A crop plant height extraction method and system based on an adaptive CSF algorithm

Abstract

This invention discloses a method and system for extracting crop plant height based on an adaptive CSF algorithm, belonging to the technical fields of precision agriculture and computer vision. The method includes: acquiring the original point cloud of the crop and filtering it to obtain a preprocessed point cloud; executing the CSF algorithm using ground point separation parameters to separate the preprocessed point cloud into ground points and non-ground points; constructing a digital elevation model based on the separated ground points, converting the elevation values ​​of the non-ground points into height values ​​relative to the ground to obtain a normalized point cloud; rasterizing the normalized point cloud to generate a canopy height model; and extracting crop plant height information from the canopy height model. This invention solves the problems of traditional methods requiring manual parameter adjustment, poor adaptability, low processing efficiency, and lack of agronomical knowledge, achieving fully automated, high-precision, and highly robust crop plant height measurement, which can be widely applied to precision agriculture, crop phenotypic analysis, and field management.

Description

Technical Field

[0001] This invention belongs to the technical fields of precision agriculture and computer vision, specifically relating to a method and system for extracting crop plant height based on the adaptive CSF algorithm. Background Technology

[0002] With the rapid development of precision agriculture technology, the accurate acquisition of crop phenotypic information has become crucial for crop breeding, field management, and yield prediction. Plant height, as an important crop phenotypic parameter, directly reflects the growth status and health of crops. Traditional methods of measuring plant height mainly rely on manual measurement, which is not only time-consuming and labor-intensive but also difficult to meet the needs of large-scale, high-frequency monitoring.

[0003] In recent years, point cloud-based 3D reconstruction technology has been widely used in the agricultural field. By acquiring farmland point cloud data through sensors such as LiDAR or depth cameras, non-contact, high-precision crop height measurement can be achieved. Among them, ground point separation is a key step in point cloud processing, which directly affects the final accuracy of plant height extraction. Existing ground point separation methods mainly include: (1) methods based on elevation thresholds, which simply identify points below a certain height threshold as ground points, but are prone to misjudgment in farmland with large terrain undulations; (2) methods based on slope filtering, which identify ground points based on the slope changes of adjacent points, but parameter setting is difficult and computational complexity is high; (3) methods based on morphological filtering, which use morphological opening operations and other methods, but have poor adaptability to ground features of different scales; (4) methods based on cloth simulation (CSF), such as existing technologies that identify ground points by simulating the cloth falling process, which has good results but the parameters are fixed and need to be manually adjusted for different scenarios.

[0004] The CSF (Cloth Simulation Filter) algorithm is an effective method for separating ground points. It identifies ground points by simulating the physical process of cloth falling under gravity and conforming to the terrain. However, the traditional CSF algorithm has the following problems: (1) Parameter settings depend on human experience and require repeated debugging for different scenarios; (2) Fixed parameters are difficult to adapt to point cloud data with large density variations in farmland environments; (3) It is inefficient when processing large-scale point clouds and cannot meet real-time requirements; (4) Improper handling of boundary areas can lead to classification errors. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned shortcomings of the prior art by providing a method and system for extracting crop plant height based on the adaptive CSF algorithm, so as to solve the problems of difficult parameter adjustment, poor adaptability, low processing efficiency and poor boundary quality of traditional methods.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, a method for extracting crop plant height based on the adaptive CSF algorithm includes the following steps: S1. Obtain the original point cloud of the crop and filter the original point cloud to obtain a preprocessed point cloud. S2. Based on the local features of the preprocessed point cloud and the agronomic parameters of the crop, dynamically determine the ground point separation parameters of the CSF algorithm, and use the ground point separation parameters to execute the CSF algorithm to separate the preprocessed point cloud into ground points and non-ground points. S3. Based on the separated ground points, construct a digital elevation model and use an interpolation method to convert the elevation values ​​of non-ground points into height values ​​relative to the ground to obtain a normalized point cloud. S4. Rasterize the normalized point cloud to generate a canopy height model; S5. Extract crop plant height information from the canopy height model.

[0007] Furthermore, step S2 includes the following sub-steps: S21. Point cloud segmentation and density calculation: The preprocessed point cloud is segmented into voxels according to spatial location to obtain multiple point cloud blocks, and the voxel density and planar density of each point cloud block are calculated. S22. Crop Information Acquisition and Density Level Classification: Acquire the crop type and key parameters of the crop, including row spacing and expected plant height; classify the point cloud blocks into density levels based on the voxel density, planar density, and crop type. S23. Initial Parameter Selection and Adaptive Adjustment Based on Growth Stage: From the multidimensional parameter mapping table, select the corresponding basic parameters of the CSF algorithm according to the density level and crop type, including cloth resolution, classification threshold, and cloth stiffness; wherein, the cloth resolution is adjusted according to the row spacing, and the classification threshold is adjusted according to the expected plant height; the crop growth stage is determined by the preliminary estimation of the canopy height model, and the cloth stiffness and the classification threshold are corrected according to the growth stage to obtain the final ground point separation parameters; S24, CSF Independent Separation: Using the final ground point separation parameters, the CSF algorithm is executed independently on each point cloud block to obtain the preliminary separated ground points and non-ground points of each point cloud block; S25. Overlapping Area Voting Fusion: For points located within the overlapping area of ​​adjacent point cloud blocks, a voting fusion mechanism is used to determine their final classification. S26. Quality check: Perform a quality check on the separation results to verify whether the proportion of ground points is within the preset range. If it exceeds the preset range, automatically adjust the classification threshold and re-execute S24 until the quality requirements are met or the preset maximum number of retries is reached.

[0008] Furthermore, in step S22, the point cloud blocks are classified into density levels based on the voxel density, planar density, and crop type, including: Sparseness: For maize, voxel density <300 points / m³ and planar density <800 points / m²; for wheat and rice, voxel density <600 points / m³ and planar density <1200 points / m². Medium: For corn, the voxel density is 300-1500 points / m³ and the planar density is 800-3500 points / m²; for wheat and rice, the voxel density is 600-2500 points / m³ and the planar density is 1200-5000 points / m². Dense: For corn, the voxel density is 1500~4000 points / m³ and the planar density is 3500~9000 points / m²; for wheat and rice, the voxel density is 2500~6000 points / m³ and the planar density is 5000~12000 points / m². Extremely dense: For corn, voxel density > 4000 points / m³ and planar density > 9000 points / m²; for wheat and rice, voxel density > 6000 points / m³ and planar density > 12000 points / m².

[0009] Furthermore, in step S23, adjusting the fabric resolution according to the row spacing includes: For corn, set the fabric resolution to 0.7 times the corn row spacing; For wheat, set the fabric resolution to 1.5 times the wheat row spacing; For rice, set the fabric resolution to 1.2 times the rice row spacing; Adjusting the classification threshold based on the expected plant height includes: For maize crops, the classification threshold is set at 6% of the expected maize plant height; For wheat crops, the classification threshold is set at 12% of the expected wheat plant height. For rice, the classification threshold is set at 10% of the expected plant height. Furthermore, in step S23, the correction of the fabric stiffness and the classification threshold based on the growth stage includes: When the extracted crop plant height is less than 20% of the expected crop plant height, it is determined to be in the seedling stage. The fabric stiffness value is adjusted to 2-3, and the classification threshold is reduced by 50%. When the extracted crop height is 20%-60% of the expected crop height, it is determined to be the jointing stage. The fabric hardness value is adjusted to 1-2, and the classification threshold is not adjusted. When the extracted crop height is greater than 60% of the expected crop height, it is determined to be in the maturity stage. The fabric stiffness adjustment value is adjusted to 1, and the classification threshold is increased by 20%.

[0010] Furthermore, in S25, the voting fusion mechanism adopts the majority voting method or the weighted voting method to classify and determine the points located in the overlapping area multiple times, and takes the majority result or the weighted result as the final classification.

[0011] Furthermore, in S26, the preset range is 10%-50% of the total number of ground points; if the proportion of ground points is less than 5%, the classification threshold is increased by 40% and S24 is re-executed; if the proportion of ground points is between 5% and 10%, the classification threshold is increased by 25% and S24 is re-executed; if the proportion of ground points is greater than 50%, the classification threshold is decreased by 20% and S24 is re-executed; if the proportion of ground points is greater than 70%, the classification threshold is decreased by 35% and S24 is re-executed.

[0012] Furthermore, in step S4, the normalized point cloud is projected onto a two-dimensional plane and rasterized. The size of each raster is adaptively set to 3-10cm according to the point cloud density. The maximum height value of all points within the range of each raster is taken as the canopy height at that location. Interpolation is performed to fill in the raster with holes, thereby generating a canopy height model.

[0013] Furthermore, in step S5, plant height statistics are extracted from the generated canopy height model, including: Maximum value, used to characterize the highest plant height of crops; The 95th percentile is used to characterize the overall height of crop growth. The 99th percentile is used to characterize the height of tall crops after excluding extreme outliers.

[0014] Secondly, a crop plant height extraction system based on the adaptive CSF algorithm includes: The point cloud preprocessing module is used to acquire the original point cloud of crops and filter the original point cloud to obtain a preprocessed point cloud. An adaptive CSF ground separation module is used to dynamically determine the ground point separation parameters of the CSF algorithm based on the local features of the preprocessed point cloud and the agronomic parameters of the crop, and to execute the CSF algorithm using the ground point separation parameters to separate the preprocessed point cloud into ground points and non-ground points. The height normalization module is used to build a digital elevation model based on the separated ground points. It uses an interpolation method to convert the elevation values ​​of non-ground points into height values ​​relative to the ground, thus obtaining a normalized point cloud. The canopy height model generation module is used to rasterize the normalized point cloud to generate a canopy height model. The plant height extraction module is used to extract crop plant height information from the canopy height model.

[0015] The crop plant height extraction method and system based on the adaptive CSF algorithm provided by this invention have the following beneficial effects: This invention solves the problem of manual parameter tuning required by traditional CSF algorithms through an innovative multi-dimensional adaptive parameter selection mechanism, achieving fully automated processing. By comprehensively considering point cloud density, crop type, growth stage, and planting pattern in its parameter mapping strategy, it significantly improves the algorithm's adaptability to different crops and growth conditions, avoiding common problems such as misjudgment of corn furrows and failure to detect lodging in wheat. Through block-based parallel processing and a voting fusion mechanism, it greatly improves the processing efficiency of large-scale point clouds while ensuring boundary quality. The integrated quality check mechanism automatically verifies the rationality of ground point separation results and performs parameter correction, improving the system's robustness. Using quantile height instead of maximum value improves the robustness of plant height measurement. Associating the cloth_resolution parameter with crop row spacing, the class_threshold parameter with expected plant height, and the rigidity parameter with growth stage gives the parameter settings clear agronomical significance. This invention has good scalability, supports multiple crop types (corn, wheat, rice, etc.) and acquisition platforms, and can be widely applied in precision agriculture, crop phenotyping, and field management. Attached Figure Description

[0016] Figure 1 This is a flowchart of the crop plant height extraction method and system based on the adaptive CSF algorithm in Example 1.

[0017] Figure 2 This is a schematic diagram of point cloud voxel segmentation and overlapping regions in Example 1. The diagram shows a large-scale point cloud divided into multiple blocks (block 1, block 2, block 3, etc.), with the size of each block adaptively determined based on the point cloud scale and density (typically 30-100 m). Overlapping regions are set between adjacent blocks, with an overlap ratio of 15-25%. Points within the overlapping regions are processed by multiple blocks, and the final classification is determined through a voting fusion mechanism to ensure the continuity and accuracy of classification at block boundaries.

[0018] Figure 3This is a diagram illustrating the density level classification and parameter mapping relationship in Example 1. The horizontal axis represents point cloud density (including both voxel density and planar density), and the vertical axis represents CSF parameter values ​​(fabric resolution, classification threshold, and fabric stiffness). The broken line in the diagram represents the difference in density thresholds between corn (solid line) and wheat / rice (dashed line). The dividing points for the four density levels—sparse, medium, dense, and extremely dense—varie depending on the crop type. The diagram also marks the parameter value range corresponding to each density level, clearly demonstrating the mapping relationship between density and parameters. For example, for corn, the dense level (voxel density 1500-4000 points / m³) corresponds to a fabric resolution of approximately 0.3m and a classification threshold of approximately 0.1m.

[0019] Figure 4 This is a schematic diagram of the voting fusion mechanism in Example 1. The diagram shows that a point in the overlapping area is processed by multiple blocks (block 1, block 2, and block 3). For a point P in the overlapping area, block 1 determines it as a ground point, block 2 determines it as a non-ground point, and block 3 determines it as a ground point. The voting fusion module counts the voting results (2 votes for ground point and 1 vote for non-ground point) and determines the final classification of the point as a ground point according to the majority voting method. The diagram also shows the case of weighted voting method, where different blocks obtain different weights according to their density and parameter confidence. Detailed Implementation

[0020] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0021] Example 1 This embodiment provides a method for extracting crop plant height based on the adaptive CSF algorithm, referencing... Figure 1 Specifically, it includes the following: S1. Obtain the original point cloud of the crop and filter the original point cloud to obtain the preprocessed point cloud. This embodiment performs radius filtering on the original point cloud generated by SLAM to remove noise points and outliers, resulting in a preprocessed point cloud that provides high-quality point cloud data for subsequent processing.

[0022] S2. Based on the local features of the preprocessed point cloud and the agronomic parameters of the crops, dynamically determine the ground point separation parameters of the CSF algorithm, and execute the CSF algorithm using the ground point separation parameters to separate the preprocessed point cloud into ground points and non-ground points; Reference Figure 2 It includes the following sub-steps: S21. Point cloud segmentation and density calculation: The preprocessed point cloud is segmented into voxels according to spatial location to obtain multiple point cloud blocks, and the voxel density (number of points / 3D volume) and planar density (number of points / 2D area) of each point cloud block are calculated. In one specific embodiment, the voxel segmentation adopts an adaptive strategy: the block size is determined based on the total size and average density of the point cloud, ranging from 30 to 100 meters; an overlap area of ​​15-25% is set between adjacent blocks to ensure that points at the boundary are processed by multiple blocks; the size of the overlap area is dynamically adjusted according to the point cloud density, with a larger overlap ratio as the density increases.

[0023] S22. Crop Information Acquisition and Density Classification: Acquire crop types and key parameters of crops, including row spacing and expected plant height; classify point cloud blocks into density levels based on voxel density, planar density, and crop type. Crop type information can be obtained through one of the following methods: (1) The crop type is input or selected by the user through the user interface (preferred method, the user can select corn, wheat, rice, etc. from the preset list to ensure accuracy); (2) Automatically identify crop types based on the geometric features of point clouds (such as row spacing pattern, plant height range, and canopy density); (3) Combine RGB images to identify crop types using image recognition technology (use deep learning models to perform image recognition of crop types); (4) Read the crop type information of the plot from the field management system or agricultural database (applicable to farms that have established crop files). Automatic identification can be used when there is no user input, but user input is preferred to ensure accuracy.

[0024] In some embodiments, point cloud patches are classified into density levels based on voxel density, planar density, and crop type, with reference to... Figure 3 ,include: Sparse: For corn, voxel density <300 points / m³ and planar density <800 points / m²; for wheat and rice, voxel density <600 points / m³ and planar density <1200 points / m²; the basic parameters are cloth_resolution=1.0m, class_threshold=0.3m, rigidity=3, but cloth_resolution will be adjusted according to the crop row spacing (approximately 0.42m for corn and approximately 0.225m for wheat), and class_threshold will be adjusted according to 6-12% of the expected plant height; suitable for airborne lidar data or long-range scanning.

[0025] Medium: For maize, the voxel density is 300-1500 points / m³ and the planar density is 800-3500 points / m²; for wheat and rice, the voxel density is 600-2500 points / m³ and the planar density is 1200-5000 points / m²; the basic parameters are cloth_resolution=0.5m, class_threshold=0.2m, rigidity=2, which can also be adjusted according to crop characteristics; suitable for vehicle-mounted scanning or medium-distance acquisition.

[0026] Dense: For maize, the voxel density is 1500~4000 points / m³ and the planar density is 3500~9000 points / m²; for wheat and rice, the voxel density is 2500~6000 points / m³ and the planar density is 5000~12000 points / m²; the basic parameters are cloth_resolution=0.3m, class_threshold=0.1m, rigidity=1, which can be adjusted according to crop characteristics; suitable for close-range handheld scanning.

[0027] Extremely dense: For maize, voxel density > 4000 points / m³ and planar density > 9000 points / m²; for wheat and rice, voxel density > 6000 points / m³ and planar density > 12000 points / m²; basic parameters are cloth_resolution=0.2m, class_threshold=0.05m, rigidity=1, which are adjusted according to crop characteristics; suitable for high-precision static scanning.

[0028] S23. Initial Parameter Selection and Adaptive Adjustment Based on Growth Stage: From the multidimensional parameter mapping table, select the corresponding basic parameters of the CSF algorithm according to the density level and crop type, including cloth resolution, classification threshold, and cloth stiffness; wherein, the cloth resolution is adjusted according to the row spacing, and the classification threshold is adjusted according to the expected plant height; the crop growth stage is determined by the preliminary estimation of the canopy height model, and the cloth stiffness and the classification threshold are corrected according to the growth stage to obtain the final ground point separation parameters; In some embodiments, the crop growth stage is initially estimated and determined using a canopy height model (CHM). For the point cloud processed for the first time, if the user provides sowing date or growth cycle information, the current growth stage is estimated based on the crop growth model (preferred option). Otherwise, an intermediate parameter value (rigidity = 2, class_threshold not adjusted) is used for the first ground separation to generate a preliminary canopy height model (CHM) and determine the growth stage. Then, the ground separation is re-executed using the adjusted parameters (alternative option). The rigidity parameter and class_threshold correction coefficient are adjusted according to the growth stage.

[0029] In some embodiments, adjusting the fabric resolution according to the row spacing includes: For crops with the same row spacing as corn (60-70cm), set the fabric resolution to 0.7 times the corn row spacing (approximately 0.42m) to avoid the furrows being misjudged as ground undulations; For crops with narrow row spacing (15-20cm), set the fabric resolution to 1.5 times the wheat row spacing (approximately 0.225m) to capture the overall terrain. For rice, set the fabric resolution to 1.2 times the rice row spacing; In some embodiments, adjusting the classification threshold based on the expected plant height includes: For maize crops, the classification threshold is set at 6% of the expected maize plant height (approximately 0.12-0.15m). For wheat crops, the classification threshold is set at 12% of the expected wheat plant height (approximately 0.08-0.11m). For rice crops, the classification threshold is set at 10% of the expected plant height (approximately 0.08-0.12m). For planting patterns with obvious ridge and furrow structures, the classification threshold is further increased to accommodate the ridge height (usually 10-20 cm). This mechanism ensures that the parameter settings have clear agronomic significance and avoids the problem of parameters being disconnected from crop characteristics in traditional methods.

[0030] In some embodiments, the fabric stiffness and classification threshold are adjusted according to the growth stage, including: When the extracted crop height is less than 20% of the expected crop height, it is determined to be in the seedling stage. At this time, the plants are sparse and the ground features are obvious. The fabric hardness value is adjusted to 2-3 to make the fabric harder to fit the real terrain. At the same time, the classification threshold is reduced by 50% to more strictly identify ground points. When the extracted crop height is 20%-60% of the expected crop height, it is determined to be the jointing stage. At this time, the plants are gradually becoming denser but have not completely covered the ground. The fabric hardness value is adjusted to 1-2, and the classification threshold is not adjusted. When the extracted crop height is greater than 60% of the expected crop height, it is determined to be in the mature stage. At this time, the canopy is dense. The cloth hardness adjustment value is adjusted to 1 to make the cloth soft so that it can penetrate the dense canopy to find the ground point. The classification threshold is increased by 20% to adapt to the canopy's shading of the ground.

[0031] In some embodiments, for point clouds processed for the first time, the following initialization strategy is adopted for growth stage determination: (1) Preferred scheme: If the user provides sowing date or growth cycle information, the system estimates the current growth stage based on the crop growth model (corn growth cycle is about 120-140 days, wheat about 180-220 days, and rice about 120-160 days); (2) Alternative scheme: The system performs the first ground separation using intermediate parameter values ​​(rigidity=2, class_threshold is not adjusted), generates a preliminary canopy height model CHM, determines the growth stage, and then performs ground separation again using the adjusted parameters; (3) Continuous monitoring scenario: The system estimates the growth stage by referring to the growth stage and time interval of the last measurement. This mechanism solves the problem that the rigidity parameter in traditional methods is only related to density and does not consider the crop growth stage, significantly improving the accuracy of ground point identification at different growth stages.

[0032] S24, CSF Independent Separation: Using the final ground point separation parameters, the CSF algorithm is executed independently on each point cloud block to obtain the preliminary separated ground points and non-ground points of each point cloud block; S25. Overlapping Area Voting Fusion: For points located within the overlapping area of ​​adjacent point cloud blocks, a voting fusion mechanism is used to determine their final classification. refer to Figure 4 The voting fusion mechanism includes majority voting and weighted voting: majority voting counts the number of times each point in the overlapping area is identified as a ground point and a non-ground point, and takes the majority result; weighted voting assigns weights according to the density and parameter confidence of each block and performs weighted voting; boundary smoothing processes perform secondary checks on areas with inconsistent voting results to ensure the continuity of classification.

[0033] S26. Quality check: Perform a quality check on the separation results to verify whether the proportion of ground points is within the preset range. If it exceeds the preset range, automatically adjust the classification threshold and re-execute S24 until the quality requirements are met or the preset maximum number of retries is reached.

[0034] Specifically, the system checks the proportion of ground points to the total number of points. Normally, this proportion should be between 10% and 50%. If the proportion is less than 10%, the judgment parameters may be too strict, leading to insufficient ground point identification. The system automatically increases the `class_threshold` parameter (by 20%-30% of the original value) and re-executes CSF ground separation. If the proportion is greater than 50%, the judgment parameters may be too lenient, causing some plant points to be misclassified as ground points. The system automatically decreases the `class_threshold` parameter. The system also checks the continuity and rationality of the CHM (Crop Height Model). If there are large areas of abnormally low values ​​(continuous areas with heights less than 10% of the expected plant height exceeding 5% of the total area), it may indicate lodging or missing seedlings. The system marks these areas and processes them separately during plant height statistics. Finally, the system verifies whether the extracted plant height is within the reasonable range (50%-150%) of the expected crop height. If it exceeds this range, a parameter re-evaluation process is triggered. The specific adjustment strategies are as follows: The preset range is 10%-50% of the total number of ground points. If the proportion of ground points is less than 10%, such as between 5% and 10%, the classification threshold is increased by 25% (taking the midpoint between 20% and 30%) and S24 is executed again. If the proportion of ground points is less than 5%, the classification threshold is increased by 40% and S24 is executed again. If the proportion of ground points is greater than 50%, the classification threshold is decreased by 20% and S24 is executed again. If the proportion of ground points is greater than 70%, the classification threshold is decreased by 35% and S24 is executed again. The maximum number of retries is 3. If the conditions are still not met, the system is marked and a manual check is prompted.

[0035] In one specific embodiment, the parameter mapping table supports optimized configurations for various crop types and growth stages: For maize, the parameter configuration includes a row spacing of 0.6m, expected plant height of 2.0-2.5m, a base coefficient of cloth_resolution of 0.7 (actual value approximately 0.42m), a class_threshold ratio of 6%, a seedling rigidity of 3 with a 50% reduction in class_threshold, a jointing stage rigidity of 2, and a maturity stage rigidity of 1 with a 20% increase in class_threshold; for wheat, the parameter configuration includes a row spacing of 0.15m, an expected plant height of 0.7-0.9m, and a cloth_resolution coefficient of 0.7. The base coefficient for lution is 1.5 (actual value approximately 0.225m), and the class_threshold ratio is 12%. The adjustment rules for growth stage parameters are similar to those for corn. The parameter configuration for rice includes row spacing of 0.2m, expected plant height of 0.8-1.2m, base coefficient for cloth_resolution of 1.2 (actual value approximately 0.24m), and class_threshold ratio of 10%. For planting methods with ridge and furrow structures, cloth_resolution needs to be greater than row spacing to avoid misjudging ridges and furrows as ground undulations, and class_threshold needs to be increased to adapt to ridge height. Users can customize parameter combinations according to actual needs and save them to the mapping table.

[0036] S3. Based on the separated ground points, construct a digital elevation model, and use inverse distance weighted (IDW) or Kriging interpolation methods to construct a continuous ground elevation surface. Convert the absolute height of non-ground points into height values ​​relative to the ground to obtain a normalized point cloud. S4. Rasterize the normalized point cloud to generate a canopy height model; The normalized point cloud is projected onto a two-dimensional plane and rasterized. The size of each grid is adaptively set to 3-10 cm according to the point cloud density. The maximum height value of all points within each grid is taken as the canopy height at that location. Grids with holes are filled by interpolation to generate a canopy height model.

[0037] The CHM grid resolution is adaptively adjusted according to the point cloud density, with 8-10cm resolution used in sparse areas and 3-5cm resolution used in dense areas.

[0038] S5. Extract crop plant height information from the canopy height model, including: Maximum value, used to characterize the highest plant height of crops; The 95th percentile is used to characterize the overall height of crop growth. The 99th percentile is used to characterize the height of tall crops after excluding extreme outliers.

[0039] This embodiment also provides a crop plant height extraction system based on the adaptive CSF algorithm, including a point cloud data acquisition unit, a data processing unit and a data storage unit; The point cloud data acquisition unit is based on SLAM (Simultaneous Localization and Mapping) technology and uses LiDAR or RGB-D depth camera to acquire 3D point cloud data in real time. It can be mounted on a variety of platforms, including drones, ground vehicles, and handheld devices.

[0040] The data processing unit includes a CPU and GPU collaborative processing architecture: the CPU is responsible for serial tasks such as control flow, parameter selection, and voting fusion; the GPU is responsible for parallel computing-intensive tasks such as point cloud filtering, CSF calculation, and rasterization; a pipelined processing mechanism is adopted to process data while it is being acquired, thereby improving overall efficiency.

[0041] The data processing unit includes the following modules: The point cloud preprocessing module is used to acquire the original point cloud of crops and filter the original point cloud to obtain the preprocessed point cloud. An adaptive CSF ground separation module is used to dynamically determine the ground point separation parameters of the CSF algorithm based on the local features of the preprocessed point cloud and the agronomic parameters of the crop, and to execute the CSF algorithm using the ground point separation parameters to separate the preprocessed point cloud into ground points and non-ground points. The height normalization module is used to build a digital elevation model based on the separated ground points. It uses an interpolation method to convert the elevation values ​​of non-ground points into height values ​​relative to the ground, thus obtaining a normalized point cloud. The canopy height model generation module is used to rasterize the normalized point cloud to generate a canopy height model. The plant height extraction module is used to extract crop plant height information from the canopy height model.

[0042] The adaptive CSF ground separation module employs an improved cloth simulation algorithm: it initializes a virtual cloth mesh with a resolution determined by the cloth_resolution parameter; it projects the point cloud onto the cloth nodes; it simulates gravity and iteratively updates the cloth node positions; it determines ground points based on the distance between a node and its nearest point, with the distance threshold determined by the class_threshold parameter; and it controls the rigidity of the cloth, affecting its adaptability to terrain undulations.

[0043] This embodiment of the system generates a 3D point cloud map using SLAM technology. It employs an innovative multi-dimensional adaptive parameter selection mechanism to separate ground points from the point cloud, thereby generating a canopy height model (CHM) and extracting crop plant height information. The core innovation lies in the adaptive CSF ground separation module. This module comprehensively considers multiple factors, including point cloud density characteristics (voxel density and planar density), crop type (corn, wheat, rice, etc.), crop growth stage (seedling stage, jointing stage, maturity stage), and planting pattern (row spacing, furrow structure), dynamically adjusting CSF algorithm parameters (cloth_resolution, class_threshold, rigidity, etc.) to achieve precise adaptive processing for different crops and growth conditions. The system also supports large-scale point cloud block processing, ensuring seamless boundary stitching through an overlapping area voting fusion mechanism, and integrating a quality check mechanism to verify and automatically correct the ground point separation results. This invention solves the problems of traditional methods requiring manual parameter tuning, poor adaptability, low processing efficiency, and lack of agronomic knowledge. It achieves fully automated, high-precision, and highly robust crop plant height measurement, and can be widely applied to precision agriculture, crop phenotypic analysis, and field management.

[0044] Example 2 - Measurement of corn plant height using a handheld device; This embodiment uses a handheld device equipped with an RGB-D depth camera to collect point cloud data and measure plant height in a cornfield. The specific steps are as follows: Step 1: Point cloud acquisition; The operator walks along the field ridges with a handheld device, while an RGB-D camera captures depth images at a frame rate of 30fps. A SLAM algorithm processes the image sequence in real time to generate a 3D point cloud map. The data acquisition area is a 20m × 30m cornfield, and the generated point cloud contains approximately 5 million points.

[0045] Step 2: Point cloud preprocessing; Radius filtering was applied to the original point cloud. The search radius was set to r = 0.05m, and the minimum number of neighbors n_min = 5. For each point, neighbors were searched within its radius r. If the number of neighbors was less than n_min, the point was considered noise and removed. After processing, the point cloud was reduced to 4.8 million points, removing approximately 4% of the noise.

[0046] Step 3: Adaptive ground point separation; (3.1) Voxel Blocking: Based on the point cloud range (X: 0-20m, Y: 0-30m) and the total number of points, determine the block size to be 50m × 50m. Since the actual area is smaller than the block size, the entire point cloud is treated as a single block. A 20% overlap is set, although not required in this example, but this step is retained to demonstrate the complete process.

[0047] (3.2) Density Analysis: Calculate the voxel density of this block. The Z-direction range of the point cloud is 0-2.5m (from the ground to the top of the corn), the volume V = 20 × 30 × 2.5 = 1500m³, and the voxel density = 4800000 / 1500 = 3200 points / m³. Calculate the planar density: the area A = 20 × 30 = 600m², and the planar density = 4800000 / 600 = 8000 points / m².

[0048] (3.3) Parameter selection: First, based on the density value, this block belongs to the DENSE level (the density threshold for maize: voxel density 1500-4000 points / m³ and planar density 3500-9000 points / m²). The crop type is maize, the row spacing is 0.6m, and the expected plant height is 2.5m. Adjust the parameters according to the crop type: cloth_resolution = row spacing × 0.7 = 0.6 × 0.7 = 0.42m (greater than the base value of 0.3m to avoid misjudgment of furrows); class_threshold = expected plant height × 6% = 2.5 × 0.06 = 0.15m (greater than the base value of 0.1m). The current plant height is estimated to be about 2.3m through preliminary CHM, which accounts for 92% of the expected plant height, and is judged to be in the maturity stage. Therefore, rigidity = 1 (soft cloth penetrates dense canopy), and class_threshold is further increased by 20% to 0.18m. Final parameters: cloth_resolution = 0.42m, class_threshold = 0.18m, rigidity = 1, iterations = 500.

[0049] (3.4) CSF Ground Separation: Initialize a cloth mesh of 30m / 0.42m × 20m / 0.42m = 72 × 48. Project the point cloud onto the mesh nodes, and record the height of the lowest point below each node. Simulate the cloth falling under gravity, updating the node position in each iteration. Rigidity=1 indicates that the cloth is relatively soft and can penetrate the dense corn canopy to find the ground. After 500 iterations, the cloth stably adheres to the ground. Traverse all points and calculate their vertical distance to the nearest cloth node. Points with a distance less than 0.18m are identified as ground points, otherwise they are non-ground points. Since the cloth_resolution (0.42m) is adjusted according to the corn row spacing, the cloth mesh span is slightly larger than the furrow spacing, effectively avoiding the identification of furrows as ground undulations. Finally, 780,000 ground points and 4,020,000 non-ground points (corn plants) were identified.

[0050] (3.5) Quality Inspection: The ground point ratio = 780,000 / 4,800,000 = 16.25%, which is within a reasonable range (10%-50%), and the verification is passed. CHM continuity was checked, and no large areas of abnormally low values ​​were found. The extracted plant height of approximately 2.35m is within the expected plant height range (50%-150% of 2.0-2.5m), the parameter settings are reasonable, and no re-execution is required.

[0051] Step 4: High normalization; A DEM was generated using 800,000 ground points. Inverse Distance Weighted (IDW) interpolation was employed, with a search radius of 2m and a weighting exponent of 2. For each location (x, y) on the plane, the ground points within the search radius were weighted according to distance and interpolated to obtain the ground elevation z_ground(x, y). A DEM raster with a resolution of 0.05m was generated. For each non-ground point (x, y, z), the ground elevation z_ground was obtained by querying the DEM, and the normalized height was calculated: h = z - z_ground.

[0052] Step 5: CHM generation; (5.1) Rasterization: Project the normalized point cloud onto a two-dimensional plane and set the raster resolution to 5cm × 5cm. For each raster, take the maximum normalized height of all points within its range as the canopy height of that raster. Generate a 400×600 CHM raster image.

[0053] (5.2) Hole filling: Some grid cells have no points due to sparse point clouds. Bilinear interpolation is used to fill these holes. For empty grid cells, search for their four nearest non-empty grid cells and interpolate based on the distance.

[0054] (5.3) Gaussian smoothing: Apply a 3×3 Gaussian kernel to smooth the CHM, sigma=1.0. Remove isolated noise peaks to make the canopy height distribution more continuous and smooth.

[0055] Step 6: Plant height extraction; Plant height information was obtained from the CHM (Constant Growth Metrics and Analysis): Maximum plant height = 2.48m (maximum value in the CHM); 95th percentile plant height = 2.35m (value at the 95th percentile after sorting); 99th percentile plant height = 2.42m; Average plant height = 2.18m. Using the 95th percentile as the representative plant height for this region provides a more robust reflection of overall growth.

[0056] Validation results: Ten corn plants were randomly selected for manual measurement, and the average plant height was 2.37m. The 95th percentile plant height extracted by the system was 2.35m, with an error of only 0.02m (0.8%), verifying the accuracy of this method.

[0057] Example 3 - Monitoring wheat plant height in the field based on unmanned aerial vehicles; This embodiment uses a drone equipped with LiDAR to perform high-altitude scanning of a 100m×100m wheat field. The total number of point clouds is approximately 20 million points, requiring a block-based processing strategy.

[0058] Step 1: Point cloud acquisition and preprocessing; The drone flew at an altitude of 50m, and the lidar scan density was approximately 200 points / m². After data collection, radius filtering (r=0.1m, n_min=3) was performed to remove approximately 2% of noise points.

[0059] Step 2: Adaptive block processing; (2.1) Determine the block size: Based on the point cloud scale and density, select a block size of 50m×50m with an overlap ratio of 20% (10m). Divide the 100m×100m region into 3×3=9 overlapping blocks.

[0060] (2.2) Block-by-block processing: The following process is executed independently for each block: a) Extract the point cloud within the block (approximately 2.2 million points / block). b) Calculate density: Voxel density is approximately 600 points / m³, planar density is approximately 2000 points / m², crop type is wheat, based on wheat density thresholds (600-2500 points / m³ voxel density and 1200-5000 points / m² planar density), this block belongs to the MEDIUM level. c) Select parameters: Crop type is wheat, row spacing is 0.15m, and expected plant height is 0.85m. Adjust according to crop type: cloth_resolution = row spacing × 1.5 = 0.15 × 1.5 = 0.225m; class_threshold = expected plant height × 12% = 0.85 × 0.12 = 0.102m. Based on preliminary CHM estimation, the current plant height is approximately 0.82m, which is 96% of the expected plant height, indicating maturity. Rigidity = 1, and class_threshold is increased by 20% to 0.122m. Final parameters: cloth_resolution = 0.225m, class_threshold = 0.122m, rigidity = 1; d) Perform CSF to identify ground points and non-ground points; e) Record the classification result for each point (ground = 0, non-ground = 1); (2.3) Voting Fusion: For points within the overlapping area (approximately 36% of the total number of points), each point is processed by 2-4 blocks. The number of times each point is classified as a ground point or a non-ground point is counted, and the final classification is determined by majority voting. For example, if a point is processed by 3 blocks, is classified as a ground point twice, and is classified as a non-ground point once, then the final classification is a ground point.

[0061] Step 3: High normalization and CHM generation; A global DEM (10cm resolution) is generated using the fused ground points, and the non-ground points are height-normalized. The normalized point cloud is rasterized (8cm resolution) to generate a 1250×1250 CHM. A 5×5 Gaussian kernel smoothing is applied.

[0062] Step 4: Plant height extraction; Extracted from CHM: maximum plant height = 0.92m, 95th percentile = 0.85m, 99th percentile = 0.88m. The error compared to ground sampling measurements is less than 3cm.

[0063] Performance Analysis: Total processing time is approximately 120 seconds (including block partitioning, CSF, merging, and CHM generation). Compared to processing 20 million points directly without partitioning (which takes approximately 600 seconds and may result in memory overflow), this represents a 5x improvement in efficiency and a reduction in memory usage to 1 / 9 of the original. The voting merging mechanism effectively avoids classification errors at block boundaries, ensuring a smooth and continuous CHM in the boundary regions.

[0064] Example 4 - Multi-density mixed scene processing; This embodiment demonstrates the adaptive capability of the present invention for mixed density point clouds, with the data collection scenarios including densely planted greenhouse areas and sparsely planted field areas.

[0065] Scene description: Area A (Greenhouse): 30m×20m, handheld scanning, point cloud density 4500 points / m³, belonging to the DENSE level.

[0066] Area B (Daejeon): 70m×80m, airborne scanning, point cloud density 300 points / m³, belonging to the SPARSE level.

[0067] Processing procedure: The system automatically divides the entire area into multiple blocks, and analyzes the density and selects parameters for each block independently. Greenhouse area block: cloth_resolution=0.3m, class_threshold=0.1m, which can accurately identify the complex terrain and dense vegetation inside the greenhouse.

[0068] Datian region block: cloth_resolution=1.0m, class_threshold=0.3m, adapted to sparse point clouds, avoiding overfitting.

[0069] Results: The CHM quality in both regions is high. The greenhouse region captures detailed features, while the field region effectively filters out noise. Using fixed parameters results in either coarse classification in the greenhouse region or numerous misclassifications in the field region. The adaptive mechanism of this invention perfectly solves this problem.

[0070] Example 5 - Measurement of maize plant height during seedling stage (adaptive verification of growth stage); This embodiment uses a handheld device to measure the plant height of corn seedlings (15 days after sowing) to verify the effectiveness of the growth stage adaptive mechanism in the seedling stage.

[0071] Step 1: Point cloud acquisition; The data collection area was a 10m × 10m cornfield. Fifteen days after sowing, the corn plants were approximately 0.3m tall (12% of the expected plant height of 2.5m). The operator walked along the rows with a handheld RGB-D camera, generating a point cloud of approximately 2 million points.

[0072] Step 2: Point cloud preprocessing; Radius filtering (r=0.05m, n_min=5) was applied to the original point cloud to remove approximately 3% of noise points, resulting in a point cloud of approximately 1.94 million points.

[0073] Step 3: Density analysis and parameter selection; (3.1) Density calculation: The point cloud range in the Z direction is 0-0.35m (from the ground to the top of the seedling), the volume V = 10 × 10 × 0.35 = 35m³, the voxel density = 1940000 / 35 = 5,543 points / m³; the planar density = 1940000 / 100 = 19400 points / m². According to the density threshold of corn, this point cloud belongs to the VERY_DENSE level.

[0074] (3.2) Crop type and parameters: The user inputs the crop type as corn, row spacing as 0.6m, and expected plant height as 2.5m. Adjust the parameters according to the crop type: cloth_resolution = row spacing × 0.7 = 0.6 × 0.7 = 0.42m; class_threshold base value = expected plant height × 6% = 2.5 × 0.06 = 0.15m.

[0075] (3.3) Growth Stage Judgment: Since the user did not provide a sowing date, the system adopted an alternative scheme. First, the intermediate parameter values ​​(rigidity=2, class_threshold=0.15m) were used to perform the first ground separation and generate a preliminary CHM. The preliminary CHM showed an average plant height of about 0.28m, accounting for 11% of the expected plant height, which was judged to be in the seedling stage (<20%).

[0076] (3.4) Seedling stage parameter adjustment: Based on the growth stage adaptive mechanism, the seedling stage parameters are adjusted as follows: rigidity = 3 (using a stiffer fabric to conform to the actual terrain), class_threshold = 0.15m × 50% = 0.075m (reduced by 50% for more rigorous identification of ground points). Final parameters: cloth_resolution = 0.42m, class_threshold = 0.075m, rigidity = 3.

[0077] Step 4: CSF ground separation and effect comparison; CSF ground separation was re-executed using the adjusted parameters (rigidity=3, class_threshold=0.075m). The initial cloth grid was 10m / 0.42m × 10m / 0.42m = 24×24. Due to the sparse and low height of the corn seedlings, ground features were obvious. The rigid cloth with rigidity=3 accurately conformed to the terrain, avoiding misidentification of some roots or soil protrusions as plants. Ultimately, 350,000 ground points (18%) and 1,590,000 non-ground points (corn seedlings) were identified.

[0078] Comparative experiment: When using maturity parameters (rigidity=1, class_threshold=0.18m), the accuracy of ground point recognition decreased by about 15%, and some sparsely planted seedlings had their roots misidentified as ground, resulting in a plant height measurement that was about 0.04m lower. This verifies the necessity of growth stage adaptive mechanisms.

[0079] Step 5: High normalization and CHM generation; A DEM (5cm resolution) was generated using 350,000 ground points. Non-ground points were height-normalized. The normalized point cloud was rasterized (4cm resolution) to generate a 250×250 CHM. A 3×3 Gaussian kernel smoothing was applied to remove noise peaks.

[0080] Step 6: Plant height extraction and verification; Plant height statistics were extracted from CHM: maximum plant height = 0.33m, 95th percentile plant height = 0.28m, 99th percentile = 0.31m, average plant height = 0.26m.

[0081] Verification Results: Twenty randomly selected maize seedlings were manually measured, with an average plant height of 0.29m. The 95th percentile plant height extracted by the system was 0.28m, with an error of only 0.01m (3.4%), verifying the accuracy of the invention during the seedling stage. Quality Inspection: The proportion of ground-level points (18%) is within a reasonable range (10%-50%), and the extracted plant height of 0.28m is within 50%-150% of the expected plant height of 2.5m (accounting for 11%, slightly lower but within the normal range for the seedling stage), indicating that the parameter settings are reasonable.

[0082] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for extracting crop plant height based on the adaptive CSF algorithm, characterized in that, Includes the following steps: S1. Obtain the original point cloud of the crop and filter the original point cloud to obtain a preprocessed point cloud. S2. Based on the local features of the preprocessed point cloud and the agronomic parameters of the crop, dynamically determine the ground point separation parameters of the CSF algorithm, and use the ground point separation parameters to execute the CSF algorithm to separate the preprocessed point cloud into ground points and non-ground points. S3. Based on the separated ground points, construct a digital elevation model and use an interpolation method to convert the elevation values ​​of non-ground points into height values ​​relative to the ground to obtain a normalized point cloud. S4. Rasterize the normalized point cloud to generate a canopy height model; S5. Extract crop plant height information from the canopy height model.

2. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Point cloud segmentation and density calculation: The preprocessed point cloud is segmented into voxels according to spatial location to obtain multiple point cloud blocks, and the voxel density and planar density of each point cloud block are calculated. S22. Crop Information Acquisition and Density Level Classification: Acquire the crop type and key parameters of the crop, including row spacing and expected plant height; classify the point cloud blocks into density levels based on the voxel density, planar density, and crop type. S23. Initial Parameter Selection and Adaptive Adjustment Based on Growth Stage: From the multidimensional parameter mapping table, select the corresponding basic parameters of the CSF algorithm according to the density level and crop type, including cloth resolution, classification threshold, and cloth stiffness; wherein, the cloth resolution is adjusted according to the row spacing, and the classification threshold is adjusted according to the expected plant height; the crop growth stage is determined by the preliminary estimation of the canopy height model, and the cloth stiffness and the classification threshold are corrected according to the growth stage to obtain the final ground point separation parameters; S24, CSF Independent Separation: Using the final ground point separation parameters, the CSF algorithm is executed independently on each point cloud block to obtain the preliminary separated ground points and non-ground points of each point cloud block; S25. Overlapping Area Voting Fusion: For points located within the overlapping area of ​​adjacent point cloud blocks, a voting fusion mechanism is used to determine their final classification. S26. Quality check: Perform a quality check on the separation results to verify whether the proportion of ground points is within the preset range. If it exceeds the preset range, automatically adjust the classification threshold and re-execute S24 until the quality requirements are met or the preset maximum number of retries is reached.

3. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 2, characterized in that, In step S22, the point cloud blocks are classified into density levels based on the voxel density, planar density, and crop type, including: Sparseness: For maize, voxel density <300 points / m³ and planar density <800 points / m²; for wheat and rice, voxel density <600 points / m³ and planar density <1200 points / m². Medium: For corn, the voxel density is 300-1500 points / m³ and the planar density is 800-3500 points / m²; for wheat and rice, the voxel density is 600-2500 points / m³ and the planar density is 1200-5000 points / m². Dense: For corn, the voxel density is 1500~4000 points / m³ and the planar density is 3500~9000 points / m²; for wheat and rice, the voxel density is 2500~6000 points / m³ and the planar density is 5000~12000 points / m². Extremely dense: For corn, voxel density > 4000 points / m³ and planar density > 9000 points / m²; for wheat and rice, voxel density > 6000 points / m³ and planar density > 12000 points / m².

4. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 2, characterized in that, In step S23, adjusting the fabric resolution according to the row spacing includes: For corn, set the fabric resolution to 0.7 times the corn row spacing; For wheat, set the fabric resolution to 1.5 times the wheat row spacing; For rice, set the fabric resolution to 1.2 times the rice row spacing; Adjusting the classification threshold based on the expected plant height includes: For maize crops, the classification threshold is set at 6% of the expected maize plant height; For wheat crops, the classification threshold is set at 12% of the expected wheat plant height. For rice crops, the classification threshold is set at 10% of the expected plant height.

5. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 2, characterized in that, In step S23, the fabric stiffness and the classification threshold are corrected according to the growth stage, including: When the extracted crop plant height is less than 20% of the expected crop plant height, it is determined to be in the seedling stage. The fabric stiffness value is adjusted to 2-3, and the classification threshold is reduced by 50%. When the extracted crop height is 20%-60% of the expected crop height, it is determined to be the jointing stage. The fabric hardness value is adjusted to 1-2, and the classification threshold is not adjusted. When the extracted crop height is greater than 60% of the expected crop height, it is determined to be in the maturity stage. The fabric stiffness adjustment value is adjusted to 1, and the classification threshold is increased by 20%.

6. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 2, characterized in that, In S25, the voting fusion mechanism adopts the majority voting method or the weighted voting method to classify points located in the overlapping area multiple times, and takes the majority result or the weighted result as the final classification.

7. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 2, characterized in that, In step S26, the preset range is 10%-50% of the total number of ground points; if the proportion of ground points is less than 5%, the classification threshold is increased by 40% and step S24 is executed again; if the proportion of ground points is between 5% and 10%, the classification threshold is increased by 25% and step S24 is executed again; if the proportion of ground points is greater than 50%, the classification threshold is decreased by 20% and step S24 is executed again; if the proportion of ground points is greater than 70%, the classification threshold is decreased by 35% and step S24 is executed again.

8. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 1, characterized in that, In step S4, the normalized point cloud is projected onto a two-dimensional plane and rasterized. The size of each raster is adaptively set to 3-10cm according to the point cloud density. The maximum height value of all points within the range of each raster is taken as the canopy height at that location. Interpolation is performed to fill in the raster with holes, thereby generating a canopy height model.

9. The method for extracting crop plant height based on the adaptive CSF algorithm according to claim 1, characterized in that, In step S5, plant height statistics are extracted from the generated canopy height model, including: Maximum value, used to characterize the highest plant height of crops; The 95th percentile is used to characterize the overall height of crop growth. The 99th percentile is used to characterize the height of tall crops after excluding extreme outliers.

10. A crop plant height extraction system based on an adaptive CSF algorithm that implements the method of any one of claims 1-9, characterized in that, include: The point cloud preprocessing module is used to acquire the original point cloud of crops and filter the original point cloud to obtain a preprocessed point cloud. An adaptive CSF ground separation module is used to dynamically determine the ground point separation parameters of the CSF algorithm based on the local features of the preprocessed point cloud and the agronomic parameters of the crop, and to execute the CSF algorithm using the ground point separation parameters to separate the preprocessed point cloud into ground points and non-ground points. The height normalization module is used to build a digital elevation model based on the separated ground points. It uses an interpolation method to convert the elevation values ​​of non-ground points into height values ​​relative to the ground, thus obtaining a normalized point cloud. The canopy height model generation module is used to rasterize the normalized point cloud to generate a canopy height model. The plant height extraction module is used to extract crop plant height information from the canopy height model.

Images