to extract the maize centroid. Then, the collated plots were used to obtain the attribute information of each material. The phenotypic parameters obtained using a high-throughput analysis algorithm were used to compare differences in parameters between varieties, for instance, critical phenotypic traits, such as plant height, cover, leaf area index, and biomass.

For data with a relatively large number of lidar point clouds, but relatively uniform point cloud density, the voxel grid point method can be employed for a down-sampling operation to reduce the number of point clouds, which is essential for achieving the real-time performance of point cloud filtering. The data before and after the point cloud down-sampling are shown in Fig. 12A and B, respectively, where it can be seen that, after the point cloud down-sampling, the morphological characteristics of maize were unchanged, but the point cloud density was substantially reduced.

When the 3D voxel grid was a cube with a side length of 1 cm, the number of points before down-sampling was 14,685,679 and the number of points after filtering was 521,477, so the compression rate was 36%. The more lidar scanning lines were in the equipment selection, the more point clouds were. Therefore, the point clouds generated by the lidar and depth cameras were selected one by one by adjusting the side length of the voxel grid according to actual requirements. According to the volume size of the directional bounding box and the number of point clouds in a unit volume, the values of 0.2, 0.5, and 1 cm were selected as grid side lengths to perform point cloud down-sampling. To evaluate the down-sampling effect, a single maize plant data were processed for different grids, and the algorithm processing effect is shown in Fig. 13.

The experimental results show that, when the grid edge length was less than 0.1 cm, the result was close to the original point cloud data, but the sampling efficiency was low. However, when the grid edge length was longer than 1 cm, the outline shape of the point cloud data before and after the sampling was unchanged, saving many computing resources in subsequent operations. The point cloud data reading time results of the lidar before and after the point cloud down-sampling are shown in Table 2.

The maize plant height is an essential indicator of maize phenotypic parameters, and it is closely related to the lodging resistance of maize stalks. After correction by the point cloud coordinate system, the ground slope was reduced. The ground 3D point cloud fluctuated less in the z axis direction than x and y axes. At the same time, the plant height of a single maize plant was calculated as a distance from the highest point of the natural growth of maize to the ground.

In the error analysis of the plant height data obtained by the 3D lidar, the plant heights corresponding to 3 density values of AD268 maize varieties of 2,500, 4,000, and 5,500 plants per area on 2020 August 14 and 27 were selected. The comparison results of the systematically and manually measured values of the plant height on the 2 d are given in Table 3. As shown in Table 3, on 2020 August 14, the average relative errors of the 3 densities were less than 2%, and on 2020 August 27, the maximum and minimum relative errors were 3.45% and 0.71%, respectively. The RMSE and R² results on 2020 August 14 and 27 were 0.99 cm and 0.9137, as well as 1.80 cm and 0.8471, respectively. In addition, the achieved linear fitting effect was good. The regression analysis results of the maize plant height are shown in Fig. 14.

The validation results for the other phenotypic parameter are shown in Fig. 15. In the comparative results for leaf length, leaf width, leaf inclination, and leaf growth height, R² and RMSE were 0.87 and 4.63 cm (leaf length), 0.85 and 0.93 cm (leaf width), 0.95° and 4.08° (leaf inclination), and 0.95 and 7.12 cm (leaf growth height), respectively. The correlation between the extracted results for leaf width was low, and the correlation for leaf length was medium, while the correlation for leaf inclination and leaf growth height was high.

This paper presents an integrated phenotypic measurement platform based on a near-surface remote sensing technology, which can effectively address the existing phenotypic data collection problems. The proposed platform provides a powerful research method for plant biology, phenomics, precision cultivation management, and breeding. However, the throughput of genotype and phenotype data can create significant data in the breeding science. In addition, for high-throughput data, there are problems with the storage, management, and analysis of massive data. Therefore, the phenotype measurement platform should be further improved and innovated, and its application fields should be expanded.

Future application and development of the proposed integrated platform should focus on the following 3 aspects: