Point cloud tree species classification method and system based on airborne LiDAR
Through the deep supervision network DSTCN, feature extraction and attention assignment of single wood slices is solved, and the characteristic noise problem at different vertical height levels in three-dimensional laser point clouds is improved, which is better than traditional methods.
Patent Information
- Application Number
- CN202310398248.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-14
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-04-14
AI Technical Summary
The existing tree species classification method based on three-dimensional laser point clouds is highly characterized due to factors such as terrain fluctuations and scanning conditions at different vertical heights, which limits the classification accuracy. It is easy to produce category imbalances in the whole feature extraction unit directly using single wood as the feature extraction unit. question.
A deep supervision network DSTCN is proposed. By extracting histogram feature descriptors on single wood slices, and combining the height and intensity information of the slices, different attention is given to different slice features, and slice features with different information gains are used to solve feature noise. Problem, improve classification accuracy.
Effectively improve the overall accuracy of tree species classification. Compared with PointNet++ and SVM methods, the macro-average F-score, Kappa coefficient and lower standard deviation of classification accuracy in multiple experiments, the stability is better, especially in a small number of Excellent performance in the training sample.
Smart Images

Figure CN116597199B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of airborne LiDAR point cloud classification, and in particular relates to a point cloud tree species classification method and system based on airborne LiDAR. Background Art
[0002] As one of the most important natural resources on earth, forest resources are also the largest ecosystem with the largest biomass on land. They play an important role in conserving water, preventing wind and sand, and regulating climate. The identification and classification of tree species are of great significance for counting forest resources, exploring the diversity of ecosystems, and even studying the evolution of organisms. Therefore, the accurate identification and statistics of tree species in forest areas are becoming more and more important. With the rapid development of digital information technology in recent years, digital forestry based on remote sensing technology has provided great convenience for the sustainable management of forest resources. Aerial photography, as a non-contact remote sensing technology, can quickly obtain the horizontal structure of forests in a large area. In the early days, tree species identification was mostly completed by experts' visual interpretation of aerial photography images combined with on-site manual inventory, but the high time and financial costs made it impossible to be widely used. Hyperspectral images are detectable for many substances that cannot be detected in optical images obtained in wide bands. In high-resolution hyperspectral images, the radiation in different wavelength regions can well reflect the biological and non-biological characteristics of tree groups. This advantage has fully developed the potential of hyperspectral images in classifying forest types and drawing forest phase maps. However, as we further study the problems in the experiment, the limitations of the hyperspectral spectrum are gradually becoming apparent. Different tree species of the same family and genus may have very similar spectral characteristics; or because the forest canopy density is too high, excessive shading may cause the spectral characteristics of samples within the species to differ greatly due to differences in crown shape. These problems will significantly reduce the classification accuracy.
[0003] In the study of high-resolution images, researchers noticed the potential of canopy contour characterization parameters in tree species classification. However, since images can only depict the horizontal distribution of forest stands, they cannot provide profile information. As an advanced active remote sensing technology, LiDAR technology has a good ability to describe the structure of the canopy. In particular, airborne LiDAR has shown great potential in forestry-related surveying and mapping. Over the years, with the hard work of many scholars, more and more characterization parameters have been proposed to help tree species classification, which are mainly divided into two categories: geometric features and radiation features. Geometric features usually include canopy height information, density information, structural scale parameters, etc., while radiation features mainly include single / multi-channel reflection intensity information, echo type features, etc. And with the development of deep learning technology, tree species classification methods based on neural network extraction of deep features are becoming more and more common. In particular, the PointNet series, as a neural network that directly processes point clouds, can provide significant accuracy gains in tree species classification.
[0004] Among the existing algorithms for tree species classification based on three-dimensional laser point clouds, there are mainly deep learning methods that directly input point clouds into neural network classification, and feature engineering methods that extract feature vectors and then classify based on traditional machine learning. From the perspective of the classifier, the former is limited by the consistency of the input dimension and requires downsampling of individual trees, which not only loses individual tree features, but also makes it difficult to determine the number of samples for data sets with large differences in individual tree morphology and point cloud density. At the same time, point clouds are implicit features, and direct feature extraction processing of point clouds often places a large operating burden on the network. The latter is limited by the performance of the classifier and cannot extract deeper features. From the perspective of samples, the above methods extract features from individual tree samples at the overall level, without considering that during the airborne LiDAR scanning process, the feature noise of intra-species samples at different vertical height levels may be increased due to multiple factors such as undulating terrain, different laser scanning angles, and changes in flight altitude. In addition, in field applications, the structure and physical properties of the canopy between the same tree species may also have high heterogeneity due to many environmental factors such as undulating terrain and seasonal changes. When extracting classification features from laser point clouds, directly taking a single tree as the feature extraction unit is prone to class imbalance problems. Summary of the invention
[0005] The present invention aims to solve the problem that in practical applications, the same sample is subject to large feature noise at different vertical height levels due to factors such as scanning conditions and geographical environment, thereby limiting the classification accuracy. A deep supervised network DSTCN (Deeply Supervised Tree Classification Network) for tree species classification is proposed. DSTCN accepts the histogram feature descriptor of a single wood slice as an input vector, and combines the height and intensity information of the slice to give different attention to the features of different single wood slices, so that the slice features with different information gains can be used more reasonably, which can effectively solve the accuracy limitation caused by the feature noise at different vertical height levels, and achieve the purpose of improving the overall classification accuracy.
[0006] To achieve the above object, the present invention provides the following scheme: a point cloud tree species classification method based on airborne LiDAR, comprising the following steps:
[0007] S1, performing ground filtering on the original point cloud to be processed to obtain ground points in the target area; removing the ground points to obtain a non-ground point cloud, and performing elevation normalization processing on the non-ground point cloud to obtain a forest point cloud;
[0008] S2. Segment the tree point cloud to obtain single tree point cloud data; and construct a training data set and a test data set based on the single tree point cloud data;
[0009] S3, slicing the single wood point cloud data, and obtaining a feature vector based on the sliced single wood point cloud data;
[0010] S4, constructing a classification neural network, and using the training data set to train the classification neural network, and classifying the test data set by tree species based on the trained classification neural network to obtain a classification result.
[0011] Preferably, the method of obtaining the ground point of the target area in S1 includes:
[0012] rasterizing the target area;
[0013] Select the lowest point of each grid as a seed point, and construct an initial triangulated network based on the seed point;
[0014] Calculate the angle and distance between each of the original point clouds and the vertex of the initial triangulated network;
[0015] Setting an iteration angle threshold and an iteration distance threshold, taking the original point cloud with the angle and the distance satisfying the iteration angle threshold and the iteration distance threshold as ground points, and updating the triangulation network;
[0016] Repeat the operation to obtain all ground points in the target area.
[0017] Preferably, S3 includes:
[0018] Slicing the single wood point cloud data;
[0019] Construct a spherical neighborhood space for each point cloud, and build a covariance matrix through the neighborhood point set to obtain the eigenvalue;
[0020] and calculating the degree value, the number of neighborhood points and the mean value of reflection intensity of the local point-like, line-like and spherical structures based on the characteristic values;
[0021] Constructing a histogram feature descriptor based on the degree value of the local obedience to the point-like, line-like and spherical structures, the number of neighborhood points and the mean value of the reflection intensity;
[0022] The feature vector is obtained based on the histogram feature descriptor.
[0023] Preferably, the S4 includes:
[0024] Constructing the classification neural network;
[0025] Extract the average reflection intensity and average height of each slice, and construct intensity series and height series;
[0026] The intensity sequence, the height sequence and the feature vector are input into the classification neural network to obtain a classification result.
[0027] Preferably, the classification neural network comprises:
[0028] Sequence weighting module, weighted encoding structure and deep feedforward structure;
[0029] The sequence weighting module includes: DFF, Softmax weight normalization function and Concatenate feature aggregation unit;
[0030] The weighted coding structure includes: a multi-head attention module, a DFF and a LayerNorm layer normalization unit;
[0031] The deep feedforward structure includes: a Linear fully connected layer, a Mish activation function, and a BatchNorm batch normalization unit.
[0032] The present invention also provides a point cloud tree species classification system based on airborne LiDAR, comprising: a filtering unit, a segmentation unit, a slicing unit and a classification unit;
[0033] The filtering unit is used to perform ground filtering on the original point cloud to be processed to obtain ground points in the target area; remove the ground points to obtain a non-ground point cloud, and perform elevation normalization on the non-ground point cloud to obtain a forest point cloud;
[0034] The segmentation unit is used to segment the forest point cloud to obtain single tree point cloud data; based on the single tree point cloud data, a training data set and a test data set are constructed;
[0035] The slicing unit is used to slice the single wood point cloud data, and obtain a feature vector based on the sliced single wood point cloud data;
[0036] The classification unit is used to construct a classification neural network, and use the training data set to train the classification neural network, and classify the test data set into tree species based on the trained classification neural network to obtain a classification result.
[0037] Preferably, the method by which the filtering unit obtains the ground point comprises:
[0038] rasterizing the target area;
[0039] Select the lowest point of each grid as a seed point, and construct an initial triangulated network based on the seed point;
[0040] Calculate the angle and distance between each of the original point clouds and the vertex of the initial triangulated network;
[0041] Setting an iteration angle threshold and an iteration distance threshold, taking the original point cloud with the angle and the distance satisfying the iteration angle threshold and the iteration distance threshold as ground points, and updating the triangulation network;
[0042] Repeat the operation to obtain all ground points in the target area.
[0043] Preferably, the method for the slicing unit to obtain the feature vector includes:
[0044] Slicing the single wood point cloud data;
[0045] Construct a spherical neighborhood space for each point cloud, and build a covariance matrix through the neighborhood point set to obtain the eigenvalue;
[0046] and calculating the degree value, the number of neighborhood points and the mean value of reflection intensity of the local point-like, line-like and spherical structures based on the characteristic values;
[0047] Constructing a histogram feature descriptor based on the degree value of the local obedience to the point-like, line-like and spherical structures, the number of neighborhood points and the mean value of the reflection intensity;
[0048] The feature vector is obtained based on the histogram feature descriptor.
[0049] Preferably, the method by which the classification unit obtains the classification result includes:
[0050] Constructing the classification neural network;
[0051] Extract the average reflection intensity and average height of each slice, and construct intensity series and height series;
[0052] The intensity sequence, the height sequence and the feature vector are input into the classification neural network to obtain a classification result.
[0053] Preferably, the classification neural network comprises:
[0054] Sequence weighting module, weighted encoding structure and deep feedforward structure;
[0055] The sequence weighting module includes: DFF, Softmax weight normalization function and Concatenate feature aggregation unit;
[0056] The weighted coding structure includes: a multi-head attention module, a DFF and a LayerNorm layer normalization unit;
[0057] The deep feedforward structure includes: Linear fully connected layer, Mish activation function and BatchNorm batch normalization unit
[0058] Compared with the prior art, the present invention has the following beneficial effects:
[0059] In the present invention, DSTCN combines the height and intensity information of the slices, gives different attention to the characteristics of different single wood slices, makes slice features with different information gains more reasonably utilized, and can effectively solve the precision limitation caused by the characteristic noise at different vertical height levels. Compared with two commonly used methods based on deep network Point Net++ and traditional machine learning SVM, the method of the present invention is 4-14 percentage points higher than Point Net++ in macro average F-score (MacroAverage F-score, MAF) and 5-10 percentage points higher than SVM in multiple experimental results; Kappa coefficient is 6-18 percentage points higher than PointNet++ and 7-13 percentage points higher than SVM; the classification accuracy standard deviation (Precision Interclass Standard Deviation, PISD), recall standard deviation (Recall InterclassStandard Deviation, RISD) and F-score standard deviation (F-score Interclass StandardDeviation, FISD) between tree species have the only minimum value. As the training samples continue to decrease, compared with the control method, the method provided by the present invention always maintains a higher accuracy. Through the analysis of qualitative experimental results, it is proved that the method provided by the present invention has the ability to mine deep feature differences from a small number of training samples, and has high superiority in resisting intra-species sample feature noise and increasing inter-species sample feature variance. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0061] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention;
[0062] Figure 2 This is a schematic diagram of feature extraction of a single tree sample according to an embodiment of the present invention;
[0063] Figure 3 A schematic diagram of obtaining single-point features from a single wood sample according to an embodiment of the present invention;
[0064] Figure 4 This is a schematic diagram of the deep supervision network DSTCN architecture of an embodiment of the present invention;
[0065] Figure 4 (a) is the DSTCN architecture; Figure 4 (b) is the SWM structure; Figure 4 (c) is the Transformer Encoder structure; Figure 4 (d) is the DFF structure;
[0066] Figure 5 This is a statistical diagram of experimental tree species parameter information in an embodiment of the present invention;
[0067] Figure 5 (a) is the tree height; Figure 5 (b) is the crown width; Figure 5 (c) is the average density (density value is points per cubic meter); Figure 5 (d) is the average reflection intensity;
[0068] Figure 6 Schematic diagram of the system structure of an embodiment of the present invention. DETAILED DESCRIPTION
[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0070] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0071] Embodiment 1
[0072] like Figure 1 As shown, it is a schematic diagram of the process flow of the airborne LiDAR point cloud tree species classification method of this embodiment, which includes the following steps:
[0073] S1, perform ground filtering on the original point cloud to be processed to obtain the ground points of the target area; remove the ground points to obtain the non-ground point cloud, perform elevation normalization on the non-ground point cloud to obtain the forest point cloud;
[0074] In this embodiment, a progressive irregular triangulated network ground filtering algorithm is used to filter the ground. Specifically,
[0075] First, the target area is rasterized at a resolution of 2 m;
[0076] The morphological opening operation is used to select the original point cloud with the lowest point in each grid as the seed point, and the initial triangulation of the entire target area is constructed based on the seed point;
[0077] Then, the angle α and distance d formed between each original point cloud and the vertex of the nearest triangulation network are calculated;
[0078] The iteration angle threshold is set to 10° and the iteration distance threshold is set to 1.2m. If α and d meet the above thresholds, the original point cloud is used as the ground point and the triangulation network is updated;
[0079] Repeat the above steps until all ground points in the target area are obtained.
[0080] All the ground points obtained above are removed to obtain the non-ground point cloud, and then the non-ground point cloud is normalized to obtain the forest point cloud.
[0081] In this embodiment, a DEM is produced by Kriging interpolation, and the elevation is normalized by subtracting it from the non-ground point cloud. Specifically,
[0082] First, obtain n(n-1) / 2 distance values between ground points and sort them into n groups from small to large. The ground points x in each group j Distance values are all Distance i ;
[0083] Substitute the average distance into the semivariance formula to calculate the experimental variation value Semivariance corresponding to each group of distances i :
[0084]
[0085] Among them, Z(x j ) is the corresponding variation value, N(Distance i ) is the number of ground points in the corresponding group at this distance value, which is m.
[0086] Draw a distance-semivariance scatter plot and fit the variation curve, calculate the nugget value (Nugget), arch height value (Sill) and range value (Range), and construct a spherical fitting function:
[0087]
[0088] The DEM with a resolution of 1m was generated based on the interpolation of the spherical fitting function. The non-ground point cloud was rasterized with a resolution of 1m and subtracted from the DEM to obtain the forest point cloud.
[0089] S2. Segment the forest point cloud to obtain single tree point cloud data; construct a training data set and a test data set based on the single tree point cloud data;
[0090] In this embodiment, professional software is used to segment the forest point cloud to obtain a single tree point cloud, and sample optimization is performed to construct a training data set and a test data set.
[0091] S3, slicing the single wood point cloud data, and obtaining a feature vector based on the sliced single wood point cloud data;
[0092] The segmented single tree point cloud data is evenly sliced vertically, and the same feature engineering is performed on each slice to extract the feature vector. Specifically,
[0093] First, the single wood point cloud data is sliced into equal-step slices. In this embodiment, the slices are sliced into 20 slices.
[0094] Construct a spherical neighborhood space for each point cloud, and build a covariance matrix through the neighborhood point set to obtain the eigenvalues λ1, λ2, and λ3;
[0095] Based on the eigenvalues, the degree values DA1, DA2, DA3, and the number of neighborhood points I of the local point, line, and spherical structures are calculated. mean And the mean reflection intensity Density, such as Figure 3 As shown;
[0096] The calculation formulas for DA1, DA2, and DA3 are as follows:
[0097]
[0098] For each slice, a histogram feature descriptor is constructed based on the five features of the degree of local compliance with point, line, and spherical structures, the number of neighborhood points, and the mean reflection intensity;
[0099] In this embodiment, there are 128 statistical intervals of the histogram, and the statistical upper and lower limits are the upper and lower limits of the feature in the entire single tree. Each single tree sample has a total of 5×20 histogram feature descriptors.
[0100] All histogram descriptors are concatenated in one dimension to obtain the feature vector of a single tree. Figure 2 shown.
[0101] S4. Construct a classification neural network, and use the training data set to train the classification neural network, and classify the tree species of the test data set based on the trained classification neural network to obtain the classification results.
[0102] like Figure 5 As shown, the experimental tree species parameter information of this embodiment is shown. Specifically, Figure 5 (a) is the tree height; Figure 5 (b) is the crown width; Figure 5(c) is the average density (density value is points per cubic meter); Figure 5 (d) is the average reflection intensity.
[0103] First, a classification neural network is constructed; Figure 4 As shown, it is a schematic diagram of the classification neural network structure of this embodiment.
[0104] In this embodiment, the network architecture is mainly composed of a weighted encoding structure Weighted Encoder (WE) composed of a sequence weighted module (SWM) and a Transformer Encoder (TE) connected by three multilayer perceptrons (MLP), and a deep feed forward structure (DFF), such as Figure 4 (a). DFF contains a Linear fully connected layer, a Mish activation function, and a BatchNorm batch normalization unit, as shown in Figure 4 (d) As shown. SWM includes DFF, Softmax weight normalization function and Concatenate feature aggregation unit, as shown Figure 4 (b) As shown in Figure 2, TE includes a multi-head attention module (Multi-headAttention, MHA), DFF, and LayerNorm layer normalization unit, as shown in Figure 2. Figure 4 (c) As shown. Two WEs are sequentially added with strength and height attention in a cascade manner, and multi-scale attention is allocated to features through a combination of SWM and TE in WE. The backbone classifier Classifier is constructed by DFF, Softmax and cross-entropy loss function, and an auxiliary branch model consisting of DFF, BatchNorm and Classifier is added after the first WE to form intermediate supervision (IS) for the backbone network.
[0105] Then, the average reflection intensity and average height of each slice are extracted to construct the intensity sequence Intensity and the height sequence Height;
[0106] The intensity sequence, height sequence and feature vector are input into the classification neural network to obtain the classification result.
[0107] In SWM, DFF is used to analyze the sequence constructed by slice information, and Softmax normalization is used to generate slice weights, and attention is allocated to different slices at the slice level. Attention is then allocated at the feature level through Transformer Encoder. The input vector is weighted encoded using the Height sequence and Intensity sequence through the above steps, and passed through a classifier composed of DFF, Softmax, and Cross-entropy Loss. DFF further extracts features, Softmax outputs category probabilities, and combines probabilities with labels, and calculates classification losses through Cross-entropy Loss. At the same time, the classifier structure is added to the middle hidden layer as relay supervision to optimize the gradient flow.
[0108] Embodiment 2
[0109] like Figure 6 As shown, this embodiment provides a point cloud tree species classification system based on airborne LiDAR, including: a filtering unit, a segmentation unit, a slicing unit and a classification unit;
[0110] The filtering unit is used to perform ground filtering on the original point cloud to be processed to obtain the ground points of the target area; remove the ground points to obtain the non-ground point cloud, perform elevation normalization on the non-ground point cloud to obtain the forest point cloud;
[0111] In this embodiment, the filtering unit uses a progressive irregular triangulated network encrypted ground filtering algorithm to filter the ground. Specifically,
[0112] First, the target area is rasterized at a resolution of 2 m;
[0113] The morphological opening operation is used to select the original point cloud with the lowest point in each grid as the seed point, and the initial triangulation of the entire target area is constructed based on the seed point;
[0114] Then, the angle α and distance d formed between each original point cloud and the vertex of the nearest triangulation network are calculated;
[0115] The iteration angle threshold is set to 10° and the iteration distance threshold is set to 1.2m. If α and d meet the above thresholds, the original point cloud is used as the ground point and the triangulation network is updated;
[0116] Repeat the above steps until all ground points in the target area are obtained.
[0117] All the ground points obtained above are removed to obtain the non-ground point cloud, and then the non-ground point cloud is normalized to obtain the forest point cloud.
[0118] In this embodiment, the filtering unit uses Kriging interpolation to produce DEM, and performs elevation normalization by subtracting it from the non-ground point cloud.
[0119] First, obtain n(n-1) / 2 distance values between ground points and sort them into n groups from small to large. The ground points x in each group j Distance values are all Distance i ;
[0120] Substitute the average distance into the semivariance formula to calculate the experimental variation value Semivariance corresponding to each group of distances i :
[0121]
[0122] Among them, Z(x j ) is the corresponding variation value, N(Distance i ) is the number of ground points in the corresponding group at this distance value, which is m.
[0123] Draw a distance-semivariance scatter plot and fit the variation curve, calculate the nugget value (Nugget), arch height value (Sill) and range value (Range), and construct a spherical fitting function:
[0124]
[0125] The DEM with a resolution of 1m was generated based on the interpolation of the spherical fitting function. The non-ground point cloud was rasterized with a resolution of 1m and subtracted from the DEM to obtain the forest point cloud.
[0126] The segmentation unit is used to segment the forest point cloud to obtain single tree point cloud data; based on the single tree point cloud data, a training data set and a test data set are constructed;
[0127] In this embodiment, the segmentation unit uses professional software to segment the forest point cloud to obtain a single tree point cloud, and performs sample optimization to construct a training data set and a test data set.
[0128] The slicing unit is used to slice the single wood point cloud data, and obtain the feature vector based on the sliced single wood point cloud data;
[0129] The slicing unit slices the segmented single wood point cloud data uniformly in the vertical direction, and performs the same feature engineering on each slice to extract the feature vector. Specifically,
[0130] First, the single wood point cloud data is sliced into equal-step slices. In this embodiment, the slices are sliced into 20 slices.
[0131] Construct a spherical neighborhood space for each point cloud, and build a covariance matrix through the neighborhood point set to obtain the eigenvalues λ1, λ2, and λ3;
[0132] Based on the eigenvalues, the degree values DA1, DA2, DA3, and the number of neighborhood points I of the local point, line, and spherical structures are calculated. mean And the mean reflection intensity Density, such as Figure 3 As shown;
[0133] The calculation formulas for DA1, DA2, and DA3 are as follows:
[0134]
[0135] For each slice, a histogram feature descriptor is constructed based on the five features of the degree of local compliance with point, line, and spherical structures, the number of neighborhood points, and the mean reflection intensity;
[0136] In this embodiment, there are 128 statistical intervals of the histogram, and the statistical upper and lower limits are the upper and lower limits of the feature in the entire single tree. Each single tree sample has a total of 5×20 histogram feature descriptors.
[0137] All histogram descriptors are concatenated in one dimension to obtain the feature vector of a single tree.
[0138] The classification unit is used to construct a classification neural network, and the classification neural network is trained using a training data set, and the test data set is classified into tree species based on the trained classification neural network to obtain a classification result.
[0139] The methods for the classification unit to obtain classification results include:
[0140] First, a classification neural network is constructed; Figure 4 As shown, it is a schematic diagram of the classification neural network structure of this embodiment.
[0141] In this embodiment, the network architecture is mainly composed of a weighted encoding structure Weighted Encoder (WE) composed of a sequence weighted module (SWM) and a Transformer Encoder (TE) connected by three multilayer perceptrons (MLP), and a deep feed forward structure (DFF), such as Figure 4 (a). DFF contains a Linear fully connected layer, a Mish activation function, and a BatchNorm batch normalization unit, as shown in Figure 4(d) As shown. SWM includes DFF, Softmax weight normalization function and Concatenate feature aggregation unit, as shown Figure 4 (b) shows that TE includes a multi-head attention module (Multi-headAttention, MHA), DFF and LayerNorm layer normalization unit, as shown in Figure 4 (c) As shown. Two WEs are sequentially added in a cascade manner with strength and height attention, and multi-scale attention is allocated to features through a combination of SWM and TE in WE. The backbone classifier Classifier is constructed by DFF, Softmax, and cross-entropy loss function, and an auxiliary branch model consisting of DFF, BatchNorm, and Classifier is added after the first WE to form intermediate supervision (IS) for the backbone network.
[0142] Then, the average reflection intensity and average height of each slice are extracted to construct the intensity sequence Intensity and the height sequence Height;
[0143] The intensity sequence, height sequence and feature vector are input into the classification neural network to obtain the classification result.
[0144] In SWM, DFF is used to analyze the sequence constructed by slice information, and Softmax normalization is used to generate slice weights, and attention is allocated to different slices at the slice level. Attention is then allocated at the feature level through Transformer Encoder. The input vector is weighted encoded using the Height sequence and Intensity sequence through the above steps, and passed through a classifier composed of DFF, Softmax, and Cross-entropy Loss. DFF further extracts features, Softmax outputs category probabilities, and combines probabilities with labels, and calculates classification losses through Cross-entropy Loss. At the same time, the classifier structure is added to the middle hidden layer as relay supervision to optimize the gradient flow.
[0145] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A point cloud tree species classification method based on airborne LiDAR, characterized in that: The following steps are involved: S1, performing ground filtering on the original point cloud to be processed to obtain ground points in the target area; removing the ground points to obtain a non-ground point cloud, and performing elevation normalization processing on the non-ground point cloud to obtain a forest point cloud; S2, segmenting the tree point cloud to obtain single tree point cloud data; Based on the single tree point cloud data, construct a training data set and a test data set; S3, slicing the single wood point cloud data, and obtaining a feature vector based on the sliced single wood point cloud data; The S3 includes: Slicing the single wood point cloud data; Construct a spherical neighborhood space for each point cloud, and build a covariance matrix through the neighborhood point set to obtain the eigenvalue; and calculating the degree value, the number of neighborhood points and the mean value of reflection intensity of the local point-like, line-like and spherical structures based on the characteristic values; Constructing a histogram feature descriptor based on the degree value of the local obedience to the point-like, line-like and spherical structures, the number of neighborhood points and the mean value of the reflection intensity; Obtaining the feature vector based on the histogram feature descriptor; S4, constructing a classification neural network, and using the training data set to train the classification neural network, and classifying the test data set by tree species based on the trained classification neural network to obtain a classification result; The S4 includes: Constructing the classification neural network; The classification neural network comprises: Sequence weighting module, weighted encoding structure and deep feedforward structure; The sequence weighting module includes: DFF, Softmax weight normalization function and Concatenate feature aggregation unit; The weighted coding structure includes: a multi-head attention module, a DFF and a LayerNorm layer normalization unit; The deep feedforward structure includes: a Linear fully connected layer, a Mish activation function, and a BatchNorm batch normalization unit; Extract the average reflection intensity and average height of each slice, and construct intensity series and height series; The intensity sequence, the height sequence and the feature vector are input into the classification neural network to obtain a classification result.
2. The point cloud tree species classification method based on airborne LiDAR according to claim 1, characterized in that: The method of obtaining the ground point of the target area in S1 includes: rasterizing the target area; Select the lowest point of each grid as a seed point, and construct an initial triangulated network based on the seed point; Calculate the angle and distance between each of the original point clouds and the vertex of the initial triangulated network; Setting an iteration angle threshold and an iteration distance threshold, taking the original point cloud with the angle and the distance satisfying the iteration angle threshold and the iteration distance threshold as ground points, and updating the triangulation network; Repeat the operation to obtain all ground points in the target area.
3. The point cloud tree species classification system based on airborne LiDAR is characterized by: include: Filtering unit, segmentation unit, slicing unit and classification unit; The filtering unit is used to perform ground filtering on the original point cloud to be processed to obtain ground points in the target area; remove the ground points to obtain a non-ground point cloud, and perform elevation normalization on the non-ground point cloud to obtain a forest point cloud; The segmentation unit is used to segment the forest point cloud to obtain single tree point cloud data; Based on the single tree point cloud data, construct a training data set and a test data set; The slicing unit is used to slice the single wood point cloud data, and obtain a feature vector based on the sliced single wood point cloud data; The method for the slicing unit to obtain the feature vector includes: Slicing the single wood point cloud data; Construct a spherical neighborhood space for each point cloud, and build a covariance matrix through the neighborhood point set to obtain the eigenvalue; and calculating the degree value, the number of neighborhood points and the mean value of reflection intensity of the local point-like, line-like and spherical structures based on the characteristic values; Constructing a histogram feature descriptor based on the degree value of the local obedience to the point-like, line-like and spherical structures, the number of neighborhood points and the mean value of the reflection intensity; Obtaining the feature vector based on the histogram feature descriptor; The classification unit is used to construct a classification neural network, and use the training data set to train the classification neural network, and classify the test data set by tree species based on the trained classification neural network to obtain a classification result; The method for the classification unit to obtain the classification result includes: Constructing the classification neural network; The classification neural network comprises: Sequence weighting module, weighted encoding structure and deep feedforward structure; The sequence weighting module includes: DFF, Softmax weight normalization function and Concatenate feature aggregation unit; The weighted coding structure includes: a multi-head attention module, a DFF and a LayerNorm layer normalization unit; The deep feedforward structure includes: a Linear fully connected layer, a Mish activation function, and a BatchNorm batch normalization unit; Extract the average reflection intensity and average height of each slice, and construct intensity series and height series; The intensity sequence, the height sequence and the feature vector are input into the classification neural network to obtain a classification result.
4. The point cloud tree species classification system based on airborne LiDAR according to claim 3, characterized in that: The method for the filtering unit to obtain the ground point includes: rasterizing the target area; Select the lowest point of each grid as a seed point, and construct an initial triangulated network based on the seed point; Calculate the angle and distance between each of the original point clouds and the vertex of the initial triangulated network; Setting an iteration angle threshold and an iteration distance threshold, taking the original point cloud with the angle and the distance satisfying the iteration angle threshold and the iteration distance threshold as ground points, and updating the triangulation network; Repeat the operation to obtain all ground points in the target area.
Citation Information
Patent Citations
Airborne LiDAR data tree species classification method based on three-dimensional deep learning
CN111898688A
Single-plant tree crown segmentation method based on deep learning and airborne laser point cloud
CN112819830A
Point cloud feature extraction method based on local point cloud position relation
CN114998604A
Single tree segmentation method based on fused LiDAR point cloud data
CN115100232A
ALS point cloud classification method based on random forest
CN115761382A
