A farmland point cloud ground separation method and system based on topological persistence analysis

Abstract

The application discloses a farmland point cloud ground separation method and system based on topological persistence analysis, and belongs to the technical field of the cross of precision agriculture and point cloud data processing technology, and comprises the following steps: pre-processing and coordinate alignment of three-dimensional point cloud data of a farmland environment; projecting the three-dimensional point cloud to a two-dimensional plane and performing rasterization processing, and then generating a two-dimensional elevation field; topologically decomposing the two-dimensional elevation field by adopting a step-by-step threshold scanning, extracting connected components in each layer of binary results, and tracing the evolution process of the connected components between adjacent elevation layers; calculating topological persistence features of each connected component; classifying the ground and crops based on the topological persistence features and the spatial features of the connected components; and outputting the classified ground point cloud and crop point cloud. The application solves the technical problems of strong parameter dependence and poor robustness of the existing method, and provides high-quality basic data for precision agriculture application.

Description

Technical Field

[0001] This invention belongs to the technical field of precision agriculture and point cloud data processing technology, specifically involving a method and system for farmland point cloud ground separation based on topological persistence analysis. Background Technology

[0002] With the development of precision agriculture technology, farmland environmental perception based on 3D point clouds has become an important means of crop monitoring, yield prediction, and intelligent management. 3D sensors such as LiDAR and depth cameras can construct farmland point cloud maps in real time using SLAM technology. Farmland SLAM point clouds typically mix multiple targets such as ground, crops, and weeds; ground-crop separation is the most basic and critical task, directly affecting the accuracy of subsequent phenotypic parameter extraction.

[0003] Existing methods for separating farmland point clouds from the ground include various technical approaches, each with its own characteristics and applicable scenarios: (1) Fixed threshold method: This method distinguishes between ground and crops by setting a fixed height threshold. This method is computationally efficient and simple to implement. However, farmland has micro-topographical undulations such as furrows and ridges, and crops are relatively low in the early stages of growth, making it difficult for the fixed threshold to adapt to complex scenarios. In undulating terrain scenarios, this method achieves an F1-score (harmonic mean of precision and recall) of approximately 65-75%.

[0004] (2) RANSAC plane fitting method: Randomly sampled points are fitted to the ground plane, and points whose distance from the plane is less than a threshold are marked as ground. This method works well for flat ground. However, when there are obvious undulations in the ground, the single plane assumption no longer holds, and the RANSAC algorithm requires multiple iterations, resulting in high computational cost. The F1-score of this method is approximately 75-85%, and the computation time is long.

[0005] (3) Cloth Simulation Method (CSF): This method simulates a piece of cloth covering the surface of a point cloud under the influence of gravity, with the cloth nodes serving as the ground. This method has clear physical meaning and relatively few parameters. However, the algorithm involves complex physical simulation, resulting in low computational efficiency. Furthermore, in densely planted crop scenarios, the cloth may become embedded in the crop canopy, affecting the separation accuracy.

[0006] (4) Progressive Triangulation (PTD): This method gradually constructs a triangulation network starting from the lowest point, and determines whether a new point is part of the ground by using angle and distance criteria. This method is suitable for scenarios with undulating terrain. However, the algorithm relies on multiple manually set threshold parameters (such as angle threshold, distance threshold, etc.), and the parameter settings have a significant impact on the results, lacking adaptability.

[0007] (5) Deep learning method: using convolutional neural networks for point cloud semantic segmentation. This method can theoretically achieve high accuracy. However, it requires a large amount of manually labeled training data, which is costly; model training and inference require a lot of computing resources; the model's generalization ability depends on the diversity of scenarios covered by the training data; and the interpretability of deep learning models is relatively weak.

[0008] The above method has the following technical defects: The core flaw of the RANSAC method lies in its single-plane assumption. This method assumes the ground can be described by a single plane or a few planes, which holds true in flat scenarios. However, in farmland environments, the ground exhibits complex three-dimensional curved surface features due to the presence of micro-topographical structures such as furrows, ridges, and field ridges. When the ground elevation change exceeds 15cm, the single-plane assumption fails, leading to a large number of ground points being misclassified as crops. Furthermore, RANSAC requires a preset distance threshold, which has a decisive impact on the separation results, but lacks an adaptive selection mechanism. Experimental data shows that in undulating terrain scenarios, the recall rate of RANSAC drops below 75%.

[0009] While the PTD method can handle undulating terrain, it suffers from severe parameter dependency. This method requires manually setting as many as 5-8 parameters, such as angle thresholds, distance thresholds, and grid resolution. The optimal values ​​for these parameters vary significantly depending on the scene, lacking universality. For example, parameter settings suitable for cornfields (row spacing 0.75m) may be completely unsuitable for wheat fields (row spacing 0.25m). Furthermore, the PTD method, based on local geometric criteria, struggles to handle discrete crop distributions—when crop spacing is large, ground areas between crops may be incorrectly merged into the crop triangulation, leading to decreased accuracy.

[0010] At a deeper level, the fundamental reason these methods fail is that they only focus on local geometric features (such as flatness, angle, and distance), while ignoring the essential differences in topological structure between the ground and crops. The ground, as the basic structure of the scene, exhibits large-scale, continuous topological features in the elevation field; crops, as discretely distributed individuals, exhibit small-scale, discontinuous topological features in the elevation field. Traditional methods cannot capture this topological difference, thus exhibiting poor robustness in complex scenes.

[0011] Topological persistence analysis, as a mathematical tool, has been applied in multiple fields. In remote sensing image processing, topological persistence is used for terrain feature extraction and image vectorization; in computer vision, it is used for shape recognition and 3D reconstruction; and in medical image processing, it is used for vascular network analysis. These applications mainly focus on 2D images or specific 3D scenes.

[0012] In the field of point cloud processing, research has also explored the application of topological methods. For example, some researchers have used persistent homology analysis to analyze the topological structure of point clouds for noise filtering and feature extraction. However, these methods mainly focus on the global topological properties of point clouds for shape matching or classification tasks, rather than semantic segmentation. Summary of the Invention

[0013] The purpose of this invention is to address the above-mentioned shortcomings of the prior art by providing a method and system for farmland point cloud ground separation based on topological persistence analysis, so as to solve the problems of strong parameter dependence and poor robustness of the existing methods in complex farmland environments.

[0014] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, a method for separating farmland point clouds from the ground based on topological persistence analysis includes the following steps: S1. Acquire 3D point cloud data of the farmland environment and preprocess and align its coordinates. S2. Project the preprocessed and coordinate-aligned 3D point cloud data onto a 2D plane and perform rasterization to generate a 2D elevation field. S3. The two-dimensional elevation field is decomposed by stepwise threshold scanning, the connected components in the binarization result of each layer are extracted, and the evolution process of the connected components between adjacent elevation layers is tracked. S4. Calculate the topological persistence characteristics of each connected component; S5. Based on the topological persistence features and the spatial features of the connected components, classify each connected component into ground and crops; S6. Based on the classification results of the connected components, perform point cloud level annotation on the original three-dimensional point cloud data, and output the classified ground point cloud and crop point cloud.

[0015] Furthermore, in S1, the preprocessing involves filtering out outliers from the 3D point cloud data. This is done by using a statistical filter to remove noise points. The number of neighborhood points of the statistical filter is 10-30, and the standard deviation multiple is 1.5-3.0. The coordinate alignment adopts the principal component analysis method. The ground principal direction is obtained by calculating the eigenvalue decomposition of the three-dimensional point cloud covariance matrix. The ground principal direction is rotated and aligned to the positive Z-axis direction, thereby completing the coordinate alignment.

[0016] Furthermore, in step S2, during the generation of the two-dimensional elevation field, an adaptive selection strategy based on point cloud density is used to determine the raster resolution, specifically as follows: Calculate the projection density of the point cloud onto a two-dimensional plane, take the elevation feature value of all points in each grid cell, and use interpolation to complete empty grids; The choice of raster resolution r is as follows: For dense point clouds, the raster resolution is r = 0.03-0.08m; For medium-density point clouds, the raster resolution is r = 0.08-0.15m; For sparse point clouds, the raster resolution is r = 0.15-0.30m.

[0017] Furthermore, in S3, the two-dimensional elevation field is topologically decomposed by step-by-step threshold scanning, including setting 50-300 thresholds from high elevation to low elevation for binarization processing to generate corresponding binary image sequences, and extracting the connected components of each layer of binary image using 4-connectivity or 8-connectivity rules. By calculating the evolution process of overlapping pixels of connected components in adjacent layers, when the number of overlapping pixels of two connected components in adjacent layers is greater than 0, it is considered that there is an inheritance relationship between the two connected components in adjacent layers, and a decomposition matrix is ​​constructed to record the spatial distribution of each continuous connected component.

[0018] Furthermore, the evolution process of the connected components between adjacent elevation layers includes: Newborn, the first appearance of the connected component; Growth leads to an increase in the area of ​​interconnected components; Merge, where multiple connected components are combined into one; Demise: The connected component disappears or is absorbed by other connected components.

[0019] Furthermore, in step S4, the topological persistence feature includes the birth time, persistence length, and component area of ​​the connected component; the birth time, persistence length, and component area of ​​the connected component are combined into a persistent barcode, which serves as the topological feature descriptor of the connected component. Wherein, the birth time is the elevation threshold at the time when the connected component first appears; The persistence length is the threshold number of layers that the connected components continue to exist. The component area is the total number of grid cells covered by the connected components.

[0020] Furthermore, in S5, a triple joint determination is adopted using the persistence length criterion, the component area criterion, and the birth height criterion. Connected components that simultaneously satisfy all three criteria are marked as ground, and the remaining connected components are marked as crops or noise.

[0021] Furthermore, the threshold of the persistence length criterion is the 60-80 percentile of the persistence length distribution of all connected components, used to filter connected components that span multiple elevation layers. The threshold for the component area criterion is 0.5-10% of the total grid area, used to filter connected components that cover a large area; The threshold for the birth height criterion is the median or mean elevation, used to filter connected components with low elevations.

[0022] Furthermore, in step S6, point cloud level annotation is performed on the original three-dimensional point cloud data based on the classification results, including: marking non-ground connected components as crop candidate regions, filtering out tiny components with a total grid area of ​​0.05-0.5% to remove noise interference, and then projecting each point in the original three-dimensional point cloud onto the corresponding grid cell and querying its classification label to complete the point cloud level annotation.

[0023] Secondly, a farmland point cloud ground separation system is characterized by comprising a hardware layer, a software layer, and an interface layer, wherein the hardware layer, software layer, and interface layer interact through a data bus and support pipelined parallel processing. The hardware layer includes a data acquisition unit, a data storage unit, and a CPU+GPU heterogeneous computing processing unit. The software layer includes a point cloud preprocessing unit, an elevation field construction unit, a topology decomposition unit, a persistence analysis unit, and a ground crop classification unit, which are linked in sequence. The interface layer includes a data input interface, a parameter configuration interface, and a result output interface.

[0024] The method and system for farmland point cloud ground separation based on topological persistence analysis provided by this invention have the following beneficial effects: 1. This invention solves the technical problems of strong parameter dependence and poor robustness in existing methods, providing high-quality basic data for precision agriculture applications. It employs adaptive threshold calculation, eliminating the need for manually setting complex parameters. In complex farmland environments, the F1-score reaches 89.5%, significantly outperforming the fixed threshold method (F1=70.3%) and the RANSAC plane fitting method (F1=79.9%). Processing 5 million point cloud data points takes only 10 seconds, meeting near real-time requirements.

[0025] 2. Strong theoretical innovation: This invention applies topological data analysis theory to the field of farmland SLAM point cloud processing. By introducing the mathematical concept of topological persistence, it discovers and utilizes the essential difference between the ground and crops in the topological persistence of the elevation field, providing a solid theoretical basis for ground-crop separation.

[0026] 3. Strong adaptability and low parameter dependence: This invention eliminates the need for manually setting complex threshold parameters and can adaptively handle different scenarios. The persistence threshold is calculated adaptively using percentiles, the raster resolution is adaptively selected based on the point cloud density, and the area threshold is set proportionally based on the total raster area. Experiments show that this invention maintains stable high accuracy under different crop types, different growth stages, and different terrain conditions.

[0027] 4. High separation accuracy and robustness: In complex farmland environments containing furrow structures and height undulations of up to 15 cm, this invention achieves an F1-score of 89.5% for ground-crop separation, with a precision of 91.2% and a recall of 87.9%, significantly outperforming the traditional fixed threshold method (F1=70.3%) and the RANSAC plane fitting method (F1=79.9%). The F1-score improvement is approximately 19.2 percentage points compared to the fixed threshold method and approximately 9.6 percentage points compared to the RANSAC method.

[0028] 5. High computational efficiency, meeting real-time requirements; This invention adopts a CPU+GPU heterogeneous parallel computing architecture, which can process 5 million point cloud data in only 9.8 seconds and 10 million point cloud data in only 19.2 seconds on a standard workstation, meeting the near real-time requirements and can be integrated into a SLAM real-time mapping system to achieve online ground separation.

[0029] 6. Highly practical for engineering applications, easy to deploy and integrate; This invention provides a complete automated process from point cloud input to classification output. Each module adopts a standardized interface design, making it easy to integrate into existing precision agriculture systems. Attached Figure Description

[0030] Figure 1 This is a flowchart of the farmland point cloud ground separation method and system based on topological persistence analysis in Example 1.

[0031] Figure 2 This is a schematic diagram of the elevation field construction in Example 1; Figure 2 (a) in the figure represents the 3D point cloud projection. Figure 2 (b) in the text represents rasterization. Figure 2 (c) in the diagram represents the interpolation process.

[0032] Figure 3 This is a schematic diagram of the binarization separation effect of the two-dimensional elevation field under different threshold layers in Example 1, wherein the total number of threshold layers is 150; Figure 3 (a) in the figure represents the separation effect at the 5th threshold layer, with a corresponding elevation threshold of 0.629 m; Figure 3 (b) in the figure shows the separation effect at the 30th threshold layer, with a corresponding elevation threshold of 0.520 m; Figure 3 (c) in the figure represents the separation effect at the 70th threshold layer, with a corresponding elevation threshold of 0.380 m. Figure 3 In the figure, (d) represents the separation effect at the 100th threshold layer, with a corresponding elevation threshold of 0.250 m; Figure 3 In the figure, (e) represents the separation effect at the 130th threshold layer, with a corresponding elevation threshold of 0.120 m. Figure 3In the figure, (f) represents the separation effect at the 150th threshold layer, with a corresponding elevation threshold of 0.050 m.

[0033] Figure 4 This is a schematic diagram illustrating the evolution of the connected components between adjacent elevation layers in Example 1. Figure 4 In the diagram, (a) represents a new component, indicating that the connected component appears for the first time. Figure 4 In the diagram, (b) represents growth, indicating that the area of ​​the connected component increases. Figure 4 (c) in the text means merging, indicating that multiple connected components are merged into one; Figure 4 In the example, (d) means extinction, indicating that the connected component disappears or is absorbed by other connected components.

[0034] Figure 5 This is a schematic diagram of the persistent barcode in Example 1. Figure 5 The horizontal axis represents the elevation threshold layer, the vertical axis represents the component number, the marker point represents the birth time of the connected component, the bar length represents the persistence length of the connected component, and the area value next to the bar represents the component area of ​​the connected component.

[0035] Figure 6 This is a schematic diagram of the classification criteria in Example 1, illustrating the logical relationship of the three criteria.

[0036] Figure 7 The diagram shows the system hardware architecture in Example 1, illustrating the data acquisition, storage, and computing units.

[0037] Figure 8 The system software architecture diagram in Example 1 shows six functional units and their data interaction relationships. Detailed Implementation

[0038] 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.

[0039] Example 1 This embodiment provides a method for farmland point cloud ground separation based on topological persistence analysis. It introduces the mathematical concept of topological persistence and, by analyzing the topological structure characteristics of the elevation field, discovers and utilizes the essential differences in topological persistence between the ground and crops: ground components, due to their wide coverage and slow elevation changes, exhibit long persistence and large-area characteristics; crop components, due to their discrete distribution and abrupt elevation changes, exhibit short persistence and small-area characteristics. Based on this, the present invention proposes a robust ground separation method that does not require manually setting complex threshold parameters and can adaptively handle terrain undulation scenarios. (Refer to...) Figure 1 Specifically, it includes the following: S1. Acquire 3D point cloud data of the farmland environment, and preprocess and align the coordinates (coordinate system) of the data. In some embodiments, the acquired data object is a 3D point cloud map generated by a farmland SLAM system. This point cloud map mixes various targets such as ground, crops, and weeds, requiring effective semantic segmentation. Ground-crop separation, as the core task of point cloud semantic segmentation, is a key link in the precision agriculture data processing flow, directly affecting the accuracy and reliability of applications such as crop phenotypic parameter extraction, growth monitoring, and yield prediction.

[0040] Based on this, the original 3D point cloud data is improved in quality and its coordinates are standardized to lay the foundation for subsequent processing; specifically, this includes two sub-tasks: outlier filtering and coordinate system alignment.

[0041] In one specific embodiment, outlier filtering employs a statistical filtering method, identifying anomalies based on the statistical characteristics of local neighborhoods. For each point in the point cloud, the k nearest neighbors are calculated and formed into a set, and the average distance to these neighbors is statistically analyzed. The global mean and standard deviation of the average distances across all point clouds are calculated, and points deviating excessively from the global mean are marked as outliers and removed. This method effectively removes anomalies caused by sensor noise, multipath effects, or motion blur, improving point cloud quality. The number of neighborhood points k and the standard deviation factor are two parameters of this method; typically, k is set to 10-30, and the standard deviation factor is set to 1.5-3.0, adjusted according to the point cloud noise level.

[0042] In one specific embodiment, coordinate system alignment ensures that the Z-axis is vertically upward, providing a standardized coordinate system for constructing the elevation field. Principal component analysis is used to obtain three principal directions by calculating the eigenvalue decomposition of the point cloud covariance matrix; the eigenvector corresponding to the smallest eigenvalue is the ground normal vector. This normal vector is aligned to the positive Z-axis using a rotation matrix, completing the coordinate system alignment. For flat ground scenes, the RANSAC plane fitting algorithm can also be used to fit the ground plane and align the plane normal vector to the Z-axis. After coordinate system alignment, the XY plane is parallel to the ground, and the Z-axis represents the elevation direction, giving subsequent elevation field analysis a clear physical meaning.

[0043] S2. Project the preprocessed and coordinate-aligned 3D point cloud data onto a 2D plane and perform rasterization to generate a 2D elevation field. refer to Figure 2 In this embodiment, a 3D point cloud is projected onto a 2D plane and rasterized to generate an elevation field matrix. The elevation field is a 2D matrix, where each element represents the elevation value of the corresponding spatial location. This transformation reduces the 3D problem to 2D, preserving crucial elevation information while significantly reducing computational complexity, enabling the algorithm to efficiently process large-scale point cloud data.

[0044] In some embodiments, the choice of raster resolution r is a key parameter in the construction of the elevation field, directly affecting the accuracy and efficiency of subsequent analysis. Too large a raster resolution leads to the loss of spatial details, while too small a resolution increases the computational burden and makes the field more susceptible to noise interference.

[0045] In one specific embodiment, the present invention employs an adaptive selection strategy based on point cloud density to select the raster resolution r: the projection density of the point cloud onto the XY plane is calculated; a smaller raster resolution is selected for dense point clouds, and a larger raster resolution is selected for sparse point clouds, ensuring that each raster cell contains an appropriate number of points. The recommended resolution range is: For dense point clouds, the raster resolution is r = 0.03-0.08m; For medium-density point clouds, the raster resolution is r = 0.08-0.15m; For sparse point clouds, the raster resolution is r = 0.15-0.30m.

[0046] In some embodiments, different elevation aggregation strategies can be employed for each non-empty grid cell: the maximum value strategy selects the maximum elevation value of all points within the region, suitable for capturing the top of the crop canopy; the average value strategy calculates the average elevation of all points, suitable for smoothing terrain undulations; and the median value strategy selects the median elevation, which is more robust to outliers. This invention prefers the maximum value strategy because farmland scenarios focus on the surface features of the ground and the upper surface of the crops.

[0047] In some embodiments, for empty grids (areas not covered by point cloud), interpolation methods are used to complete them, avoiding holes from affecting subsequent topology analysis. An inverse distance-weighted interpolation method is used to search for neighboring points around the empty grid. Weights are calculated based on the inverse square of the distance, and a weighted average is used to obtain the interpolated elevation. The search radius is set to 2-5 times the raster resolution to ensure sufficient neighboring points participate in the interpolation. If the number of neighboring points within the search radius is too small, the grid is marked as invalid and ignored in subsequent processing.

[0048] S3. The two-dimensional elevation field is decomposed by stepwise threshold scanning, the connected components in the binarization result of each layer are extracted, and the evolution process of the connected components between adjacent elevation layers is tracked. refer to Figure 3 This embodiment performs topological decomposition on the elevation field through stepwise threshold scanning to reveal its topological structure and applies the theory of topological persistence to the analysis of the elevation field. The idea of ​​topological decomposition comes from Morse theory: a function can be understood by analyzing its critical points and level sets to understand its topological structure.

[0049] In some embodiments, a series of elevation thresholds are first set, uniformly distributed from the highest to the lowest elevation. The choice of the number of threshold layers *m* needs to be balanced between computational efficiency and analysis accuracy: too few layers will lead to loss of topological information, while too many layers will increase the computational burden and result in diminishing returns. Preferably, *m* is set to 50-300 layers, adjusted according to the elevation range and processing performance requirements. The threshold interval is equal to the elevation range divided by the number of layers.

[0050] Each elevation threshold layer is binarized: grids with elevations greater than or equal to the threshold are marked as 1 (foreground), and those below the threshold are marked as 0 (background). As the threshold decreases from high to low, the foreground region in the binary image gradually increases, reflecting the hierarchical structure of the elevation field. In the highest threshold layer, only a few highest points are foreground (crop canopy tops); in the middle threshold layer, the foreground includes part of the ground and crops; in the lowest threshold layer, almost the entire area is foreground (complete ground coverage).

[0051] In one specific embodiment, connected components of each layer of the binary image are extracted using 4-connectivity or 8-connectivity rules. A connected component refers to a set of foreground pixels that are interconnected in the image, and each connected component corresponds to a topological feature in the elevation field (such as a mountain peak, plateau, etc.). Breadth-first search or depth-first search algorithms are used to label connected regions, and each connected region is assigned a unique identifier.

[0052] Tracking the evolution of connected components between adjacent elevation layers is key to topology decomposition; as the threshold decreases, the reference... Figure 4 Connected components will undergo four types of evolution: Newborn (component appears for the first time); Growth (increased component area); Merge (integrate multiple components into one); Demise (the component disappears or is swallowed up by other components).

[0053] The inheritance relationship of components is established by calculating the overlapping pixels of adjacent layer components. When there are overlapping pixels between the k-th layer component and the (k+1)-th layer component, they are considered to belong to the same continuous component, exhibiting different behaviors at different elevation layers.

[0054] That is, when the number of overlapping pixels between two connected components in adjacent layers is greater than 0, it is assumed that there is an inheritance relationship between the two connected components in adjacent layers, and a decomposition matrix is ​​constructed to record the spatial distribution of each continuous connected component; connected component tracking can be efficiently implemented using the Union-Find data structure to maintain the equivalence relationship of connected component identifiers.

[0055] A decomposition matrix is ​​constructed to record the spatial distribution of each persistently connected component. The decomposition matrix is ​​a matrix of the same size as the elevation field, where each element records the number of times that location is covered by the corresponding component. The decomposition matrix reflects the cumulative spatial distribution of the connected components and is used for subsequent calculation of component areas. After topological decomposition, a series of persistently connected components are obtained, each corresponding to a topological feature in the elevation field, preparing for the next step of persistence analysis.

[0056] S4. Calculate the topological persistence features of each connected component and generate a persistent barcode as the topological feature descriptor of the component. Topological persistence features include the birth time, persistence length, and component area of ​​connected components; the birth time, persistence length, and component area of ​​connected components are combined into a persistent barcode, which serves as the topological feature descriptor of the connected components; Among them, birth time refers to the elevation threshold at which the component first appears. Birth time reflects the elevation position of the component; ground-based components are usually born at lower elevations, while crop components are born at higher elevations.

[0057] Persistence length refers to the threshold number of layers in which a component persists. Persistence length reflects the continuity of a component along the elevation direction and is the most important characteristic in persistence theory. Ground-based components, due to their wide coverage and slow elevation changes, extend from low to high elevation layers, resulting in a long persistence length. Crop-based components, on the other hand, are discretely distributed and experience abrupt elevation changes, existing only in a few elevation layers, thus exhibiting a short persistence length. This difference in persistence length is the theoretical basis for the ground-crop separation achieved in this invention.

[0058] Component size (size): refers to the total number of different grid cells covered by a component. Component size reflects the spatial scale of the component; ground components cover a large area, while crop components have a smaller area. Component size and persistence length together characterize the topological importance of a component. Components with large area and long persistence correspond to primary structures in the scene (such as the ground), while components with small area and short persistence correspond to secondary structures or noise (such as crops and weeds).

[0059] refer to Figure 5This invention combines three features—birth time, persistence length, and component area—into a persistent barcode, which serves as a topological feature descriptor for the component. The persistent barcode comprehensively characterizes the component's topological persistence features: birth time describes where the component originated, persistence length describes how long the component has existed, and component area describes its size. In topological data analysis theory, persistent barcodes are a visual representation of persistent homology; each bar corresponds to a topological feature, and the length of the bar represents the persistence of that feature. In this invention, persistent barcodes are used to distinguish between two types of targets: ground and crops; long bars correspond to ground, and short bars correspond to crops.

[0060] S5. Based on topological persistence features and spatial features of connected components, classify each connected component into ground and crops, and convert topological features into semantic labels. In some embodiments, reference Figure 6 A triple criterion is used to classify connected components. Connected components that meet all three conditions are marked as ground, otherwise they are marked as crops or noise.

[0061] The first criterion is the persistence length criterion, which filters out components with long persistence; The second criterion is the component area criterion, which filters out components with larger areas; The third criterion is the birth height criterion, which filters out low-elevation components.

[0062] Because ground components cover the entire scene, they are characterized by long persistence, large area, and low birth height; crop components are discretely distributed, with short persistence, small area, and high birth height. The combined use of the three criteria improves the robustness of classification and avoids misclassification that may be caused by a single criterion.

[0063] In some embodiments, the adaptive threshold calculation of the triple criterion is as follows: The persistence threshold is calculated using an adaptive method, eliminating the need for manual setting. Specifically, the method involves calculating the set of persistence lengths for all connected components, and then calculating the percentile of this set as the persistence threshold. Percentiles are used instead of fixed multiples because persistence distributions can vary significantly across different scenarios, and percentiles can adaptively select an appropriate threshold based on the data distribution. A threshold of 60-80 percentiles is recommended, with 70 percentiles being a typical value.

[0064] The area threshold is set as a certain percentage of the total grid area. This threshold is based on prior knowledge of farmland scenarios: the ground typically covers most of the area, while individual crops or weeds occupy a very small area. The recommended area threshold is 0.5-10% of the total grid area, with a typical value of 2%.

[0065] The birth height threshold is determined using either the median or the mean elevation. The median is used because it is insensitive to outliers and can more robustly distinguish between low-lying ground and high-lying crops.

[0066] The triple criterion in this embodiment is not a simple feature superposition, but achieves high-precision classification through the synergistic effect of complementary features. Its synergistic mechanism is reflected in the following three levels: First layer: complementary filtering; The persistence length criterion identifies ground features from a topological continuity perspective, filtering out continuous components spanning multiple elevation layers, corresponding to the vertical continuity of the ground. The component area criterion identifies ground features from a spatial scale perspective, filtering out components covering a large area, corresponding to the horizontal extension of the ground. The birth height criterion identifies ground features from an elevation location perspective, excluding high-altitude non-ground structures. These three criteria correspond to three different dimensions: topology, space, and elevation, and are independent yet complementary.

[0067] Second layer, enhanced robustness; Single criteria are prone to misclassification: using only the persistence criterion, tall corn plants may be misclassified as ground due to their long persistence; using only the area criterion, large canopies formed by densely planted crops may be misclassified as ground; using only the height criterion, low weeds or mounds of soil may be misclassified as ground. The AND logic (simultaneous fulfillment) of the triple criterion significantly reduces the misclassification rate: a component must simultaneously possess long persistence (topological continuity), large area (spatial extension), and low birth height (located at the bottom layer) to be identified as ground. The intersection of these three conditions greatly reduces the misclassification space. Experimental data shows that compared to a single criterion, the triple criterion improves accuracy by 8-13 percentage points.

[0068] Third layer, boundary case handling; In the transition zone between the ground and crops, there are some boundary case components, such as mixed components partially covered by the ground and crops, and local protrusions or depressions in the ground. The triple criterion effectively handles these boundary cases through synergy: for mixed components, if their main part is the ground, they will satisfy the area and persistence criteria but may not satisfy the height criterion (because the crop portion raises the birth height), thus being correctly classified as crops; for local ground protrusions, although they may satisfy the height criterion, they do not satisfy the area criterion due to their limited spatial extent, thus avoiding misclassification. This multi-dimensional cross-validation mechanism makes the classification boundaries clearer and more accurate.

[0069] In particular, the synergistic effect of the three criteria is especially evident in complex scenarios. In farmland with undulating terrain, a single criterion struggles to simultaneously handle variations in elevation and the discrete distribution of crops. The three criteria, however, capture the continuity of the terrain (even with elevation variations) through topological persistence, the integrity of the terrain through area, and eliminate high-level interference through height. Working together, these three criteria achieve accurate classification of complex scenarios. This synergistic mechanism is not a simple feature fusion, but rather based on a profound understanding of the essential differences between the terrain and crops in the three-dimensional feature space of topology, space, and elevation.

[0070] S6. Based on the classification results of connected components, perform point cloud level annotation on the original 3D point cloud data and output the classified ground point cloud and crop point cloud.

[0071] The non-ground components are further processed to filter noise, label point clouds, and output the final result.

[0072] Specifically, micro-component filtering: Components with excessively small areas among non-ground components are marked as noise and removed. This threshold is set based on the following considerations: interference such as weeds and rocks appear as small, isolated components in the elevation field; by setting a lower area limit, these interferences can be effectively removed. A recommended filtering threshold is 0.05-0.5% of the total grid area.

[0073] Point cloud level annotation: Mapping the raster-level classification results back to the original 3D point cloud, assigning a semantic label to each point. For each point in the original point cloud, calculate its projected raster index, look up the classification label of that raster, and assign that label to the point. The annotated point cloud data contains 3D coordinates and semantic category information.

[0074] Output Results: The final output consists of point cloud data containing classification labels, which can be further separated into ground point cloud subsets and crop point cloud subsets. These two subsets provide high-quality foundational data for subsequent precision agriculture applications. For example, digital ground models can be built based on ground point clouds for crop height measurement, while crop point clouds can be used for crop phenotypic parameter extraction, growth monitoring, and yield prediction.

[0075] In some embodiments, the present invention provides a farmland point cloud ground separation system, with reference to Figure 7 and Figure 8 It includes the hardware layer, software layer, and interface layer; The hardware layer includes: a data acquisition unit for acquiring 3D point cloud data from sensors such as LiDAR and depth cameras; a data storage unit for storing raw point clouds, intermediate results, and final outputs; and a computing and processing unit, including a CPU for serial logic processing and a GPU for parallel computing such as elevation field rasterization and connected component extraction.

[0076] The software layer comprises six functional units: The point cloud preprocessing unit is used to acquire three-dimensional point cloud data of the farmland environment and to preprocess and align the coordinates of the data. The elevation field construction unit is used to project preprocessed and coordinate-aligned 3D point cloud data onto a 2D plane and perform rasterization processing to generate a 2D elevation field. The topology decomposition unit is used to perform topology decomposition on the two-dimensional elevation field using a stepwise threshold scan, extract connected components in the binarization result of each layer, and track the evolution of connected components between adjacent elevation layers. The persistence analysis unit is used to calculate the topological persistence characteristics of each connected component; Ground separation unit is used to classify ground and crops in each connected component based on topological persistence features and spatial features of connected components; The crop extraction unit is used to perform point cloud-level annotation on the original 3D point cloud data based on the classification results of connected components, and output the classified ground point cloud and crop point cloud.

[0077] The interface layer includes: a data input interface that supports point cloud formats such as PCD, PLY, LAS, and ROS; a parameter configuration interface that supports adjusting parameters such as raster resolution and threshold layer number; and a result output interface that supports multiple output formats such as classified point clouds and visualized images.

[0078] The units interact via a high-speed data bus, supporting pipelined parallel processing. The system can process 5 million point cloud data points within 10 seconds on a standard workstation, meeting near real-time requirements, and can be integrated into a SLAM real-time mapping system to achieve online ground separation.

[0079] This invention is the first to apply topological persistence analysis to the specific task of separating the ground and crops from farmland point clouds. The method is designed specifically for the characteristics of farmland scenarios (undulating ground and discrete crop distribution), proposing a combined approach using persistence, area, and birth height as triple criteria. Existing topological methods mainly use a single persistence feature; this invention significantly improves classification accuracy and robustness through multi-feature fusion. Simultaneously, an adaptive threshold calculation method based on percentiles is proposed, eliminating the need for manually setting complex parameters. Existing methods typically use fixed thresholds or require extensive parameter tuning; this invention's adaptive method improves the method's practicality. Converting the 3D point cloud into a 2D elevation field for topological analysis preserves crucial elevation information while significantly reducing computational complexity, enabling the method to handle large-scale point cloud data.

[0080] Example 2 For cornfield ground-crop separation, it includes the following: I. Dataset Description This embodiment uses SLAM point cloud data from a cornfield for experimental verification. This dataset was collected by a ground robot equipped with a 16-line mechanical rotating LiDAR at a farm in Hebei Province in July 2024 (corn jointing stage). The specific parameters of the point cloud data are as follows: scene size 200m long × 150m wide, total area 30,000 square meters; total number of points in the point cloud approximately 5 million; point cloud density 167 points / m²; crop row spacing 0.75m; crop height 1.2-1.8m, average 1.5m; ground has furrow structures with a furrow spacing of 0.75m and a furrow depth of approximately 0.15m.

[0081] II. Detailed Processing Steps Step 1: Point cloud preprocessing; A statistical filter was used with a neighborhood number of 20 and a standard deviation factor of σ = 2.0. Approximately 40,000 outliers were removed, accounting for 0.8% of the total number of points. The PCA algorithm was used for coordinate system alignment.

[0082] Step 2: Elevation field construction; Point cloud density is 167 points / m², and raster resolution r=0.1m is selected; the XY plane is divided into 3 million grid cells; elevation range: minimum 0.12m, maximum 1.98m, median 0.85m.

[0083] Step 3, Topological decomposition; The elevation range was 1.86m, the threshold layer number was set to m=150, and the threshold interval was 1.24cm; connected components were extracted using the 8-connectivity rule; 68 persistent components were identified, including 3 ground components and 65 crop / noise components.

[0084] Step 4: Persistence analysis; Ground component J1: birth time 0.28m, persistence length 142 layers, area 1.86 million grids (62%); Crop component J5: birth time 1.65m, persistence length 48 layers, area 8500 grids (0.28%).

[0085] Step 5: Ground separation; The persistence threshold is 40 layers (70th percentile), the area threshold is 60,000 grids (2% of the total area), and the birth height threshold is 0.85m; 3 ground components and 65 crop components are identified.

[0086] Step 6: Crop extraction; With a filtering threshold of 3000 grids, 18 noise components were filtered out, and 47 crop components were retained. Statistical results: 2.42 million ground points (48.8%) and 2.54 million crop points (51.2%).

[0087] III. Performance Evaluation By manually labeling 500 random sampling points as true values, the classification performance indicators were calculated as follows: precision 91.2%, recall 87.9%, F1-score 89.5%, and processing time 9.8 seconds, which meets the near real-time requirements.

[0088] Example 3 For wheat field ground-crop separation, it specifically includes the following: I. Dataset Description This embodiment uses SLAM point cloud data from a wheat field. The dataset was collected in April 2024 (wheat heading stage) at a farm in Shandong Province. The scene is 150m long × 100m wide, with a total area of ​​15,000 square meters; the total number of points in the point cloud is approximately 3.2 million; the point cloud density is 213 points / m²; the crop row spacing is 0.25m (dense planting); the crop height is 0.6-0.9m, with an average of 0.75m; the ground is relatively flat with no obvious furrow structure.

[0089] II. Processing Results The same processing steps as in Example 2 were used, but the parameters were adjusted according to the characteristics of the scene: raster resolution r = 0.08m (high-density point cloud); threshold layer number m = 120; persistence threshold 35 layers (70th percentile); area threshold 45,000 grids (2% of total area).

[0090] Classification results: 2 ground components and 89 crop components were identified; Statistics: 1.72 million ground-based sampling points (53.8%), and 1.48 million crop-based sampling points (46.2%). Performance metrics: Precision 88.7%, Recall 90.3%, F1-score 89.5%, Processing time 6.2 seconds.

[0091] This embodiment verifies the adaptability of the method to different crop types (wheat vs. maize), different planting densities (dense planting vs. medium row spacing), and different terrain conditions (flat vs. undulating). Despite the significantly different scene characteristics, the method maintains stable high accuracy, demonstrating the effectiveness of adaptive threshold calculation.

[0092] Example 4 For the separation of early growth stages in soybean fields, the specific content includes the following: I. Dataset Description This example addresses a challenging scenario during the early growth stage of crops (when plant height is relatively low). The dataset was collected from a farm in Heilongjiang Province in June 2024 (soybean seedling stage). The scenario measures 180m long × 120m wide, with a total area of ​​21,600 square meters; the point cloud contains approximately 4.1 million points; the point cloud density is 190 points / m²; the crop row spacing is 0.60m; the crop plant height is only 0.15-0.30m, with an average of 0.22m (early growth); the ground has a ridge-furrow structure with a ridge height of 0.10-0.15m.

[0093] II. Challenges and Solutions The main challenge in the early growth stage is that the crop height is close to the ground furrow height, which can easily lead to confusion. This method effectively addresses this issue through topological persistence analysis: although the crop and furrow elevations are similar, they have fundamental differences in topological persistence—furrows, as part of the ground, exhibit a continuous distribution in the elevation field and have long persistence; crops, on the other hand, are discretely distributed and have short persistence.

[0094] III. Processing Results Parameter settings: raster resolution r=0.10m; threshold number of layers m=80 (small elevation range); persistence threshold 28 layers (70th percentile); area threshold 65,000 grids (3% of total area, appropriately relaxed to avoid misclassifying furrows as crops).

[0095] Classification results: 4 ground components (including furrow areas) and 132 crop components were identified.

[0096] Statistics: 2.45 million ground-based points (59.8%), 1.65 million crop-based points (40.2%).

[0097] Performance metrics: Precision 87.3%, Recall 85.1%, F1-score 86.2%, Processing time 7.5 seconds.

[0098] This embodiment verifies the effectiveness of the method in challenging scenarios (low crop height, close to ground elevation). Although the F1-score is slightly lower than that of the cornfield example, it is still significantly better than traditional methods (the fixed threshold method has an F1 score of only 58.3% in this scenario, making it difficult to distinguish between low-lying crops and undulating ground).

[0099] Ablation experiment: Validation of the effectiveness of the triple criterion; To verify the necessity of using the triple criteria in combination, an ablation experiment was conducted on the cornfield data of Example 2 to compare the performance of different criterion combinations.

[0100] Experiment 1: Using only the persistence length criterion; Results: Precision 78.5%, Recall 93.2%, F1-score 85.2%; Analysis: The recall rate was high but the precision rate was low because some large crop components (such as tall corn plants) also have a long persistence and were misclassified as ground.

[0101] Experiment 2: Using only the component area criterion; Results: Precision 82.1%, Recall 81.7%, F1-score 81.9%; Analysis: The performance is moderate because there is some overlap between the area distribution of the ground and the crop, making it difficult to completely distinguish them using a single area criterion.

[0102] Experiment 3: Using a dual criterion of persistence and area; Results: Precision 87.6%, Recall 85.3%, F1-score 86.4%; Analysis: Performance has been significantly improved, but some elevated ground areas (such as mounds) are still misidentified as crops.

[0103] Experiment 4: Using a triple criterion of persistence, area, and birth height (complete method); Results: Precision 91.2%, Recall 87.9%, F1-score 89.5%; Analysis: The best performance was achieved. The birth height criterion effectively excluded high ground areas, and the combined use of the three criteria significantly improved the accuracy and robustness of the classification.

[0104] Ablation experiment conclusions: Compared with single or dual criteria, the combined use of triple criteria improved the F1-score by 3.1-7.6 percentage points, proving the necessity and effectiveness of multi-feature fusion.

[0105] Parameter sensitivity analysis; To verify the effectiveness of the adaptive threshold calculation method, the performance at different percentiles was tested on the cornfield data of Example 2.

[0106] Percentile experiments on persistence thresholds: Test the 50th, 60th, 70th, 80th, and 90th percentiles; The results showed that the F1-score was 85.7% at the 50th percentile (the threshold was too low, and some crops were misidentified as ground); 88.3% at the 60th percentile; 89.5% at the 70th percentile (optimal); 88.9% at the 80th percentile; and 86.2% at the 90th percentile (the threshold was too high, and some ground edges were misidentified as crops).

[0107] Area threshold ratio experiment: testing 0.5%, 1%, 2%, 5%, 10%; The results showed that the F1-score was 87.1% at 0.5%; 88.7% at 1%; 89.5% (optimal) at 2%; 88.2% at 5%; and 85.9% at 10%.

[0108] Parameter sensitivity analysis results: The method of this invention maintains stable high performance within the recommended parameter range (60th-80th percentile of durability, 1-5% of area), demonstrating the robustness of the method. The recommended typical value (70th percentile, 2% of area) achieved optimal or near-optimal performance in multiple embodiments.

[0109] Comparative experiment: compared with existing methods; On the datasets of Examples 2, 3, and 4, the method of the present invention is systematically compared with existing mainstream methods.

[0110] Method 1: Fixed threshold method (threshold set as median elevation); F1=70.3% for cornfield, F1=72.1% for wheatfield, F1=58.3% for soybeanfield; average F1=66.9%.

[0111] Method 2: RANSAC plane fitting method (1000 iterations); F1=79.9% for corn field, F1=84.2% for wheat field, and F1=71.5% for soybean field; average F1=78.5%, with a relatively long processing time (32 seconds for corn field).

[0112] Method 3: Fabric simulation method (CSF, using default parameters); F1=75.6% for corn field, F1=79.8% for wheat field, and F1=68.9% for soybean field; average F1=74.8%; processing time is relatively long (45 seconds for corn field).

[0113] Method 4: The method of the present invention; F1=89.5% for corn field, F1=89.5% for wheat field, and F1=86.2% for soybean field; average F1=88.4%; processing time: 9.8 seconds for corn field, 6.2 seconds for wheat field, and 7.5 seconds for soybean field.

[0114] The comparison shows that the method of this invention significantly outperforms existing methods in separation accuracy, with an average F1-score improvement of 9.9 percentage points compared to the best existing method (RANSAC) and 21.5 percentage points compared to the fixed threshold method. Simultaneously, the processing speed of this method is 3-5 times that of RANSAC and 4-7 times that of CSF, meeting near real-time requirements. More importantly, the method of this invention exhibits stable performance under different scenarios, with a standard deviation of only 1.9%, while the standard deviation of the fixed threshold method reaches 7.1%, demonstrating the robustness of this method.

[0115] 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 farmland point cloud ground separation based on topological persistence analysis, characterized in that, Includes the following steps: S1. Acquire 3D point cloud data of the farmland environment and preprocess and align its coordinates. S2. Project the preprocessed and coordinate-aligned 3D point cloud data onto a 2D plane and perform rasterization to generate a 2D elevation field. S3. The two-dimensional elevation field is decomposed by stepwise threshold scanning, the connected components in the binarization result of each layer are extracted, and the evolution process of the connected components between adjacent elevation layers is tracked. S4. Calculate the topological persistence characteristics of each connected component; S5. Based on the topological persistence features and the spatial features of the connected components, classify each connected component into ground and crops; S6. Based on the classification results of the connected components, perform point cloud level annotation on the original three-dimensional point cloud data, and output the classified ground point cloud and crop point cloud. In step S2, during the generation of the two-dimensional elevation field, an adaptive selection strategy based on point cloud density is used to determine the raster resolution, specifically as follows: Calculate the projection density of the point cloud onto a two-dimensional plane, take the elevation feature value of all points in each grid cell, and use interpolation to complete empty grids; The choice of raster resolution r is as follows: For dense point clouds, the raster resolution is r = 0.03-0.08m; For medium-density point clouds, the raster resolution is r = 0.08-0.15m; For sparse point clouds, the raster resolution is r = 0.15-0.30m; In S3, the two-dimensional elevation field is decomposed topologically by step-by-step threshold scanning, including setting 50-300 thresholds from high elevation to low elevation for binarization processing to generate corresponding binary image sequences, and using 4-connectivity or 8-connectivity rules to extract connected components of each layer of binary image. By calculating the evolution process of overlapping pixels of connected components in adjacent layers, when the number of overlapping pixels of two connected components in adjacent layers is greater than 0, it is assumed that there is an inheritance relationship between the two connected components in adjacent layers, and a decomposition matrix is ​​constructed to record the spatial distribution of each continuous connected component. In step S4, the topological persistence feature includes the birth time, persistence length, and component area of ​​the connected component; the birth time, persistence length, and component area of ​​the connected component are combined into a persistent barcode, which serves as the topological feature descriptor of the connected component. Wherein, the birth time is the elevation threshold at the time when the connected component first appears; The persistence length is the threshold number of layers that the connected components continue to exist. The component area is the total number of grid cells covered by the connected components; In S5, a triple joint determination is adopted using the persistence length criterion, the component area criterion, and the birth height criterion. Connected components that simultaneously satisfy all three criteria are marked as ground, and the remaining connected components are marked as crops or noise. In step S6, point cloud level annotation is performed on the original three-dimensional point cloud data based on the classification results, including: marking non-ground connected components as crop candidate regions, filtering out tiny components with a total grid area of ​​0.05-0.5% to remove noise interference, and then projecting each point in the original three-dimensional point cloud onto the corresponding grid cell and querying its classification label to complete the point cloud level annotation.

2. The method for farmland point cloud ground separation based on topological persistence analysis according to claim 1, characterized in that, In step S1, the preprocessing involves filtering out outliers from the 3D point cloud data. This is done by using a statistical filter to remove noise points. The number of neighborhood points in the statistical filter is 10-30, and the standard deviation factor is 1.5-3.

0. The coordinate alignment adopts the principal component analysis method. The ground principal direction is obtained by calculating the eigenvalue decomposition of the three-dimensional point cloud covariance matrix. The ground principal direction is rotated and aligned to the positive Z-axis direction, thereby completing the coordinate alignment.

3. The method for farmland point cloud ground separation based on topological persistence analysis according to claim 1, characterized in that, The evolution process of the connected components between adjacent elevation layers includes: Newborn, the first appearance of the connected component; Growth leads to an increase in the area of ​​interconnected components; Merge, where multiple connected components are combined into one; Demise: The connected component disappears or is absorbed by other connected components.

4. The method for farmland point cloud ground separation based on topological persistence analysis according to claim 1, characterized in that, The threshold for the persistence length criterion is the 60-80 percentile of the persistence length distribution of all connected components, used to filter connected components that span multiple elevation layers. The threshold for the component area criterion is 0.5-10% of the total grid area, used to filter connected components that cover a large area; The threshold for the birth height criterion is the median or mean elevation, used to filter connected components with low elevations.

5. A farmland point cloud ground separation system based on topological persistence analysis that implements the method of any one of claims 1-4, characterized in that, It includes a hardware layer, a software layer, and an interface layer, which interact with each other via a data bus and support pipelined parallel processing; The hardware layer includes a data acquisition unit, a data storage unit, and a CPU+GPU heterogeneous computing processing unit. The software layer includes a point cloud preprocessing unit, an elevation field construction unit, a topology decomposition unit, a persistence analysis unit, and a ground crop classification unit, which are linked in sequence. The interface layer includes a data input interface, a parameter configuration interface, and a result output interface.

Images