Single-tree Dynamic Monitoring Method and System Based on Multi-period Point Clouds and Deep Learning

Abstract

The present invention discloses a method and system for single-tree dynamic monitoring based on multi-temporal point clouds and deep learning, including collecting multi-temporal multi-spectral images and lidar point cloud data of a research area and performing preprocessing; registering and fusing the preprocessed multi-temporal multi-spectral images and lidar point cloud data to generate multi-temporal fused point clouds; performing single-tree segmentation on the basis of the multi-temporal fused point clouds using a deep learning model; inversely calculating the crown diameter and tree height of a single tree based on the single-tree segmentation result, and analyzing the growth parameters of the single tree according to the inversion results of the crown diameter and tree height. By fusing multi-temporal point clouds and multi-spectral image data and combining with a deep learning model, accurate segmentation and dynamic growth monitoring of single trees are achieved, overcoming the defects in the prior art such as low accuracy of single-tree segmentation, insufficient fusion of multi-source data, and inability to capture the growth changes of forest trees. By adopting the inversion models of crown diameter and tree height, the accuracy of growth parameter estimation is effectively improved, providing a scientific basis for forest resource management.

Description

Technical Field

[0001] The present invention belongs to the technical field of spectral imaging and laser point cloud, and particularly relates to a single-tree dynamic monitoring method and system based on multi-temporal point cloud and deep learning. Background Art

[0002] Single-tree dynamic monitoring is an important part of forest resource management and ecological research, which can provide scientific basis for evaluating the growth status of forest trees, calculating carbon sinks, and preventing and controlling pests and diseases. At present, common single-tree monitoring methods mainly rely on manual measurement and canopy extraction technology based on remote sensing images. However, manual measurement is inefficient, costly, and limited by operability in complex environments, while although remote sensing images can cover large areas of forest land, they are limited by resolution and it is difficult to accurately extract the three-dimensional structure information of forest trees. In addition, traditional point cloud data processing methods usually rely on rule-based models or threshold-based segmentation algorithms, and these methods show great limitations in dealing with the complexity of tree morphology and environmental noise. Especially in the case of dense stands and overlapping canopies, it is difficult to achieve accurate segmentation and dynamic growth monitoring of single trees.

[0003] With the development of deep learning technology, deep learning models based on point cloud provide new ideas for single-tree segmentation and growth monitoring. However, existing technologies often lack an effective mechanism for fusing multi-source data (such as laser point cloud and multi-spectral images) when dealing with multi-temporal point cloud, resulting in large registration errors and insufficient feature extraction. In addition, the monitoring of single-tree growth dynamics mostly stays in the stage of static feature inversion and cannot accurately capture the growth change rules of forest trees at different time nodes. Therefore, how to fully fuse multi-temporal point cloud and multi-spectral image data and combine deep learning models to achieve accurate dynamic monitoring of single trees and inversion of growth parameters has become an urgent problem to be solved in the current technical field. Summary of the Invention

[0004] The purpose of the present invention is to provide a single-tree dynamic monitoring method and system based on multi-temporal point cloud and deep learning to solve the problem of inaccurate inversion accuracy in the prior art.

[0005] To achieve one of the above-mentioned invention purposes, an embodiment of the present invention provides a single-tree dynamic monitoring method based on multi-temporal point cloud and deep learning, and the method includes,

[0006] Collect multi-temporal multi-spectral images and laser point cloud data of the research area and perform preprocessing;

[0007] Register and fuse the preprocessed multi-temporal multi-spectral images and laser point cloud data to generate multi-temporal fused point cloud;

[0008] Based on the multi-temporal fused point cloud, use a deep learning model for single-tree segmentation;

[0009] Based on the individual tree segmentation results, the crown width and tree height of individual trees are inverted, and the growth parameters of individual trees are analyzed according to the inversion results of the crown width and tree height.

[0010] As a further improvement of an embodiment of the present invention, the method further includes that the registration and fusion step includes:

[0011] Convert the preprocessed multi-spectral image into a multi-spectral image point cloud, and perform scale normalization on each point cloud;

[0012] Extract feature lines from the multi-spectral image point cloud and the laser point cloud respectively; the feature lines include edge feature lines representing the forest boundary lines and corner feature lines representing structural key points;

[0013] Based on the extracted feature lines, perform registration and fusion by calculating the spatial correspondence between the feature lines to generate a multi-temporal fusion point cloud; the multi-temporal fusion point cloud is a comprehensive point cloud dataset containing multi-temporal three-dimensional spatial information and spectral features.

[0014] As a further improvement of an embodiment of the present invention, the method further includes that the individual tree segmentation step includes:

[0015] Based on the multi-temporal fusion point cloud, use a deep learning model to perform individual tree segmentation and output a point cloud subset of each individual tree , where, is a single point in the point cloud, represents the individual tree number, represents the segmentation label of the i-th point;

[0016] The use of the deep learning model for individual tree segmentation includes,

[0017] Use the PointNet++ model to segment the multi-temporal fusion point cloud , where each point , is the spatial coordinate, is the spectral feature.

[0018] As a further improvement of an embodiment of the present invention, the method further includes that the inversion step of the crown width includes,

[0019] Extract the feature vector from through the feature extraction function;

[0020] Calculate the crown width , where is the neural network model adopted;

[0021] When training the neural network model, optimize the mean square error loss ;

[0022] Among them, is the total number of individual trees, is the true crown width.

[0023] As a further improvement of an embodiment of the present invention, the method further includes that the inversion step of the crown width further includes,

[0024] Defining a temporal smoothing term by regularization to capture the smooth trend of crown width changes, expressed as:

[0025] ,

[0026] Among them, is the predicted value of the crown width of the k-th individual tree in the t-th period, is the point cloud subset of the k-th individual tree in the t-th period 's eigenvector, T is the number of time periods;

[0027] Combining the environmental variables with to form an extended eigenvector , and using a neural network model integrating environmental data to calculate the crown width ; among them, is the annual average rainfall, is the average temperature;

[0028] Dynamically adjusting the regression model parameters by optimizing the total loss function, and the total loss function is expressed as:

[0029] ,

[0030] Among them, is the true crown width, and λ is the smoothing weight coefficient.

[0031] As a further improvement of an embodiment of the present invention, the method further includes that the inversion step of the tree height includes,

[0032] Classifying the tree species or identifying the tree shape of the point cloud subset , and outputting the tree species through the classification model;

[0033] Based on the tree species selecting a curve to match the crown shape; the types of the curve include parabola, high-order polynomial or spline curve;

[0034] Projecting onto the XZ plane to obtain a point set ;

[0035] Fitting the curve Generate a canopy height distribution model, and the optimization objective is:

[0036] ,

[0037] wherein, represents the number of points in, is the height of the projection point;

[0038] The calculation of tree height is expressed as:

[0039] ,

[0040] where, is the ground height.

[0041] As a further improvement of an embodiment of the present invention, the method further includes that the tree height inversion step further includes

[0042] optimizing the fitting effect by embedding an adaptive adjustment mechanism for growth model parameters, specifically including

[0043] introducing and the maximum height potential into the tree height fitting, and adjusting the fitting function to:

[0044] ,

[0045] where, is the tree age, is the spatial distribution function;

[0046] Extract the feature vector from the point cloud subset , and combine the external environmental data and tree species information to calculate the growth parameters through a regression model or a rule mapping:

[0047] ,

[0048] ,

[0049] where, is a prediction model based on a neural network or an empirical formula, and adaptively updates the parameters according to the current point cloud and environmental conditions;

[0050] Based on the adjusted , optimize the objective through the least squares method, generate a dynamic canopy height distribution model, and calculate the tree height;

[0051] The optimization objective is expressed as:

[0052] .

[0053] As a further improvement of an embodiment of the present invention, the method further includes that analyzing the growth parameters of a single tree according to the inversion results of crown width and tree height includes:

[0054] Performing spatio-temporal matching based on the multi-period single-tree segmentation results, where the set of single trees in the t-th period is ;

[0055] where is the total number of single trees in the t-th period, and each single tree corresponds to a point cloud subset ;

[0056] Calculating the center point of the single tree ;

[0057] The spatio-temporal matching formula is defined as: if , then , it is determined that the matching is successful, otherwise 0; where is the single-tree number in the (t + 1)-th period, is the distance threshold;

[0058] If the matching is successful, calculate the tree height growth rate , and the crown width growth rate ;

[0059] where is the tree height inversion result, is the crown width inversion result, is the time interval.

[0060] As a further improvement of an embodiment of the present invention, the method further includes that analyzing the growth parameters of a single tree according to the inversion results of crown width and tree height further includes

[0061] Improving the detection accuracy by introducing a closed-loop feedback mechanism, specifically including

[0062] Recording the sets of single trees in the previous periods , the matching results , the tree height and the crown width as historical data;

[0063] Calculating the matching success rate and the historical growth rate Adjusting the distance threshold:

[0064] ,

[0065] where is the adjustment coefficient, is a penalty factor to adapt to the changes in tree growth and forest area characteristics;

[0066] Dynamically adjust the inversion parameters using historical tree height sequences and crown width sequences, and recalculate based on the adjusted parameters and ;

[0067] Feed back the matching results and inversion parameters of the current period to the next period analysis to achieve a closed-loop process of monitoring, feedback, and optimization, thereby improving the success rate of spatio-temporal matching and the accuracy of growth rate calculation.

[0068] To achieve one of the above invention purposes, an embodiment of the present invention further provides a single-tree dynamic monitoring system based on multi-period point clouds and deep learning. The system includes a data acquisition module, a data fusion module, a segmentation module, and an analysis module;

[0069] The data acquisition module is used to collect multi-period multi-spectral images and laser point cloud data of the research area and perform preprocessing;

[0070] The data fusion module is used to register and fuse the preprocessed multi-period multi-spectral images and laser point cloud data to generate multi-period fused point clouds;

[0071] The segmentation module is used to perform single-tree segmentation based on the multi-period fused point clouds using a deep learning model;

[0072] The analysis module is used to invert the crown width and tree height of a single tree based on the single-tree segmentation results, and analyze the growth parameters of the single tree according to the inversion results of the crown width and tree height.

[0073] Compared with the prior art, the single-tree dynamic monitoring method and system based on multi-period point clouds and deep learning provided by the present invention achieve precise segmentation and dynamic growth monitoring of single trees by fusing multi-temporal point clouds and multi-spectral image data and combining deep learning models, overcoming the defects of low accuracy of single-tree segmentation, insufficient fusion of multi-source data, and inability to capture the growth changes of forest trees in the prior art. By using the inversion models of crown width and tree height, the accuracy of growth parameter estimation is effectively improved, providing a scientific basis for forest tree growth monitoring and forest resource management. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is the overall flowchart of the single-tree dynamic monitoring method based on multi-period point clouds and deep learning according to the present invention.

[0075] Figure 2 is the schematic architecture diagram of the single-tree dynamic monitoring system based on multi-period point clouds and deep learning according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0076] The present invention will be described in detail below in conjunction with the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, method, or functional transformation made by those of ordinary skill in the art based on these embodiments is included in the protection scope of the present invention.

[0077] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention and should not be construed as limiting the present invention.

[0078] In the first embodiment of the present invention, the present invention provides a method for single-tree dynamic monitoring based on multi-temporal point clouds and deep learning, as Figure 1 shown, the method includes,

[0079] S1: Collect multi-temporal multi-spectral images and lidar point cloud data of the research area and perform preprocessing;

[0080] S2: Register and fuse the preprocessed multi-temporal multi-spectral images and lidar point cloud data to generate multi-temporal fused point clouds;

[0081] S3: Based on the multi-temporal fused point clouds, use a deep learning model to perform single-tree segmentation;

[0082] S4: Based on the single-tree segmentation results, invert the crown width and tree height of the single tree, and analyze the growth parameters of the single tree according to the inversion results of the crown width and tree height.

[0083] In a specific embodiment of the present invention, the registration and fusion step is specifically as follows:

[0084] Convert the preprocessed multi-spectral image into a multi-spectral image point cloud and perform scale normalization on each point cloud;

[0085] Extract feature lines from the multi-spectral image point cloud and the lidar point cloud respectively; the feature lines include edge feature lines representing the forest boundary line and corner feature lines representing the structural key points;

[0086] Based on the extracted feature lines, perform registration and fusion by calculating the spatial correspondence between the feature lines to generate multi-temporal fused point clouds; the multi-temporal fused point clouds are a comprehensive point cloud dataset containing multi-temporal three-dimensional spatial information and spectral features.

[0087] It should be noted that in the present invention, multi - period fused point cloud refers to a comprehensive point cloud dataset containing multi - temporal three - dimensional spatial information and spectral features generated after spatially registering and information - fusing multi - period multi - spectral image point clouds with corresponding lidar point cloud data. Specifically, multi - period multi - spectral images refer to multi - spectral image data of the study area collected at different time intervals (such as annually or quarterly), and lidar point cloud data is three - dimensional point cloud data obtained through lidar technology within the corresponding time period. The goal of registration and fusion is to align these multi - period data spatially and integrate their spatial coordinates and spectral information to provide a unified high - precision data basis for subsequent individual tree segmentation and dynamic monitoring. In the implementation process, first, the pre - processed multi - period multi - spectral images are converted into multi - spectral image point clouds. Among them, each point cloud contains spatial coordinate information (x, y, z) and corresponding spectral feature information c. The spatial coordinates (x, y, z) are generated through geometric correction and depth estimation of the image, representing the position of the point in three - dimensional space; the spectral feature c comes from the reflectance of each band of the multi - spectral image (such as visible light, near - infrared band values), characterizing the spectral characteristics of the forest trees. To ensure the consistency of the point clouds from different data sources in spatial distribution and scale, scale normalization processing is performed on the generated multi - spectral image point clouds and lidar point clouds. Specifically, scale normalization reduces the registration error caused by differences in data acquisition equipment or resolution by calculating the spatial range of the point cloud and normalizing the coordinate values of all points to a unified scale interval (such as [0, 1]).

[0088] Furthermore, feature extraction is respectively performed on the multi - spectral image point cloud and the lidar point cloud. The extracted feature lines include edge feature lines and corner feature lines. "Feature line" refers to a set of linear features that can characterize the structural outline of forest trees, extracted from the point cloud data by analyzing the spatial distribution characteristics of the point cloud. Specifically, the edge feature line is the forest tree boundary line determined by calculating the gradient change rate of the point cloud, that is, identifying areas in the point cloud where there are significant changes in height or density (such as the boundary between the tree crown and the background), and connecting the points in these areas into continuous line segments; the corner feature line is a line segment formed by identifying key points with sudden curvature changes in the point cloud (such as the top of the tree crown or the intersection of branches). These key points usually have relatively high local curvature values, which can be extracted and connected into lines through curvature calculation formulas (such as principal curvature estimation based on neighborhood points). This process ensures that the feature lines extracted from the multi - spectral image point cloud and the lidar point cloud have sufficient discrimination and robustness, can accurately reflect the spatial structure characteristics of forest trees, and provide a reliable reference for subsequent registration.

[0089] Furthermore, the registration and fusion are achieved through the "feature line matching algorithm". This algorithm aims to establish the spatial correspondence relationship between the feature lines of the multi - spectral image point cloud and the lidar point cloud, and finally generate the fused point cloud by optimizing and adjusting their relative positions. In this embodiment, the "feature line matching algorithm" includes the following steps:

[0090] Feature line extraction and description: Edge feature lines and corner feature lines are extracted from the multi-spectral image point cloud and the laser point cloud respectively, and descriptors are generated for each feature line. The descriptor can be a vector representation based on line segment length, direction, curvature distribution, or neighborhood point statistical features (such as point density), and is used to quantify the geometric and spatial attributes of the feature line.

[0091] Calculation of the correspondence relationship between feature lines: By comparing the descriptors of the two sets of feature lines, the similarity between them is calculated to determine the matching pairs. The "correspondence relationship between feature lines" refers to finding the pair of feature lines that represent the same forest structure (such as the same tree crown edge or vertex) in space in the multi-spectral image point cloud and the laser point cloud by calculating the Euclidean distance or cosine similarity as the similarity metric. By traversing all pairs of feature lines, a preliminary set of matching relationships is generated.

[0092] Registration optimization: Based on the set of matching relationships, the least squares method or the Iterative Closest Point (ICP) algorithm is used to optimize the registration result. The least squares method calculates the best rigid transformation of the two point clouds by minimizing the sum of the squares of the distances between the matching feature line pairs; the ICP algorithm further refines the registration accuracy by iteratively finding the closest point pairs and updating the transformation parameters. After optimization, the spatial positions of the multi-spectral image point cloud and the laser point cloud are aligned, and their point data are integrated in chronological order to form a multi-temporal fused point cloud.

[0093] In a specific embodiment of the present invention, the single-tree segmentation step is specifically as follows.

[0094] Based on the multi-temporal fused point cloud, a deep learning model is used for single-tree segmentation, and the point cloud subset of each single tree is output. , where is a single point in the point cloud, represents the single-tree number, represents the segmentation label of the i-th point;

[0095] The single-tree segmentation using the deep learning model includes

[0096] Using the PointNet++ model to segment the multi-temporal fused point cloud where each point , is the spatial coordinate, is the spectral feature.

[0097] Furthermore, the segmentation loss function is expressed as:

[0098] ,

[0099] where is the true class distribution, is the prediction probability, is the total number of individual trees.

[0100] It should be noted that the goal of individual tree segmentation is to accurately separate the point cloud data of each tree from the multi-temporal fused point cloud, laying a foundation for subsequent parameter inversion and dynamic analysis. Specifically, this step uses the multi-temporal fused point cloud as input and performs individual tree segmentation through a deep learning model, finally outputting the point cloud subset of each individual tree. The multi-temporal fused point cloud is generated by registering and fusing multi-spectral image point cloud and lidar point cloud, containing the three-dimensional spatial information and spectral attributes of the study area at different time periods. Compared with the traditional segmentation method based on single-phase data, this step makes full use of the spatio-temporal continuity of multi-phase data and realizes high-precision individual tree recognition through deep learning technology. It is applicable to dynamic monitoring scenarios and can capture the morphological changes of trees during the growth process.

[0101] Furthermore, at the initial stage of individual tree segmentation, the multi-temporal fused point cloud generated based on the aforementioned registration and fusion step is used as input data, and an advanced deep learning model is adopted for individual tree segmentation processing. The multi-temporal fused point cloud is a comprehensive data set obtained by registering and fusing multi-spectral image point cloud and lidar point cloud data of multiple time phases, containing the three-dimensional spatial information (coordinates) and spectral characteristics (reflectivity) of forest trees in the study area at different time nodes. By analyzing this data set through a deep learning model, the spatial boundaries and class affiliations of each individual tree can be effectively identified, and finally the point cloud subset of each individual tree is output. This segmentation method is not only applicable to single-time-phase data, but also can capture the dynamic change characteristics of trees through the spatio-temporal continuity of multi-phase data, laying a foundation for the subsequent analysis of growth parameters.

[0102] Furthermore, the PointNet++ model segments the multi-temporal fused point cloud through a hierarchical feature extraction and classification process. Its core steps include the following aspects: Local feature extraction: The model first groups the point cloud based on the spatial neighborhood of points, and uses the spherical query or KNN algorithm to determine the neighborhood point set within the local area for each point. Then, the spatial geometric features and spectral features of each local area are extracted through a multi-layer perceptron to form a preliminary feature representation. Multi-scale feature aggregation: PointNet++ adopts a multi-scale sampling and grouping strategy, and aggregates local features into global features by gradually reducing the number of sampled points and expanding the neighborhood range. This hierarchical structure can capture diverse information from the details of the tree crown to the overall morphology, and is particularly suitable for the segmentation task of overlapping or dense forest areas. Classification and label output: After the feature extraction is completed, the model generates the class probability distribution for each point through a fully connected layer and a softmax function, and outputs the segmentation label . Finally, after the label of each point is determined, the point cloud subset Automatically generated by filtering points.

[0103] Furthermore, to train the PointNet++ model and optimize its segmentation performance, a cross-entropy loss function is adopted to measure the difference between the prediction result and the ground truth label. During the training process, by minimizing , the model uses the backpropagation algorithm and an optimizer (such as the Adam optimizer) to adjust the network parameters and gradually improve the segmentation accuracy. The training data usually comes from a manually labeled point cloud dataset, where the point cloud subset and label of each single tree have been determined in advance. Through multiple iterations, the model can learn the complex mapping relationship between the spatial structure and spectral features of the point cloud and the single tree category.

[0104] In a specific embodiment of the present invention, the steps for inverting the crown width are specifically as follows:

[0105] Extract feature vectors from through a feature extraction function;

[0106] Calculate the crown width , where is the neural network model adopted;

[0107] When training the neural network model, optimize the mean squared error loss ;

[0108] where, is the total number of single trees, is the true crown width.

[0109] It should be noted that the steps for inverting the crown width are aimed at accurately estimating the crown width of each tree from the point cloud subset obtained by single tree segmentation. The specific implementation process first uses a machine learning regression model to directly regress the crown width from the point cloud subset of a single tree. By introducing a machine learning regression model, the present invention can integrate various features of the point cloud and overcome the limitations of traditional methods when the crown shape is irregular or the point cloud data is sparse. This data-driven approach significantly improves the accuracy and robustness of the estimation by learning the complex mapping relationship between the point cloud and the crown width.

[0110] Furthermore, before regressing the crown width, it is necessary to extract effective feature representations from the point cloud subset . Specifically, the feature extraction process usually includes the following aspects: First, extract spatial geometric features, such as the boundary range of the point cloud, height distribution statistics (mean, variance, etc.), and three-dimensional shape descriptors (such as the eigenaxes obtained by principal component analysis); Second, combine spectral features, such as the mean reflectance of each band in the multi-spectral image point cloud or texture information. These features are integrated into a multi-dimensional vector , fully characterize the spatial structure and spectral properties. The diversity and comprehensiveness of feature extraction provide rich input information for the subsequent neural network model, ensuring that crown width estimation can adapt to the differences of different tree species and growth environments.

[0111] Furthermore, based on the extracted feature vectors , use a neural network model to calculate the crown width of a single tree . The neural network model here can adopt a multi-layer perceptron (MLP), a convolutional neural network (CNN) or other regression network structures. Its core function is to transform the feature vectors into crown width values . The input layer of the model receives the feature vectors , and after being processed by the hidden layer (such as the non-linear transformation of the activation function ReLU) and parameter weight adjustment, finally generates the predicted value of the crown width at the output layer . Compared with traditional geometric measurement methods, the neural network can learn the implicit relationship between point cloud features and crown width, and shows higher flexibility and accuracy especially when dealing with non-circular or irregular tree crowns. In addition, if there is noise or missing parts in the point cloud data, the generalization ability of the neural network can also effectively alleviate the impact of these problems.

[0112] It should be noted that to ensure the prediction accuracy of the neural network model, the model parameters are adjusted by optimizing the mean square error loss function during the training stage. The goal of the mean square error loss function is to quantify the gap between the predicted value and the true value. By minimizing , the model gradually learns the precise mapping relationship between the point cloud features and the crown width. During the training process, the backpropagation algorithm and an optimizer (such as Adam or SGD) are used to update the network weights, usually combined with batch training and learning rate scheduling strategies to accelerate convergence and avoid overfitting. In addition, to enhance the generalization ability of the model, regularization techniques (such as L2 regularization) or data augmentation means (such as point cloud rotation, scaling) can be introduced to improve its performance on different datasets.

[0113] Preferably, in a specific embodiment of the present invention, the crown width inversion framework further incorporates a temporal smoothing term and environmental data feedback to achieve dynamic adaptive adjustment of the regression model parameters.

[0114] It should be noted that the inversion framework for crown width aims to accurately estimate the width of the tree crown from a subset of single-tree point clouds, which is based on predicting the crown width through a machine learning regression model. However, a single static regression model may not be able to fully adapt to the dynamic changes during the tree growth process or the diversity of environmental conditions. Therefore, a temporal smoothing term and an environmental data feedback mechanism are further integrated to form a dynamic adaptive inversion framework. This framework not only relies on the spatial and spectral features of the point clouds but also optimizes the parameter adjustment process of the regression model by introducing time series information and external environmental variables, thereby improving the accuracy and stability of crown width estimation. Such a design can better reflect the changing trend of the crown width of trees at different growth stages and ecological conditions, providing more reliable data support for dynamic monitoring.

[0115] Furthermore, in traditional crown width inversion, the regression model usually independently predicts the crown width value based on single-phase point cloud data. . However, tree growth has temporal continuity, and the crown width changes between multiple phases should show a smooth trend rather than sudden changes. To capture this temporal characteristic, a temporal smoothing term is introduced into the framework, specifically incorporated into the loss function through regularization. Assume that the predicted crown width value at the t-th phase is , where is the feature vector of the point cloud subset . The temporal smoothing term is defined as the difference penalty of the crown width predicted values in adjacent time periods:

[0116] ,

[0117] where T is the number of time periods, encourages the smoothness of the model output in the time dimension, avoiding jumps in crown width prediction caused by point cloud noise or short-term anomalies.

[0118] Furthermore, the growth of the tree crown is not only affected by its own structure but also closely related to external environmental conditions, such as light, rainfall, temperature, and soil humidity. To further improve the adaptability of the regression model, an environmental data feedback mechanism is integrated into the framework. Specifically, environmental variables (such as the annual average rainfall R t , the average temperature T temp，t ) are used as additional inputs and concatenated with the point cloud feature vector to form an extended feature vector . The regression model is updated to , where is a neural network model that integrates environmental data. The introduction of environmental data enables the model to perceive the ecological background in which the tree grows. For example, it predicts a smaller crown width expansion in dry years and a larger crown growth in wet years. This feedback mechanism dynamically adjusts the model output through external conditions, enhancing its response ability to environmental changes.

[0119] Furthermore, the optimized total loss function combines the original mean squared error and a temporal smoothing term:

[0120] ,

[0121] wherein, is the true crown width, and λ is the smoothing weight, which is used to balance precision and smoothness. This temporal constraint enables the model to learn the long-term variation law of the crown width and improves the consistency of multi-temporal data prediction.

[0122] Furthermore, based on the temporal smoothing term and environmental data feedback, the inversion framework realizes dynamic adaptive adjustment of the regression model parameters. During the training phase, the model learns the initial parameters by optimizing the total loss function ; while in the inference phase, for each period of data, the model dynamically updates the parameters according to the environmental variable and the predicted value of the previous period . For example, an adaptive learning rate or an online learning strategy can be introduced to adjust the model weights over time and environmental conditions:

[0123] ,

[0124] wherein, is the model parameter of the t-th period, η is the learning rate, is the loss gradient. This dynamic adjustment mechanism enables the model to adapt to new data in real time and overcomes the performance degradation problem of static models in long-term monitoring. In addition, if there are significant changes in the environmental data (such as abnormal drought), the model can quickly correct the prediction deviation through the feedback mechanism to ensure the continuous accuracy of crown width estimation.

[0125] In a specific embodiment of the present invention, the inversion steps of the tree height are specifically as follows:

[0126] Perform tree species classification or tree shape recognition on the point cloud subset , and output the tree species through the classification model;

[0127] Based on the tree species select a curve to match the crown shape; the types of the curve include parabola, high-order polynomial or spline curve;

[0128] Project onto the XZ plane to obtain a point set ;

[0129] Generate a crown height distribution model by fitting the curve using the least squares method, and the optimization objective is:

[0130] ,

[0131] Among them, , represents the number of points in is the height of the projection point;

[0132] Calculating the tree height is expressed as:

[0133] ,

[0134] Among them, is the ground height.

[0135] It should be noted that the first step of tree height inversion is to classify the point cloud subset of individual trees or identify the tree shape to determine the biological characteristics of each tree. Specifically, through the classification model outputs the tree species identification from the point cloud subset . The classification model can adopt a deep learning network (such as PointNet or a random forest classifier), and its input is the geometric features (such as height distribution, point density) and spectral features (such as multi-spectral band values) of , and the output is the tree species category. The purpose of tree species classification is to provide a basis for subsequent curve selection, because the tree shapes of different tree species (such as conical, circular or irregular) have a significant impact on the height distribution. By identifying the tree species or tree shape, this step can provide a personalized modeling basis for tree height estimation, significantly improving the adaptability and accuracy compared with the traditional inversion method that uniformly assumes the tree shape.

[0136] Furthermore, after determining the tree species , select the appropriate curve type according to its characteristics to adapt to the tree species morphology and improve the fitting accuracy. The curve types include various forms such as parabolas, high-order polynomials or spline curves, and the specific selection depends on the typical tree shape characteristics of the tree species. For example, for coniferous trees (such as pine trees), their tree shapes are usually conical and suitable for fitting with parabolas (such as ); for broad-leaved trees (such as oak trees), the tree crowns may be wider and more irregular, and higher-order polynomials (such as cubic or quartic polynomials) or spline curves can be selected to capture the complex vertical profiles. Traditional tree height inversion methods often assume fixed geometric models (such as cylinders or simple straight lines) and cannot adapt to diverse tree shapes, while this step realizes adaptive tree height modeling through tree species-specific curve selection. This flexibility not only improves the fitting accuracy but also enhances the generality of the method in different forest types.

[0137] It should be noted that for curve fitting, the point cloud subset Project onto the XZ plane to generate a two-dimensional point set The projection process retains the horizontal position x and height z of the point cloud, ignores the y coordinate to simplify the calculation, and forms the side profile of the tree in the vertical direction. This projection method is based on the definition of tree height (i.e., the difference in height between the top of the tree crown and the ground), aiming to extract the height distribution characteristics of the point cloud. The projected point set provides direct two-dimensional data input for subsequent curve fitting, simplifying the conversion process from three-dimensional point cloud to one-dimensional height curve.

[0138] Furthermore, based on the projected point set , fit the selected curve type by the least squares method to generate a height distribution model and provide accurate tree height estimation. The core of the least squares method is to minimize the sum of the squared errors between the fitted curve and the observed points by adjusting the curve parameters (such as a, b, c of a parabola). This process can effectively smooth the noise or discreteness in the point cloud data and generate a continuous tree height profile curve. Compared with the simple estimation of directly taking the highest point as the tree height in traditional methods, this step captures the overall height trend of the tree crown through curve fitting, avoids the interference of local outliers on the results, and improves the reliability of tree height estimation.

[0139] It should be noted that the tree height is calculated by the difference between the vertex of the curve and the ground height, accurately reflecting the actual height of a single tree. This curve-based definition of tree height not only considers the overall shape of the tree crown but also eliminates the influence of terrain undulation through correction. Compared with the rough estimation that only relies on a single highest point in traditional methods, this method calculates the tree height through the maximum value of the curve, which can better handle the situation of missing or unevenly distributed top points of the tree crown.

[0140] Preferably, in a specific embodiment of the present invention, an adaptive adjustment mechanism for growth model parameters is further embedded in the tree height inversion to optimize the fitting effect.

[0141] It should be noted that the goal of tree height inversion is to accurately estimate the height of a tree from the subset of single-tree point clouds, usually achieved by fitting curves (such as parabolas or high-order polynomials). However, traditional static fitting methods often assume that the tree height distribution follows a fixed model, ignoring the dynamic characteristics during the tree growth process and the differences under different tree species and environmental conditions. To improve the fitting effect and enhance the adaptability of the inversion method, an adaptive adjustment mechanism for growth model parameters is further embedded in the tree height inversion. This mechanism optimizes the accuracy and robustness of curve fitting by introducing the parameters of the tree growth model and dynamically adjusting these parameters in combination with the point cloud data and external information. Such a design not only improves the accuracy of tree height estimation but also provides a more reliable basis for dynamic trend analysis of multi-temporal data.

[0142] Furthermore, in tree height inversion, growth model parameters are embedded in the curve fitting process to describe the law of tree height change over time or environment. Traditional tree height fitting directly uses geometric curves to optimize the point cloud projection point set while this mechanism further introduces growth model parameters such as growth rate and maximum height potential . These parameters are derived from the biological characteristics of tree growth, and their basic form is:

[0143] ,

[0144] where is the tree height, is the tree age, is the initial time, is the growth rate. To adapt to point cloud data, the growth model is combined with geometric curves, and the adjusted fitting function can be expressed as:

[0145] ,

[0146] where is the spatial distribution function (such as a parabola), and the growth parameters dynamically regulate the overall scale of the tree height.

[0147] Furthermore, the adaptive adjustment mechanism of growth model parameters is dynamically optimized through point cloud features and external data , and other parameters. Specifically, first, feature vectors are extracted from the point cloud subset , including height statistical values (such as maximum height, distribution variance) and spectral features (such as canopy density index); at the same time, external inputs are combined, such as environmental data and tree species information . These information predict the initial growth parameters through regression models or rule mappings, for example:

[0148] ,

[0149] ,

[0150] where can be a deep learning-based artificial neural network model or a mathematical expression based on domain knowledge and data. The adaptive adjustment mechanism updates the parameters according to the current point cloud and environmental conditions during each fitting to ensure can reflect the actual growth state of the tree. For example, under drought conditions, may decrease, reflecting the slowdown of the growth rate.

[0151] Furthermore, based on the adaptively adjusted growth parameters, the least squares method is still used as the optimization objective, but the fitting function is updated to a dynamic form:

[0152] ,

[0153] wherein, incorporates the growth model parameters. The optimization process adaptively adjusts the and parameters through iteration to make the curve better fit the point cloud distribution. The final tree height is calculated as:

[0154] ,

[0155] The adaptability of the growth parameters enables the fitting curve to dynamically adapt to the growth stage of the tree and environmental changes, such as the rapid growth in the sapling stage or the height saturation of mature trees, significantly improving the accuracy of the fitting effect.

[0156] In a specific embodiment of the present invention, the growth parameters of individual trees are analyzed based on the inversion results of crown width and tree height. Specifically,

[0157] spatiotemporal matching is performed based on the multi-period individual tree segmentation results. Among them, the set of individual trees in the t-th period is ;

[0158] wherein, is the total number of individual trees in the t-th period, and each individual tree corresponds to the point cloud subset ;

[0159] Calculate the center point of the individual tree;

[0160] The spatiotemporal matching formula is defined as: if , then , it is determined as a successful match, otherwise 0; wherein, is the individual tree number in the (t + 1)-th period, is the distance threshold;

[0161] If the match is successful, calculate the tree height growth rate , and the crown width growth rate ;

[0162] wherein, is the tree height inversion result, is the crown width inversion result, is the time interval.

[0163] It should be noted that the core objective of dynamic analysis is to monitor the growth changes of individual trees through multi-period data. The first step is to perform spatio-temporal matching based on the multi-period individual tree segmentation results to identify the corresponding relationships of the same tree in different time periods. The process of spatio-temporal matching aims to determine which individual trees belong to the same tree in the time series from multi-period data, providing a basis for subsequent growth parameter calculation. Traditional methods may rely on manual marking or simple spatial proximity judgment, while this step realizes automated and efficient individual tree tracking through a systematic center point matching algorithm, which is particularly suitable for large-scale forest monitoring scenarios.

[0164] Furthermore, to achieve spatio-temporal matching, first, the center point of each individual tree is calculated as a representative of its spatial position. The calculation of the center point comprehensively considers the spatial distribution of the point cloud and can effectively characterize the position characteristics of the individual tree. Even if the crown shape is irregular or the point cloud density is uneven, it can provide a stable reference point.

[0165] Furthermore, based on the individual tree center points, a spatio-temporal matching formula is used to judge whether the individual tree in the t-th period corresponds to the individual tree in the (t + 1)-th period as the same tree. Among them, is a preset distance threshold. By traversing all the individual trees in the (t + 1)-th period , find the one that is closest to and less than to determine the matching relationship. If the distance is less than , then it is determined that the matching is successful (match = 1), indicating that and are the same tree; otherwise, it is determined that the matching fails (match = 0). The selection of needs to be determined according to the tree density and growth rate of the study area. For example, in sparse forest areas, it can be set to a smaller value (such as 1 - 2 meters), while in dense forest areas, it can be appropriately relaxed (such as 3 - 5 meters) to balance the accuracy and fault tolerance of the matching. This matching algorithm realizes the efficient identification of individual tree identities by quantifying the spatial distance and combining the spatio-temporal continuity of multi-period data, avoiding complex morphological comparison or manual correction in traditional methods.

[0166] Preferably, in a specific embodiment of the present invention, a closed-loop feedback mechanism is further introduced in the dynamic analysis to adjust the spatio-temporal matching and inversion parameters in real time according to historical monitoring data to improve the overall monitoring accuracy.

[0167] It should be noted that to improve the overall monitoring accuracy, a closed-loop feedback mechanism is further introduced to adjust the spatio-temporal matching and inversion parameters in real time using historical monitoring data. This mechanism optimizes the parameter settings of the matching threshold and inversion model by constructing a dynamic feedback loop, taking the previous monitoring results as input and feeding them back to the current analysis process. Such a design not only improves the accuracy of individual tree tracking but also enhances the reliability of parameter estimation such as tree height and crown width, providing more accurate technical support for long-term forest dynamics monitoring.

[0168] Furthermore, the core of the closed-loop feedback mechanism is to use historical monitoring data (such as the individual tree matching results and inversion parameters of previous periods) as the basis for real-time adjustment, forming a cycle of "monitoring - feedback - optimization". Specifically, assuming that the set of individual trees in the t-th period is , the tree height , crown width and matching relationship are obtained through spatio-temporal matching and inversion calculations. These results are stored as historical data and, together with the current input point cloud , are input into the feedback model. The feedback model evaluates the effectiveness of the current parameters based on the regularity of historical data (such as growth rate trends and matching success rates) and dynamically adjusts the distance threshold of spatio-temporal matching and the weight of the inversion model. Through continuous learning of historical information, the closed-loop feedback enables the system to adaptively respond to tree growth and environmental changes, enhancing the continuity and consistency of monitoring.

[0169] Furthermore, in the traditional method where spatio-temporal matching relies on the calculation of the distance from the center point, is a fixed value and it is difficult to adapt to the differences in tree density or growth rate in different forest areas. By using the closed-loop feedback mechanism to dynamically adjust with historical matching data, the matching success rate is defined, and the spatial movement range of the tree is estimated based on the historical growth rate . The adjustment formula can be expressed as:

[0170] ,

[0171] where is the adjustment coefficient and is the penalty factor. If the historical matching success rate is low, is appropriately relaxed to increase the matching rate; if the growth rate is large, is increased to adapt to the shift of the center point caused by the change in tree height. This real-time adjustment mechanism enables spatio-temporal matching to dynamically adapt to the actual situation of the monitoring area, reducing the phenomena of mis-matching or missed-matching.

[0172] Further, the inversion parameters (such as and of the tree height fitting curve or of the crown width regression model) are also optimized through a closed-loop feedback mechanism. Historical monitoring data provides a time series of inversion results (such as ), and the feedback model analyzes its trends and errors to dynamically update the current parameters. For example, for tree height inversion, the growth parameters can be adjusted according to historical tree heights:

[0173] ,

[0174] where is the predicted tree height at the t-th period, and is the learning rate. If the historical prediction is too low, is increased to correct the bias. Similarly, the weight of the crown width regression model is updated through online learning:

[0175] ,

[0176] This parameter optimization uses historical errors to correct the model in real time, ensuring that the inversion results are consistent with the actual growth trend and improving the accuracy of long-term monitoring.

[0177] Further, the closed-loop feedback mechanism forms a dynamically optimized monitoring system through spatio-temporal matching and coordinated adjustment of inversion parameters. After each monitoring is completed, the results are recorded and fed back to the next period for analysis, and parameters such as , , are iteratively optimized. The overall improvement in accuracy is reflected in: the success rate of spatio-temporal matching is increased, avoiding the interruption of single-tree tracking; the self-adaptability of inversion parameters is enhanced, reducing the systematic errors in tree height and crown width estimation. In multi-period monitoring, this mechanism can smooth the change curve of growth parameters, reduce the influence of noise, and provide more reliable calculation results of growth rates (such as , ).

[0178] In the second embodiment of the present invention, the present invention provides a single-tree dynamic monitoring system based on multi-period point clouds and deep learning. As shown in Figure 2 , the system includes a data acquisition module 1, a data fusion module 2, a segmentation module 3, and an analysis module 4;

[0179] The data acquisition module 1 is used to collect multi-period multi-spectral images and laser point cloud data of the research area and perform preprocessing;

[0180] The data fusion module 2 is used to register and fuse the preprocessed multi-period multi-spectral images and laser point cloud data to generate multi-period fused point clouds;

[0181] The segmentation module 3 is used to perform individual tree segmentation on the basis of the multi-temporal fused point cloud using a deep learning model;

[0182] The analysis module 4 is used to invert the crown width and tree height of individual trees based on the individual tree segmentation results, and analyze the growth parameters of individual trees according to the inversion results of the crown width and tree height.

[0183] In summary, the present invention provides a method and system for dynamic monitoring of individual trees based on multi-temporal point clouds and deep learning. By fusing multi-temporal point cloud and multi-spectral image data and combining with a deep learning model, accurate segmentation and dynamic growth monitoring of individual trees are achieved, overcoming the defects of low accuracy of individual tree segmentation, insufficient fusion of multi-source data, and inability to capture the growth changes of forest trees in the prior art. The use of the inversion models for crown width and tree height effectively improves the accuracy of growth parameter estimation, providing a scientific basis for forest tree growth monitoring and forest resource management.

[0184] It should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative manner of the specification is only for clarity. Those skilled in the art should regard 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.

[0185] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described modules can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0186] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0187] In addition, the functional modules in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware, or in the form of a combination of hardware and software functional modules.

[0188] The integrated modules implemented in the form of software functional modules can be stored in a computer-readable storage medium. The above-mentioned software functional modules are stored in a storage medium and include several instructions for causing a computer system (which can be a personal computer, a server, or a network system, etc.) or a processor to execute some steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of the present application.

Claims

1. A single tree dynamic monitoring method based on multi-period point cloud and deep learning, characterized by: include, Collect multi-period multispectral images and laser point cloud data of the study area and perform pre-processing; The pre-processed multi-period multispectral images are registered and fused with the laser point cloud data to generate a multi-period fused point cloud; The registration and fusion step comprises: Convert the preprocessed multispectral image into a multispectral image point cloud and standardize the scale of each point cloud; Extracting feature lines from the multispectral image point cloud and the laser point cloud respectively; the feature lines include edge feature lines representing tree boundary lines and corner feature lines representing structural key points; Based on the extracted feature lines, registration and fusion are performed by calculating the spatial correspondence between the feature lines to generate a multi-phase fused point cloud; the multi-phase fused point cloud is a comprehensive point cloud data set containing multi-phase three-dimensional spatial information and spectral features; Based on the multi-phase fused point cloud, a deep learning model is used to perform single tree segmentation; The single wood splitting step comprises: Based on multi-phase fusion point clouds, a deep learning model is used to segment individual trees and output a point cloud subset for each individual tree. ,in, is a single point in the point cloud, Indicates the single tree number, Represents the segmentation label of the i-th point; The use of a deep learning model to perform single tree segmentation includes: Use PointNet++ model to fusion point cloud Segmentation is performed, where each point , is the spatial coordinate, is the spectral feature; Based on the single tree segmentation result, the crown width and tree height of the single tree are inverted, and the growth parameters of the single tree are analyzed according to the inversion results of the crown width and tree height; The growth parameters of a single tree analyzed based on the inversion results of crown width and tree height include: Based on the multi-period single tree segmentation results, time-space matching is performed, where the single tree set in period t is ; in, is the total number of trees in period t, and each tree Corresponding point cloud subset ; Calculate the center point of a single tree ; The time-space matching formula is defined as: ,but , it is determined to be a successful match, otherwise 0; among them, is the single tree number of the t+1 period, is the distance threshold; If the match is successful, calculate the tree height growth rate , and crown growth rate ; in, is the tree height inversion result, is the crown width inversion result, For the time interval.

2. The single tree dynamic monitoring method based on multi-period point cloud and deep learning according to claim 1 is characterized by: The crown width inversion step includes: Through the feature extraction function from Extract feature vectors from ; Calculate crown width ,in is the neural network model adopted; When training the neural network model, optimize the mean square error loss ; in, is the total number of single trees, The actual crown width.

3. The single tree dynamic monitoring method based on multi-period point cloud and deep learning according to claim 2 is characterized by: The crown width inversion step further comprises: The time series smoothing term is defined by regularization to capture the smooth trend of crown amplitude change, which is expressed as: in, is the predicted value of the crown width of the kth tree in the tth period, is the point cloud subset of the kth tree in the tth period The characteristic vector of , T is the time period; Use environment variables with Concatenate to form an extended feature vector and calculate the crown width using a neural network model that integrates environmental data ;in, is the average annual rainfall, is the average temperature; The dynamic adjustment of regression model parameters is achieved by optimizing the total loss function, which is expressed as: in, is the true crown width, and λ is the smoothing weight coefficient.

4. The single tree dynamic monitoring method based on multi-period point cloud and deep learning according to claim 1 is characterized by: The tree height inversion step includes: Point cloud subset Perform tree species classification or tree shape recognition, and output tree species through classification models ; Based on tree species Select Curve Matching the crown shape; the type of the curve includes a parabola, a high-order polynomial or a spline curve; Will Project to the XZ plane and get the point set ; The curve was fitted by the least squares method Generate a crown height distribution model with the optimization goal as follows: in, ,express The number of points in is the height of the projection point; The calculation of tree height is expressed as: in, is the ground height.

5. The single tree dynamic monitoring method based on multi-period point cloud and deep learning according to claim 4 is characterized by: The tree height inversion step further comprises, The fitting effect is optimized by embedding the adaptive adjustment mechanism of the growth model parameters, including: Introduced in tree height fitting and maximum height potential , adjust the fitting function to: in, For tree age, is the spatial distribution function; From point cloud subset Extract feature vectors from , and combined with external environmental data and tree species information, growth parameters are calculated through regression models or rule mapping: in, For prediction models based on neural networks or empirical formulas, the parameters are adaptively updated according to the current point cloud and environmental conditions; Based on the adjusted , optimize the target by the least square method, generate a dynamic crown height distribution model, and calculate the tree height; The optimization objective is expressed as: 。 6. The single tree dynamic monitoring method based on multi-period point cloud and deep learning according to claim 1 is characterized by: The analysis of the growth parameters of a single tree based on the inversion results of the crown width and the tree height further includes: By introducing a closed-loop feedback mechanism to improve detection accuracy, specifically including: Record the single wood collection of the previous issues , matching results , tree height and crown width As historical data; Calculate the matching success rate based on historical data and historical growth rate Adjust the distance threshold: in, is the adjustment coefficient, is a penalty factor to adapt to changes in tree growth and forest area characteristics; Use historical tree height sequences and crown width sequences to dynamically adjust inversion parameters, and recalculate based on the adjusted parameters and ; The matching results and inversion parameters of the current period are fed back to the next analysis period to realize a closed-loop process of monitoring, feedback and optimization, thereby improving the success rate of spatiotemporal matching and the accuracy of growth rate calculation.

7. A single tree dynamic monitoring system based on multi-period point cloud and deep learning, applied to the single tree dynamic monitoring method based on multi-period point cloud and deep learning as claimed in claim 1, characterized in that: It includes data acquisition module, data fusion module, segmentation module and analysis module; The data acquisition module is used to collect multi-period multispectral images and laser point cloud data of the research area and perform preprocessing; The data fusion module is used to register and fuse the preprocessed multi-period multispectral images with the laser point cloud data to generate a multi-period fused point cloud; The segmentation module is used to perform single tree segmentation using a deep learning model based on the multi-phase fused point cloud; The analysis module is used to invert the crown width and tree height of the single tree based on the single tree segmentation result, and analyze the growth parameters of the single tree according to the inversion results of the crown width and tree height.

Images