A single-tree segmentation radar point cloud denoising method

Abstract

The application provides a single-tree segmentation radar point cloud noise reduction method, and belongs to the technical field of laser radar data processing and forest resource monitoring. The method is based on an unsupervised deep learning architecture, and realizes high-precision point cloud noise reduction through a double noise processing mechanism and local and global feature extraction modules. A data set of 133 forest lands in Guangdong Province is used for verification, and traditional methods such as morphological filtering, progressive triangulation network filtering, statistical filtering and radius filtering are compared. The experimental results show that the method is superior to the traditional algorithms in the four indexes of MSE, SNR, Hausdorff distance and SSIM, can significantly improve the noise reduction effect, clearly retain the key features such as tree height, diameter at breast height and the like, and has high consistency and strong robustness on different sample point cloud data.

Description

Technical Field

[0001] This invention relates to the field of lidar data processing and forest resource monitoring technology, and particularly to a method for noise reduction of radar point clouds segmented from a single tree. Background Technology

[0002] Forests are a vital component of Earth's ecosystems, providing a wide range of ecosystem services, such as biodiversity, carbon storage, and cultural value. First, it is estimated that forest ecosystems possess more than half of the world's biodiversity and include species of immense conservation value. Second, exploring the carbon sequestration capacity of forest ecosystems is crucial for achieving carbon neutrality, as they absorb nearly one-third of annual anthropogenic carbon dioxide emissions. Furthermore, in addition to providing food, maintaining soil, and ensuring climate stability, forest ecosystems can influence human physiology and psychology through individual participation, bringing psychological and cultural benefits to individuals and society. In the future development of humankind, forest ecosystems will play an extremely important role, making research into forest physiology and ecology, as well as effective forest management, indispensable.

[0003] LiDAR plays a significant role in forest physiological and ecological research, providing a wealth of data for forest surveys and management. Airborne LiDAR can accurately extract the vertical structure of the canopy, obtaining forest parameters at the stand and individual tree scales. Airborne LiDAR can control scanning errors to the centimeter or even millimeter level, offering significant advantages in digital ecosystem construction, aboveground biomass retrieval, and long-term environmental monitoring of forest stands. Using airborne LiDAR to acquire forest ecological data to strengthen forest ecological construction has become a global consensus among forestry scholars. Currently, image segmentation based on the Canopy Height Model (CHM) and point cloud segmentation based on normalized point cloud spatial clustering are two typical methods for acquiring forest ecological data using UAV airborne radar. Point cloud segmentation based on UAV airborne radar data is crucial for solving problems such as canopy height measurement, individual tree segmentation and identification, leaf area index and canopy closure estimation, and biomass calculation.

[0004] However, due to the limitations of scanning equipment and the influence of the acquisition environment, point cloud data inevitably contains many noise points and outliers. These noise points reduce data accuracy and severely affect the effectiveness of subsequent data processing and use. To reduce noise in point cloud data and obtain clean point clouds, many scholars have conducted in-depth research and discussions. Currently, the most widely used and technically mature denoising technique is filtering. Choudhury et al. proposed a trilateral filtering method based on bilateral filtering for denoising three-dimensional mesh models. Sun et al. limited the filtering window to the approximate normal vector region for normal vector filtering. Liang Xinhe et al., based on the SUN algorithm, obtained an improved discrete point model filtering method by filtering both the normal vector of the discrete point model and the position of the points. Lu Dongdong et al. explored the application of statistical filtering algorithms and radius filtering algorithms in practical denoising work and compared and analyzed the advantages and disadvantages of these two typical filters in terms of denoising performance. Filtering methods are also widely used in the field of LiDAR single-tree segmentation. For example, Guo Yushan et al. used morphological filtering methods to ensure better preservation of contour information when studying single-tree crown extraction; Wu Xiaokang compared multi-level surface filtering, slope filtering, and CSF filtering algorithms, and the results showed that CSF filtering was more accurate than slope filtering and multi-level surface filtering methods in the field of lidar single-tree segmentation. In addition, Ma Kaisen, Li Wei, and others used an improved progressive triangular network filtering method to process radar point cloud data. However, traditional algorithms have certain limitations in terms of adaptability and noise reduction, especially when processing different types of data, where the effects vary significantly.

[0005] With the development of various technologies, point cloud denoising based on deep learning and machine learning has gradually attracted the attention of many scholars. ChaojingDuan et al. proposed a neural network-based 3D point cloud denoising architecture that can accurately estimate and remove noise in point clouds. Zhang Jie et al. proposed a density prior-guided unsupervised deep point cloud denoising algorithm, which improves the performance of unsupervised deep point cloud denoising algorithms. ShitongLuo et al. proposed a neural network architecture for estimating correlation scores using only noisy point clouds as input, and developed a denoising algorithm based on the estimated scores.

[0006] Currently, point cloud denoising technology is still in a stage of parallel development between deep learning methods and traditional methods. However, compared with traditional algorithms, deep learning methods, through the use of advanced architectures such as Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs), can automatically learn and extract complex features of point clouds, significantly reducing the workload of manual feature design. Through training on large amounts of data, deep learning models can adapt well to different types of noise and variations in point cloud data. For example, models can learn specific noise patterns from the training set, thus performing denoising more effectively. Furthermore, deep learning methods perform particularly well in handling complex scenes and diverse point cloud data, especially in complex situations such as non-uniform distribution, high-dimensional data, and multi-view point cloud synthesis. These methods not only improve denoising performance but also preserve more point cloud details and structural information. However, in the field of LiDAR single-tree segmentation, traditional algorithms still dominate, and deep learning methods have not completely replaced traditional methods when dealing with more detailed and complex data requirements. This indicates that there is still room for further optimization and development of deep learning methods in single-tree segmentation tasks. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] This paper aims to develop an unsupervised deep learning algorithm for LiDAR single-tree segmentation, based on existing deep learning algorithms. Specific objectives include: collecting and preprocessing LiDAR point cloud data for single-tree segmentation; selecting a suitable unsupervised learning algorithm and improving and optimizing its design; training and optimizing the model to enhance denoising performance; conducting detailed quantitative and qualitative analysis of the algorithm's denoising results; and, through optimizing the denoising effect, laying the foundation for future data processing and applications, enabling more effective utilization of the data in subsequent analysis and research.

[0009] (II) Technical Solution

[0010] A method for denoising radar point clouds segmented from a single tree includes the following steps:

[0011] Acquire forest point cloud data collected by airborne lidar, wherein the point cloud data includes individual tree structure information;

[0012] The point cloud data is preprocessed, including point cloud matching, coordinate normalization, and sample segmentation;

[0013] An unsupervised deep learning noise reduction network is constructed, the network including an input layer, a feature extraction module, a displacement vector prediction module, and an output layer;

[0014] Add dual independent noise to the point cloud data to form a point cloud with dual noise as network input;

[0015] The feature extraction module extracts local geometric features and global structural features of the point cloud.

[0016] The displacement vector prediction module predicts a displacement vector for each point and corrects the position of the point cloud based on the displacement vector.

[0017] Output denoised point cloud data for single tree segmentation, tree height extraction, or diameter at breast height estimation.

[0018] Preferably, the feature extraction module includes a local feature extraction submodule and a global feature aggregation submodule;

[0019] The local feature extraction submodule uses the k-nearest neighbor algorithm to construct a local map of the point cloud and extracts neighborhood features through a graph convolutional network.

[0020] The global feature aggregation submodule aggregates local features using max pooling or average pooling to form a global descriptor.

[0021] Preferably, the local feature extraction submodule adopts the DGCNN architecture, and the global feature aggregation submodule draws on the symmetric function structure of PointNet.

[0022] Preferably, the displacement vector prediction module includes a multilayer perceptron (MLP) network and a residual connection structure;

[0023] The MLP network performs a nonlinear mapping on the feature vectors and outputs a three-dimensional displacement vector for each point.

[0024] The residual connection adds the original point coordinates to the predicted displacement vector to obtain the corrected point coordinates.

[0025] Preferably, the method for adding the dual independent noise includes:

[0026] First, generate a first noise that is identical to but independent of the original noise distribution, and add it to the point cloud to obtain the first noise point cloud;

[0027] A second noise, independent and identically distributed to the first noise, is generated and added to the first noise cloud to obtain a second noise cloud as network input.

[0028] Preferably, the noise intensity is experimentally set to 0.1, and the noise distribution is either Gaussian or uniform.

[0029] Preferably, the loss function used during network training is:

[0030]

[0031] in, For mean square error loss, For repulsive force loss, It is the influencing factor for controlling repulsive force loss, and is set to 0.0005.

[0032] Preferably, the repulsive force loss is achieved by constructing a pseudo-clean point cloud and calculating the maximum distance between points, in order to avoid excessive aggregation of the point cloud.

[0033] Preferably, the method further includes a network hyperparameter setting step:

[0034] The learning rate was adjusted to 0.001 using a cosine annealing algorithm;

[0035] The network structure is an encoder-decoder type, with a layer structure of 3-256-512-1024-2048;

[0036] The training rounds consist of 200 rounds, and the batch size is 128.

[0037] Preferably, the method further performs visual and quantitative evaluations after noise reduction, and the evaluation indicators include at least one of the following: mean square error, signal-to-noise ratio, Hausdorff distance, and structural similarity index.

[0038] A point cloud denoising system for performing any of the methods described above, comprising:

[0039] The data acquisition module is used to acquire the raw lidar point cloud;

[0040] The preprocessing module is used for data cleaning, normalization, and noise addition;

[0041] A noise reduction network module is used to perform unsupervised deep learning noise reduction.

[0042] The post-processing module is used to output and evaluate the noise reduction results.

[0043] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.

[0044] An electronic device includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method as described in any of the preceding claims.

[0045] The beneficial effects of this invention are:

[0046] 1. The present invention has high noise reduction accuracy, and its four indicators, MSE, SNR, Hausdorff distance, and SSIM, are all superior to those of traditional methods;

[0047] 2. This invention can effectively preserve key forest tree structural features such as tree height and diameter at breast height;

[0048] 3. This invention requires no data annotation, is highly adaptable, and performs consistently across different forest land data.

[0049] 4. The network structure of this invention is lightweight, training is stable, and it is suitable for practical engineering applications. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this disclosure, the accompanying drawings used in some embodiments of this disclosure will be briefly described below. Obviously, the drawings described below are only drawings of some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings. In addition, the drawings described below can be regarded as schematic diagrams and are not intended to limit the actual size of the product, the actual flow of the method, the actual timing of the signals, etc. involved in the embodiments of this disclosure.

[0051] Figure 1 Schematic diagram of the sample plot;

[0052] Figure 2 Sampling comparison before and after noise reduction on S10;

[0053] Figure 3 S5 point cloud frontal view comparison;

[0054] Figure 4 S10 point cloud frontal view comparison;

[0055] Figure 5 S5 point cloud overhead view comparison;

[0056] Figure 6 Comparison of S10 point cloud overhead views. Detailed Implementation

[0057] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0058] In the description of the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0059] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0060] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] The present invention will be further described in detail below through embodiments.

[0063] Example

[0064] This embodiment provides a method for noise reduction of radar point clouds segmented from a single tree, including the following steps:

[0065] Acquire forest point cloud data collected by airborne lidar, wherein the point cloud data includes individual tree structure information;

[0066] The point cloud data is preprocessed, including point cloud matching, coordinate normalization, and sample segmentation;

[0067] An unsupervised deep learning noise reduction network is constructed, the network including an input layer, a feature extraction module, a displacement vector prediction module, and an output layer;

[0068] Add dual independent noise to the point cloud data to form a point cloud with dual noise as network input;

[0069] The feature extraction module extracts local geometric features and global structural features of the point cloud.

[0070] The displacement vector prediction module predicts a displacement vector for each point and corrects the position of the point cloud based on the displacement vector.

[0071] Output denoised point cloud data for single tree segmentation, tree height extraction, or diameter at breast height estimation.

[0072] In this embodiment, the feature extraction module includes a local feature extraction submodule and a global feature aggregation submodule;

[0073] The local feature extraction submodule uses the k-nearest neighbor algorithm to construct a local map of the point cloud and extracts neighborhood features through a graph convolutional network.

[0074] The global feature aggregation submodule aggregates local features using max pooling or average pooling to form a global descriptor.

[0075] In this embodiment, the local feature extraction submodule adopts the DGCNN architecture, and the global feature aggregation submodule draws on the symmetric function structure of PointNet.

[0076] In this embodiment, the displacement vector prediction module includes a multilayer perceptron (MLP) network and a residual connection structure;

[0077] The MLP network performs a nonlinear mapping on the feature vectors and outputs a three-dimensional displacement vector for each point.

[0078] The residual connection adds the original point coordinates to the predicted displacement vector to obtain the corrected point coordinates.

[0079] In this embodiment, the method for adding the dual independent noise includes:

[0080] First, generate a first noise that is identical to but independent of the original noise distribution, and add it to the point cloud to obtain the first noise point cloud;

[0081] A second noise, independent and identically distributed to the first noise, is generated and added to the first noise cloud to obtain a second noise cloud as network input.

[0082] In this embodiment, the noise intensity is preferably set to 0.1 through experiments, and the noise distribution is either Gaussian or uniform.

[0083] In this embodiment, the loss function used during network training is:

[0084]

[0085] in, For mean square error loss, For repulsive force loss, It is the influencing factor for controlling repulsive force loss, and is set to 0.0005.

[0086] In this embodiment, the repulsive force loss is achieved by constructing a pseudo-clean point cloud and calculating the maximum distance between points, in order to avoid excessive aggregation of the point cloud.

[0087] In this embodiment, the method further includes a network hyperparameter setting step:

[0088] The learning rate was adjusted to 0.001 using a cosine annealing algorithm;

[0089] The network structure is an encoder-decoder type, with a layer structure of 3-256-512-1024-2048;

[0090] The training rounds consist of 200 rounds, and the batch size is 128.

[0091] In this embodiment, the method performs visual and quantitative evaluations after noise reduction. The evaluation indicators include at least one of the following: mean square error, signal-to-noise ratio, Hausdorff distance, and structural similarity index.

[0092] A point cloud denoising system for performing any of the methods described above, comprising:

[0093] The data acquisition module is used to acquire the raw lidar point cloud;

[0094] The preprocessing module is used for data cleaning, normalization, and noise addition;

[0095] A noise reduction network module is used to perform unsupervised deep learning noise reduction.

[0096] The post-processing module is used to output and evaluate the noise reduction results.

[0097] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.

[0098] An electronic device includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method as described in any of the preceding claims.

[0099] Algorithm Method

[0100] Datasets and Data Preprocessing

[0101] The data used in this paper are from 17 forest plots in Qingyuan City, Guangdong Province. The actual sample plots were 80×80m in size, and the data were obtained through UAV flight scanning after manual ground surveys. Data collection took place in March 2024 using a BB4 UAV equipped with an AS-1300HL lidar system. The lidar scanner was a RieglVUX-1LR with a wavelength of 1550nm, a pulse length of 3.5ns, and a laser beam divergence angle of 0.5mrad. The pulse repetition frequency was 50kHz, the maximum scanning angle was 30°, and the scanning frequency was 49Hz. The UAV flew in a crisscross pattern, with a point cloud overlap of 50%, an average flight speed of 10m / s, and an average point cloud density of 110pts / m². 2 Sample plot illustration as follows Figure 1 As shown.

[0102] This paper selects 17 forest plots in Qingyuan City, Guangdong Province as experimental data, primarily for the following two reasons: First, Guangdong Province spans tropical and subtropical climate zones, and the forest land in this region is typically covered with abundant ground vegetation and other materials. These attachments often introduce significant data errors when segmenting individual trees, thus affecting the accuracy of forest land analysis. Using this algorithm to process such complex data can better demonstrate the algorithm's effectiveness in handling complex data scenarios. Second, forest resource management in Guangdong Province is crucial. The diversity of the province's topography, climate, and soil conditions has formed a symbiotic ecosystem of flora and fauna dominated by forests, making it an important gene pool for rare and endangered plants and animals in southern China. Currently, Guangdong Province has recorded 6,658 species of wild higher plants and 1,052 species of terrestrial vertebrate wildlife. Among them, 10 plant species, including cycads and water pines, are listed as national first-class protected plants, while 63 wild animal species, including the Chinese pangolin and the Chinese alligator lizard, are listed as national first-class protected animals, becoming "permanent residents" of Guangdong Province. This rich biodiversity makes Guangdong occupy an important position in China's ecological protection and research field.

[0103] Compared to some public datasets and datasets provided by research, the dataset used in this paper has undergone actual manual ground surveys and includes a large amount of field imagery. These field images can be matched one-to-one with the point cloud data, allowing for adjustments and comparisons based on the imagery during post-processing, thereby improving the accuracy of the data and the precision of the processing.

[0104] Before data cleaning, this paper first performs preliminary processing and classification of the collected data, matching point cloud data with actual imagery to ensure consistency between the point cloud data and actual geographical locations. During processing, specific functions are used to process the point cloud data, generating arrays containing the point cloud data, which are then stored in the required format for use in subsequent steps. After segmenting the plot data, 60 samples were selected as experimental subjects.

[0105] To improve the stability of model training, accelerate model convergence, reduce bias, and facilitate comparison of various algorithms in later stages, this paper also performs normalization processing on the point cloud data. This step ensures the comparability of data among different algorithms and models, and improves the reliability and applicability of the research results.

[0106] Basic idea

[0107] This paper designs a solution based on the Noise4Denoise algorithm, redesigning the algorithm and appropriately organizing its steps to adapt it to the needs of point cloud noise reduction for single-tree segmentation using LiDAR. The Noise4Denoise algorithm was chosen for two reasons: firstly, it cites the algorithm's core concepts and has demonstrated good experimental results; secondly, it is an unsupervised algorithm, which does not require manually labeled training data and can learn directly from unlabeled data. This makes it particularly suitable for large-scale, complex datasets, especially when labeling is costly or unavailable. Unsupervised algorithms can work effectively in various environments, regardless of whether the data is high-dimensional, unstructured, or lacks explicit classification or labels. This gives it broad application potential in various data analysis tasks.

[0108] The algorithm designs its basic model during the training process based on the following relationship: A clean point cloud with n points can be considered as... The three-dimensional coordinates are:

[0109]

[0110] in This represents the coordinates of a point in three-dimensional space. Therefore, a point cloud with noise can be defined as:

[0111]

[0112] in For noise. Redefine a new noise. , resampling range and Same, but and They are independent and not entirely the same. join in When another point cloud is obtained :

[0113]

[0114] here This results in double noise, containing twice the amount of noise. The algorithm assumes that a clean point cloud can be estimated based on the input noise. Set as a priori value, representing the observable double-noise point cloud, through the expected value. This represents the estimation of the overall surface of the noise point cloud. According to formula (2), we have:

[0115]

[0116] because and They are independent and identically distributed, and all sampled from the same distribution range, therefore their expected values ​​are the same. That is, Therefore, it can be deduced that:

[0117]

[0118] Therefore, it can be deduced that:

[0119]

[0120]

[0121] Here it is used , , , They represent The predicted value. Substituting into formula (6), we can obtain:

[0122]

[0123] Based on the conditional expectation rule and the discrete characteristics of point clouds, we can obtain Assuming If the values ​​in the table have a one-to-one correspondence, then it can be... For each point, a noisy displacement vector is estimated, and these vectors are used to construct an approximate clean point cloud. This relationship can be expressed as:

[0124]

[0125] in Point The predicted displacement vectors can be used to directly estimate the position of the corresponding clean point cloud.

[0126] exist Adding another set of displacement vectors will yield the predicted value. In other words:

[0127]

[0128] Based on the assumption of one-to-one correspondence, by substituting the value into formula (7), it can be proven that the direct prediction can be made. It is twice as much Equivalent:

[0129]

[0130]

[0131]

[0132]

[0133] Model Architecture

[0134] Noise4Denoise proposes an unsupervised point cloud denoising method. Its core idea is to utilize a deep neural network to learn a mapping from noisy point clouds to noise-free point clouds. Specifically, it learns a displacement vector field that allows each noisy point to move to its predicted clean location. This process is achieved by training the network to minimize the difference between the denoised point cloud and certain criteria. The model mainly consists of the following modules: Input layer: The input is noisy point cloud data, typically input in a double-noise form (i.e., adding two different noises to the original point cloud). Feature extraction module: Uses a deep learning-based network to extract local and global features of the point cloud. Displacement vector prediction module: Based on the extracted features, the model learns to predict the displacement vector of each point, thus achieving the mapping from noisy points to clean points. Output layer: The output is the denoised point cloud, obtained by displacement correction of the noisy points.

[0135] The feature extraction module draws inspiration from network architectures such as PointNet and DGCNN, capturing the geometric structure of point clouds by learning local and global features. To ensure the preservation of point cloud details during noise reduction, the module includes local neighborhood feature extraction, determining the neighborhood of each point using the k-nearest neighbor (k-NN) method, and then extracting local geometric features through graph convolution. Global feature aggregation is then performed, integrating local features into a global description through global pooling.

[0136] The displacement vector prediction module receives the features output by the feature extraction module and generates a displacement vector for each point. To preserve the structure and details of the point cloud, the model is designed to use an MLP network to perform a nonlinear transformation on the extracted features to generate the displacement vector; a residual connection is added between the generated displacement vector and the original point coordinates to ensure the model's stability and expressive power.

[0137] To preserve the structure and details of the point cloud during noise reduction, the model designs a specific symmetric loss function to ensure that the model does not destroy the overall geometry of the point cloud; through local feature extraction and the interrelationships between neighboring points, the model can identify and preserve local structures; and regularization techniques such as Dropout or weight decay are used to prevent the model from overfitting, thereby better preserving the details of the original point cloud.

[0138] Loss function design

[0139] use In Formula 10, for Predicting actual values ​​to guide displacement The estimation was performed. A mean squared error loss function was designed, which can be defined as:

[0140]

[0141] To ensure appropriate spacing between denoised point clouds and avoid potential clustering problems, the algorithm employs a repulsive loss mechanism to assist in the distribution of clean point clouds. First, a pseudo-clean point cloud is defined. for

[0142]

[0143] in It doesn't actually have a clean appearance. Next, we'll use the defined process to query pseudo-clean point clouds.

[0144]

[0145] Here, due to the assumption Point clouds are less affected by noise, so they are used. Come to Perform a query, and then perform a query on the point set. Normalization is performed. Using this, the loss function can be defined as...

[0146]

[0147] Then the loss function It can be defined as

[0148]

[0149] in It is the influencing factor for controlling repulsive force loss, and is set to 0.0005.

[0150] Experimental environment

[0151] Hardware environment

[0152] The hardware configuration used in this study is as follows: a computer equipped with an Intel(R) Core(TM) i5-10500 processor with a base frequency of 3.1GHz, 6 cores and 12 threads; an NVIDIA GeForce RTX 2070 SUPER graphics card with 8GB of video memory; and a computing platform equipped with 32GB of RAM and 4TB of storage space to support data processing and experimental calculations.

[0153] Software environment

[0154] This paper describes the computational experiments conducted under the Windows 11 Pro operating system, using CUDA 12.5 as the parallel computing framework. To facilitate model training management, PyCharm was used as the development environment, and CloudCompare was employed for point cloud data visualization. Specific hardware and software configurations are detailed in Table 1.

[0155] Table 1. List of hardware and software environments required for this experiment

[0156]

[0157] Hyperparameter settings

[0158] In this experiment, all parameters were determined through inference experiments. For the learning rate setting experiment, we used the cosine annealing algorithm to initially estimate the learning rate, ranging from 0.1 to 0.0001, and ultimately selected 0.001 as the optimal learning rate. Furthermore, the weight parameter in the loss function was set to 0.0005 according to the original algorithm. Regarding the noise addition level, after testing different noise intensities of 0.01, 0.02, ..., 0.1, 0.1 was ultimately selected as the optimal value. In the encoder and decoder structure design, a layer-by-layer optimization approach was adopted, ultimately determining a 3-256-512-1024-2048 architecture, achieving a good balance between experimental results and computation time. This study conducted a total of 200 training rounds, with each training batch containing 128 samples.

[0159] Common evaluation metrics for point cloud noise reduction

[0160] Common evaluation methods for radar point cloud denoising include mean square error (MSE), signal-to-noise ratio (SNR), Hausdorff distance, and structural similarity index (SSIM). This paper chooses MSE, SNR, Hausdorff distance, and SSIM as evaluation criteria because they comprehensively measure the effectiveness of point cloud denoising from different dimensions. MSE evaluates the algorithm's accuracy, SNR measures signal sharpness, Hausdorff distance evaluates geometric accuracy, and SSIM focuses on structural fidelity. By combining these four metrics, the performance of the denoising algorithm can be comprehensively and accurately evaluated in various aspects, thereby better understanding the algorithm's effectiveness and limitations in practical applications.

[0161] (1) Mean Squared Error (MSE) refers to the expected value of the squared difference between the estimated and true parameter values, used to describe the degree of difference in point clouds before and after denoising. It quantifies the prediction error of the algorithm or the accuracy of the model by calculating the average of the sum of squared errors between the predicted and true values. Its formula can be expressed as:

[0162]

[0163] in It is the actual value. This represents the predicted value. The smaller the MSE, the closer the model's prediction is to the true value. Therefore, in the evaluation of point cloud denoising, a lower MSE indicates a better denoising effect.

[0164] (2) Signal-to-noise ratio (SNR) measures the ratio of signal energy to noise energy, usually expressed in decibels (dB). In point cloud processing, SNR describes the relative ratio of signal strength to noise strength in a point cloud, reflecting the clarity of the signal and the degree of noise interference. The common formula for the ratio of signal power to noise power is the logarithm multiplied by 10, as follows:

[0165]

[0166] in The power of the signal. The logarithmic operation in this formula amplifies the difference between the signal and noise power, allowing the SNR to intuitively represent the quality of the signal in decibels.

[0167] (3) The Hausdorff distance is a measure of the similarity between two sets of points. It measures the maximum distance between two point clouds and is often used to evaluate the geometric differences between point clouds before and after denoising. Specifically, the Hausdorff distance is defined as the maximum distance between a point in one set and the nearest point in another set. If we calculate point cloud sets (1) and (2), then the one-way Hausdorff distance between these two point sets is:

[0168]

[0169] in, express and The Euclidean distance between them Also known as the forward Hausdorff distance

[41] . The Hausdorff distance emphasizes the minimum distance between the farthest points, thus it can sensitively capture the maximum difference between point clouds. This makes it a powerful tool for evaluating the denoising effect of point clouds, especially in applications where geometric accuracy needs to be guaranteed.

[0170] (4) The structural similarity index (SSIM) is an important metric for measuring the similarity between two images, especially for capturing changes in structural information. It is also widely used to evaluate the structural similarity between two point clouds. In 3D point cloud processing, the 3D point cloud can be projected onto a 2D plane for calculation, or a voxelization method can be used to segment and calculate the point cloud. In this paper, we adopt a neighborhood-based method to directly calculate the SSIM in 3D space. Its basic formula is:

[0171]

[0172] in, and Images and The average value; and Images and The variance; For image and Covariance between them; and These are two constants added to avoid the denominator being too small, among which This represents the dynamic range of pixel values. In 3D point cloud applications, SSIM can measure the structural similarity of point clouds before and after denoising by analyzing their local structure and global distribution. This method allows for a more effective evaluation of the ability of point cloud processing algorithms to preserve geometric structure and detail.

[0173] Results Analysis

[0174] Quantitative analysis

[0175] This paper analyzes the results of sixty experiments and additionally presents the point cloud data processing results for two sample plots (named S5 and S10, respectively). To evaluate the accuracy, versatility, and effectiveness of the algorithm used in single-tree segmentation radar point cloud processing, this paper selects several common denoising methods for comparative analysis, including morphological filtering, progressive triangulation filtering, statistical filtering, and radius filtering. These methods cover different processing dimensions, making the comparison more comprehensive and helping to demonstrate the adaptability of deep learning algorithms in processing diverse data features. By comparing with these traditional methods, the robustness and adaptability of deep learning algorithms in various denoising scenarios can be evaluated. Through comparison with these traditional filtering methods, this paper can more comprehensively demonstrate the innovation and effectiveness of deep learning algorithms in point cloud denoising, while clarifying their advantages over traditional methods in practical applications. This comparative analysis not only helps to understand the performance of the new algorithm but also provides important references for its further optimization and practical application.

[0176] Plots 5 and S10 were selected for analysis due to their complexity, which makes them highly representative. First, these two plots encompass both dense and relatively sparse forest areas, reflecting diverse forest conditions. Second, their rugged terrain further highlights their suitability as representative samples.

[0177] To quantify the denoising effectiveness of these methods, this paper calculates the mean squared error (MSE), signal-to-noise ratio (SNR), Hausdorff distance, and structural similarity index (SSIM) after processing by each method. The results of several algorithms are compared. MSE and SNR provide different but complementary denoising quality metrics; combining these two metrics allows for a comprehensive evaluation of the algorithm's performance. The SSIM and Hausdorff distance evaluate the point cloud denoising process from different perspectives, providing a more comprehensive assessment of the algorithm's performance. By comparing SSIM and Hausdorff distance together, we can evaluate the denoising algorithm's performance in terms of structural fidelity (via SSIM) and geometric accuracy (via Hausdorff distance). This dual analysis provides a more comprehensive evaluation of the algorithm's effectiveness, ensuring that the algorithm not only maintains the visual and structural similarity of the point cloud but also preserves its accurate geometric construction.

[0178] This paper calculates the standard deviation and mean deviation of different algorithms on four indicators, and the results are shown in Table 2.

[0179]

[0180] In Table 2, firstly, regarding the MSE value, DEN4 has the lowest mean MSE (0.0094) and the smallest standard deviation (0.0008), indicating the smallest error and the highest stability. Other methods (such as Morphological and SOR) have mean MSE values ​​in the range of 0.0297-0.0300, significantly higher than DEN4. Secondly, regarding the SNR value, DEN4 has the highest mean (149.1570), indicating the best signal quality after noise reduction. Its smallest standard deviation (0.5628) indicates stable SNR performance. Next, regarding the Hausdorff value, DEN4 has the lowest mean (0.8503), indicating the best preservation of point cloud geometry. Its smallest standard deviation (0.0947) indicates stable geometric characteristics. Finally, regarding the SSIM value, DEN4 has the highest mean (0.8399) and the smallest standard deviation (0.0054), indicating the best structural similarity performance and the smallest fluctuation.

[0181] In the S10 point cloud, the algorithm's MSE value was reduced by 70.2% compared to the asymptotic triangular network filter and by 37.8% compared to the radius filter, reaching 0.0105. This indicates that the algorithm effectively maintains the accuracy of the point cloud data while reducing noise. The algorithm's SNR value is 1-5 dB higher than other algorithms, reaching 149.4042, which is the highest level, indicating that the algorithm has a significant advantage in enhancing signal quality. The algorithm's Hausdorff distance is significantly lower than other methods, at 0.8466, indicating that it has very high accuracy in preserving the point cloud geometry. The algorithm's SSIM value is slightly lower than that of morphological filtering and statistical filtering, but the difference is small. Furthermore, the algorithm demonstrated high consistency and robustness in processing point cloud data from different sample plots (S5 and S10). The MSE value changed by 0.0001, the SNR by 0.0523, the Hausdorff value by 0.0350, and the SSIM value by 0.130. The overall error was within the range of 0% to 5%, showing excellent stability compared to other algorithms.

[0182] This paper uses visualization methods to further present the experimental data. Figure 2 The mean and standard deviation of different algorithms on four key metrics (MSE, SNR, Hausdorff distance, and SSIM) are shown. Figure 3 The Euclidean distance distributions of five algorithms (DEN4, Morphological, PTD, SOR, and ROR) are shown.

[0183] exist Figure 2Among the results, DEN4's distribution is highly concentrated, mainly within a small distance range (between 0 and 1), and it exhibits the highest density peak. This indicates that DEN4's denoising results are more compact, with smaller inter-point distances, demonstrating superior denoising performance and higher stability. In contrast, Morphological, PTD, SOR, and ROR's distributions are more dispersed, especially with higher density over a larger distance range (between 2 and 5), suggesting that these algorithms have weaker denoising effects and may lead to point cloud diffusion or discretization.

[0184] exist Figure 3 Among the algorithms, DEN4 exhibits the lowest MSE and the smallest standard deviation, demonstrating its significant advantage in minimizing error. Other algorithms (such as Morphological and PTD) have higher MSEs and greater error fluctuations. Furthermore, DEN4's SNR is significantly higher than other algorithms, indicating that it produces the best signal quality after noise reduction. Although the SNR values ​​of other algorithms are not significantly different, they are all lower than those of DEN4.

[0185] In terms of Hausdorff distance, DEN4 has the lowest value, indicating that it performs best in preserving the geometry of point clouds. Other algorithms, especially Morphological and PTD, have significantly higher Hausdorff distances, indicating that they perform poorly in geometry preservation. In addition, DEN4 has the highest SSIM value with the least fluctuation, further validating its superior ability to preserve the structural similarity of point clouds.

[0186] Visual assessment

[0187] exist Figure 4 In this paper, we tested the effect of denoising on S10 point clouds before and after processing by sampling. We randomly selected 10,000 points after processing for demonstration, and the results were significant. In the original point cloud, the areas marked with black boxes showed uneven thickness and a large amount of noise. However, after processing, the point cloud exhibited a more consistent thickness, and the distance between points reached a more ideal state. The denoising process significantly improved the overall smoothness of the point cloud, reduced the aggregation of large noisy point clouds, and also reduced floating noise points, making the point cloud data more uniform and clear.

[0188] exist Figure 5 and Figure 6In the diagram, we present the results of different algorithms processing S5 and S10 point cloud data, along with a frontal view of the original point clouds. Compared to other methods, our algorithm better renders the outline of each tree and preserves key features such as crown width, tree height, and diameter at breast height (DBH) to the greatest extent possible. Specifically, it is noticeable that our algorithm uses darker colors in higher areas, indicating a denser point cloud with less noise in the outer contour. The point cloud outside the crown is also denser, demonstrating superior crown recognition and more complete feature preservation. It also performs well in trunk recognition, clearly displaying the overall outline of the trunk. In contrast, other algorithms fall short in analyzing the overall tree structure and outline, leading to loss and distortion of some details. These algorithms fail to fully preserve the natural outline of trees when dealing with complex morphologies, particularly in the preservation of crown and trunk features. Our algorithm, however, better preserves the overall tree structure, avoiding significant distortion and information loss, resulting in more accurate and realistic processing results.

[0189] Figures 7 and 8 show the results of different algorithms processing S5 and S10 point cloud data, along with an overhead view of the original point clouds, to compare the overall shape and crown width of the point clouds after denoising. The method used in this paper performs excellently in preserving the overall crown width features, effectively removing most of the noise and making the shape of the main part of the point cloud clearer and more complete. Compared with other algorithms, this method not only better preserves the crown width features but also significantly reduces the impact of noise on the shape of the point cloud, thereby improving the visual effect after denoising.

[0190] right Figure 5 Comparing the various top-down views in section .6, it's clear that the point clouds processed by SOR, ROR, and Morphology show lighter colors in the tree crowns and small patches of missing color compared to the original point clouds. The overall tree shapes are also less distinct. While the point cloud processed by PTD shows better crown coverage, the appearance of the tree trunks and other parts of the ground cloud is significantly compromised. In contrast, the point cloud processed by this algorithm exhibits more distinct overall tree shapes, better crown coverage, fewer missing elements, and better preservation of both details and overall shape.

[0191] In this study, we propose an innovative unsupervised deep learning point cloud denoising algorithm, DEN4, which aims to improve the denoising performance of radar point clouds in single-tree segmentation, thereby improving the effect of subsequent single-tree segmentation and related research.

[0192] DEN4 incorporates a multi-level noise separation module in its network architecture, enabling efficient differentiation between signal and noise. This mechanism ensures that while removing large noise, the fine structure of the point cloud is preserved, significantly reducing errors and improving the signal-to-noise ratio (SNR). This dual noise processing design makes DEN4 particularly suitable for LiDAR point cloud denoising in complex scenarios, demonstrating stable performance across various data environments. Experimental results show that DEN4's mean MSE is only 0.0094, significantly lower than Morphological (0.0299) and PTD (0.0297), with a standard deviation of only 0.0008, indicating a substantial reduction in denoising error and stable results. On the S10 dataset, DEN4 reduces MSE by 70.2% compared to asymptotic triangulation filtering and by 37.8% compared to radius filtering. Furthermore, DEN4's mean SNR reaches 149.1570, significantly higher than Morphological (146.1743) and PTD (146.1301). On the S5 dataset, DEN4's SNR was further improved to 149.4042, demonstrating excellent signal preservation capabilities.

[0193] DEN4 borrows the architecture from PointNet and DGCNN, and by optimizing the local feature extraction module, it can accurately capture the subtle features of point clouds. This design preserves the local geometric details of the point cloud (such as tree height and diameter at breast height) during the noise reduction process, effectively avoiding the detail loss and over-smoothing problems common in traditional methods. Through this technique, DEN4 can maintain the consistency of the core structure of the point cloud at different scales, significantly improving its ability to preserve geometry. Experimental data show that DEN4's mean Hausdorff distance is 0.8503, significantly lower than Morphological (5.3568), PTD (5.5593), and SOR (5.3034). In the S5 dataset, DEN4's Hausdorff distance is further reduced to 0.8116, demonstrating its excellent ability to preserve complex geometric structures.

[0194] By combining global and local feature learning, DEN4 not only focuses on details during denoising but also maintains the consistency of the overall point cloud structure. Its global feature learning capability effectively avoids the overall shape distortion problem common in traditional filtering methods, ensuring that the denoised point cloud structure remains highly consistent with the original data. The unsupervised learning mechanism further enhances the algorithm's adaptability to different datasets, resulting in excellent performance across various environments. This characteristic is reflected in the SSIM metric: DEN4's mean SSIM is 0.8399, outperforming other algorithms in most experiments, with minimal fluctuation (standard deviation of only 0.0054), indicating stable overall structure preservation in the denoised point cloud.

[0195] As an unsupervised algorithm, DEN4 does not require pre-labeled data and can adaptively capture features from different datasets. This characteristic enables DEN4 to exhibit consistency and robustness across diverse forest point cloud datasets. Especially in complex data environments, unsupervised learning ensures a high degree of consistency between the denoised point cloud data and the original data, enhancing the algorithm's versatility in various scenarios. Final experimental results show that DEN4's MSE standard deviation is only 0.0008, significantly lower than other algorithms, indicating extremely low error fluctuations during the denoising process. Furthermore, DEN4's SNR standard deviation is 0.5628, far lower than Morphological (1.1729) and PTD (1.1254), indicating high signal quality stability.

[0196] This study introduces a comprehensive and innovative design, including dual noise processing, local feature extraction, global feature learning, and unsupervised learning, to comprehensively improve the quality and stability of point cloud denoising. The superior performance of these designs is reflected in significant advantages in key metrics: dual noise processing optimizes MSE and SNR performance; local feature extraction significantly reduces Hausdorff distance while preserving geometric structure; global feature learning improves the stability of SSIM; and the unsupervised learning mechanism ensures consistent denoising results across different datasets.

[0197] The organic combination of these design elements enables DEN4 to generate extremely clean and accurate point clouds while preserving key details, providing strong support for applications such as single-tree segmentation and forest resource management. These multi-dimensional improvements not only validate the algorithm's robustness but also demonstrate its transformative application potential in complex ecological environments.

[0198] LiDAR point cloud data is a crucial tool for forest surveys and management, providing important parameters and indicators for forest resource monitoring and management. However, traditional point cloud processing algorithms have shortcomings in terms of accuracy, robustness, and contour preservation. To address these issues, this paper conducts an in-depth study based on a dataset of 17 forest plots in Qingyuan City, Guangdong Province. After data preprocessing, this paper introduces and improves an unsupervised deep learning point cloud denoising algorithm and redesigns its structure to better adapt to the needs of single-tree segmentation. By adding different levels of noise to the same dataset and combining it with multi-layer deep learning processing, this paper successfully generates relatively clean point cloud data. Experimental results show that the improved algorithm can not only adapt to different single-tree segmentation LiDAR point cloud data but also significantly improves the denoising effect, making the tree contours clearer while effectively preserving key features such as tree height and diameter at breast height. Compared with traditional algorithms, this algorithm performs excellently in both quantitative analysis and visual effects, fully demonstrating its application potential in the field of single-tree segmentation LiDAR point cloud denoising.

[0199] While this study demonstrates the superiority of the algorithm in single-tree segmentation for radar point cloud denoising, there is still room for improvement. Since the dataset used primarily contains plots of a single tree species, the algorithm's effectiveness in handling other single-species and multi-species scenarios has not yet been verified. Further research is recommended to explore the algorithm's potential applications in complex mixed forests or large-scale forest surveys, particularly investigating the impact of different seasons and environmental factors (such as light, humidity, and vegetation density) on algorithm performance to enhance its universality and robustness. Therefore, future research should expand the diversity of datasets to test the algorithm's applicability to different tree species and complex forest environments. Furthermore, future research should focus on realizing a full-process deep learning algorithm for single-tree segmentation and integrating it with other optimization algorithms. This will help better address challenges in practical applications, especially when processing large-scale forest data. By integrating the advantages of multiple algorithms, research can develop more efficient and accurate point cloud processing methods to meet the growing needs of forest resource surveys and management.

[0200] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0201] The foregoing has described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0202] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for noise reduction of radar point clouds segmented from a single tree, characterized in that, Includes the following steps: Acquire forest point cloud data collected by airborne lidar, wherein the point cloud data includes individual tree structure information; The point cloud data is preprocessed, including point cloud matching, coordinate normalization, and sample segmentation; An unsupervised deep learning noise reduction network is constructed, the network including an input layer, a feature extraction module, a displacement vector prediction module, and an output layer; Add dual independent noise to the point cloud data to form a point cloud with dual noise as network input; The feature extraction module extracts local geometric features and global structural features of the point cloud. The displacement vector prediction module predicts a displacement vector for each point and corrects the position of the point cloud based on the displacement vector. Output denoised point cloud data for single tree segmentation, tree height extraction, or diameter at breast height estimation.

2. The method according to claim 1, characterized in that, The feature extraction module includes a local feature extraction submodule and a global feature aggregation submodule; The local feature extraction submodule uses the k-nearest neighbor algorithm to construct a local map of the point cloud and extracts neighborhood features through a graph convolutional network. The global feature aggregation submodule aggregates local features using max pooling or average pooling to form a global descriptor.

3. The method according to claim 2, characterized in that, The local feature extraction submodule adopts the DGCNN architecture, while the global feature aggregation submodule draws on the symmetric function structure of PointNet.

4. The method according to claim 1, characterized in that, The displacement vector prediction module includes a multilayer perceptron (MLP) network and a residual connection structure; The MLP network performs a nonlinear mapping on the feature vectors and outputs a three-dimensional displacement vector for each point. The residual connection adds the original point coordinates to the predicted displacement vector to obtain the corrected point coordinates.

5. The method according to claim 1, characterized in that, The method for adding the dual independent noise includes: First, generate a first noise that is identical to but independent of the original noise distribution, and add it to the point cloud to obtain the first noise point cloud; A second noise, independent and identically distributed to the first noise, is generated and added to the first noise cloud to obtain a second noise cloud as network input.

6. The method according to claim 5, characterized in that, The noise intensity is preferably set to 0.1 through experiments, and the noise distribution is either Gaussian or uniform.

7. The method according to claim 1, characterized in that, The loss function used in the network training process is: The loss function used in the network training process is: in, For mean square error loss, For repulsive force loss, It is the influencing factor for controlling repulsive force loss, and is set to 0.0005.

8. The method according to claim 7, characterized in that, The repulsive force loss is achieved by constructing a pseudo-clean point cloud and calculating the maximum distance between points, in order to avoid excessive aggregation of the point cloud.

9. The method according to claim 1, characterized in that, The method also includes a network hyperparameter setting step: The learning rate was adjusted to 0.001 using a cosine annealing algorithm; The network structure is an encoder-decoder type, with a layer structure of 3-256-512-1024-2048; The training rounds consist of 200 rounds, and the batch size is 128.

10. The method according to claim 1, characterized in that, The method further performs visual and quantitative evaluations after noise reduction, and the evaluation indicators include: Mean squared error, signal-to-noise ratio, Hausdorff distance, and structural similarity index are all considered.

11. A point cloud noise reduction system for performing the method according to any one of claims 1 to 10, characterized in that, include: The data acquisition module is used to acquire the raw lidar point cloud; The preprocessing module is used for data cleaning, normalization, and noise addition; A noise reduction network module is used to perform unsupervised deep learning noise reduction. The post-processing module is used to output and evaluate the noise reduction results.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 10.

13. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method as described in any one of claims 1 to 10.

Images