Forest landscape visual quality evaluation method and system based on point cloud data and machine learning

Abstract

The invention is suitable for the technical field of forest landscape evaluation, and provides a forest landscape visual quality evaluation method and system based on point cloud data and machine learning, and the method comprises the steps: obtaining the three-dimensional point cloud data of a forest land through a laser radar, and carrying out the data preprocessing based on the obtained three-dimensional point cloud data; extracting landscape 3D measurement features based on the preprocessed data; constructing a landscape visual quality evaluation system based on the extracted landscape 3D measurement features, determining a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D measurement feature, and determining a sample data set of machine learning based on the visual quality evaluation score or the visual quality evaluation grade; and training the preset landscape visual quality evaluation model by using the sample data set, and performing forest landscape visual quality evaluation on any new three-dimensional point cloud data by using the trained landscape visual quality evaluation model, so that the requirements of high efficiency and high precision in forest landscape visual quality evaluation are met.

Description

Technical Field

[0001] The present invention relates to the technical field of forest landscape assessment, and in particular to a method and system for assessing the visual quality of a forest landscape based on point cloud data and machine learning. Background Art

[0002] With the rise of ecotourism and the increasing importance of forest management, the visual quality assessment of forest landscapes has received increasing attention. Forest landscape refers to the sensory space created by the forest as the main body and the surrounding environment. The landscape appreciation is reflected in people's perception of the landscape in terms of vision, hearing, taste, touch and smell, among which vision plays a dominant role. Visual quality is the most important factor affecting the appreciation of landscapes, and it is also the actual need and ultimate goal for the improvement and utilization of forest ecosystems. The assessment of landscape visual quality helps to guide the configuration of forest landscapes, which is of great significance for improving the structural stability of forests and the diversity of ecosystems.

[0003] Landscape visual quality is "the state of landscape space, function and visual structure within a given time", which contains two meanings: landscape visual characteristics and human visual perception. The evaluation of landscape visual characteristics focuses more on the intrinsic attributes of the landscape. The visual quality of forest landscape is closely related to the characteristic attributes of forests, and viewing preferences are affected by the quality or structure of the forest stand, such as plant height, diameter at breast height, stand density, configuration pattern, etc. However, the physical attributes of forest landscapes usually use 2D and 2.5D indicators based on photos and videos, such as GIS-based 2D landscape evaluation models and 2.5D terrain attributes extracted from digital elevation models. These indicators are usually evaluated from a top-down bird's-eye view, which cannot reflect the real scenes that users see or experience in a given landscape at the human eye level, nor can they meet the high requirements of high efficiency and high precision in the evaluation of forest landscape visual quality.

[0004] In view of this, a forest landscape visual quality assessment method and system based on point cloud data and machine learning is needed. Summary of the invention

[0005] The embodiments of the present application provide a method and system for forest landscape visual quality assessment based on point cloud data and machine learning, which are used to solve the problem that the existing technology cannot meet the high requirements of high efficiency and high precision in forest landscape visual quality assessment.

[0006] The first aspect of the embodiment of the present application provides a forest landscape visual quality assessment method based on point cloud data and machine learning, comprising:

[0007] Use laser radar to obtain three-dimensional point cloud data of the forest land, and perform data preprocessing based on the obtained three-dimensional point cloud data;

[0008] Extracting 3D metric features of the landscape based on the preprocessed data;

[0009] Constructing a landscape visual quality evaluation system based on the extracted landscape 3D metric features, determining a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D metric feature according to the landscape visual quality evaluation system, and determining a sample data set for machine learning based on the visual quality evaluation score or the visual quality evaluation grade;

[0010] The sample data set is used to train a preset landscape visual quality assessment model, and the trained landscape visual quality assessment model is used to perform forest landscape visual quality assessment on any new three-dimensional point cloud data.

[0011] Furthermore, the extracting of landscape 3D metric features based on the preprocessed data includes:

[0012] Use the PointNet model to normalize and voxelize the preprocessed data to obtain the global feature information of the point cloud;

[0013] The voxelized data is input into the convolutional neural network to extract the local feature information of the voxel data;

[0014] Fusing the global feature information of the point cloud with the local feature information of the voxel data to obtain a fused feature vector;

[0015] The landscape 3D metric features are calculated based on the fused feature vector.

[0016] Furthermore, the global feature information of the point cloud is fused with the local feature information of the voxel data to obtain a fused feature vector, and the expressions include:

[0017]

[0018] G f =σ(G s W f +b f )

[0019]

[0020] Among them: F fusion is the fusion feature vector, G f is the transformed PointNet feature, is the feature after the convolutional neural network transformation, σ is the activation function, G s is the key global feature, W f is the weight matrix of the fully connected layer, b f is the bias vector, is the feature vector converted by global average pooling, h and w are the height and width of the feature map, is the eigenvalue of the jth feature map at the (x, y) position.

[0021] Furthermore, the calculating of the landscape 3D metric features based on the fused feature vector includes:

[0022] The landscape 3D metric features include landscape diversity, landscape shape, landscape connectivity, color uniformity, openness, and topography;

[0023] Landscape diversity:

[0024]

[0025] Where: HVDD is the diversity index, N is the number of groups, V i is the volume of the i-th group, is the average volume of all groups;

[0026] Landscape shape:

[0027]

[0028] Where: MSI is the morphological index, SVR is the surface area to volume ratio, S is the actual surface area of ​​the landscape, S min is the minimum surface area of ​​landscapes of the same size, and V is the volume of the landscape;

[0029] Landscape Connectivity:

[0030] NC=M

[0031]

[0032] Where: NC is the number of clusters, CD is the cluster density, CS is the cluster size, CDist is the cluster distance, M is the number of clusters obtained after statistical clustering, H a is the horizontal area of ​​the landscape, V h is the maximum height of the landscape, |C i For each cluster C i The number of voxels included, d(C i ,C j ) is the distance between the centroid voxel points of the i-th cluster and the j-th cluster;

[0033] Color uniformity:

[0034]

[0035] Among them: CD v is the color distance, Dave is the average color distance, C v is the color change index, v r 、v g and vb are the values ​​of the red, green and blue color channels of the voxel, respectively, v is the total number of voxels;

[0036] Openness:

[0037]

[0038] Where: VD is the visual distance, K is the number of groups, d min (g k ) is the minimum distance between the viewpoint of the kth group and any voxel in the group;

[0039] terrain:

[0040]

[0041] Where: AS is the average slope, GV is the gradient change, N g is the number of ground grids, s i is the slope of the ith ground grid, g i is the gradient of the ith ground grid, g ave is the average value of all grid gradients.

[0042] Furthermore, the landscape visual quality evaluation system is constructed based on the extracted landscape 3D metric features, a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D metric feature is determined according to the landscape visual quality evaluation system, and a sample data set for machine learning is determined based on the visual quality evaluation score or the visual quality evaluation grade, including:

[0043] Analyze and screen the extracted landscape 3D metric features to determine the key feature subsets related to landscape visual quality in each landscape 3D metric feature;

[0044] Collect expert scores for the selected subset of key features and determine the visual quality level or visual quality score corresponding to each landscape 3D metric feature;

[0045] The landscape 3D metric feature vector corresponding to the key feature subset screened out from each sample is used as the feature data of the data set, and the score or rating of the landscape visual quality of each landscape 3D metric feature is used as the label data of the data set to obtain a sample data set for machine learning.

[0046] Furthermore, the expression for collecting and filtering the expert scores of the key feature subset includes:

[0047]

[0048] in: is the normalized value of the landscape visual quality score of the pth subject for the i-th key feature subset in the landscape 3D metric feature X, The landscape visual quality score of the p-th subject for the i-th key feature subset of the landscape 3D metric feature X, is the average of the landscape visual quality scores of the pth subject for all subsets of the landscape 3D metric features X, is the standard deviation of the landscape visual quality ratings of the pth subject for all subsets of the landscape 3D metric features X, N p is the total number of subjects, is the final normalized landscape visual quality value of the i-th key feature subset in the landscape 3D metric feature X.

[0049] Furthermore, the collecting of expert scores of the selected key feature subsets and determining the visual quality level or visual quality score corresponding to each landscape 3D metric feature includes:

[0050] According to the isometric method, the landscape visual quality is divided into five levels, namely very good, good, average, poor, and very poor:

[0051] Very good: [(LVD max ―LVD min )×0.8+LVD min ,LVD max ]

[0052] Good: [(LVD max ―LVD min )×0.6,(LVD max ―LVD min )×0.8+LVD min ]

[0053] generally:

[0054] [(LVD max ―LVD min )×0.4+LVD min ,(LVD max ―LVD min )×0.6+LVD min ]

[0055] Difference:

[0056] [(LVD max ―LVD min )×0.2+LVD min ,(LVD max ―LVD min )×0.4+LVD min ]

[0057] Very poor: [LVD min ,(LVD max ―LVD min )×0.2+LVD min ]

[0058] Among them, LVD max Indicates the maximum landscape visual quality value, LVD min Indicates the minimum landscape visual quality value.

[0059] Furthermore, the collecting of expert scores of the selected key feature subsets and determining the visual quality level or visual quality score corresponding to each landscape 3D metric feature includes:

[0060]

[0061] Where: S is the visual quality score, k is the number of key feature subsets, w i is the weight of the i-th key feature subset.

[0062] Furthermore, the method of using the sample data set to train a preset landscape visual quality assessment model, and using the trained landscape visual quality assessment model to perform forest landscape visual quality assessment on any new three-dimensional point cloud data, includes:

[0063] The model learns the mapping relationship between the key feature subsets in each landscape 3D metric feature and the corresponding landscape 3D metric feature, as well as the mapping relationship between each landscape 3D metric feature and the landscape visual quality, and determines the trained landscape visual quality assessment model based on the learned mapping relationship.

[0064] A second aspect of the embodiment of the present application provides a forest landscape visual quality assessment system based on point cloud data and machine learning, comprising:

[0065] A data acquisition and preprocessing unit, used for acquiring three-dimensional point cloud data of the forest land by using a laser radar, and performing data preprocessing based on the acquired three-dimensional point cloud data;

[0066] A landscape 3D metric feature extraction unit, used for extracting landscape 3D metric features based on the preprocessed data;

[0067] A sample data set determination unit, configured to construct a landscape visual quality evaluation system based on the extracted landscape 3D metric features, determine a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D metric feature according to the landscape visual quality evaluation system, and determine a sample data set for machine learning based on the visual quality evaluation score or the visual quality evaluation grade;

[0068] The landscape visual quality assessment model training unit is used to train a preset landscape visual quality assessment model using the sample data set, and use the trained landscape visual quality assessment model to perform forest landscape visual quality assessment on any new three-dimensional point cloud data.

[0069] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0070] The present invention combines point cloud data with machine learning. The evaluation based on the extracted 3D landscape metrics can more objectively and truly reflect the landscape visual quality. Moreover, the automated evaluation process greatly improves the evaluation efficiency and saves manpower, material resources and time costs. The three-dimensional point cloud data of the forest land is obtained by using laser radar, and data preprocessing is performed based on the obtained three-dimensional point cloud data. The landscape 3D metric features are extracted based on the preprocessed data. A landscape visual quality evaluation system is constructed based on the extracted landscape 3D metric features, and a sample data set for machine learning is determined according to the landscape visual quality evaluation system. Finally, the sample data set is used to train the preset landscape visual quality evaluation model, and the trained landscape visual quality evaluation model is used to evaluate the visual quality of the forest landscape on any new three-dimensional point cloud data. The model can be continuously optimized according to new data and requirements, adapt to the evaluation of different scenarios and requirements, effectively quantify the visual quality of the forest landscape, improve the objectivity and efficiency of the landscape visual quality evaluation, and provide a new technical means for the landscape visual quality evaluation.

[0071] Other advantages, objectives, and features of the present invention will be set forth in part in the following description, and in part will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a schematic diagram of an embodiment of a method for assessing visual quality of forest landscape based on point cloud data and machine learning in the present invention;

[0073] Figure 2 It is a laser radar scan image for collecting forest land data in the present invention;

[0074] Figure 3 Schematic diagram of the landscape visual quality evaluation system constructed based on the extracted landscape 3D metric features in the present invention. DETAILED DESCRIPTION

[0075] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "corresponding to" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0076] Embodiment 1

[0077] The implementation method in this embodiment can be implemented in the system, in the server, or in the terminal, and the specific implementation is not clearly limited. The following is an introduction to the forest landscape visual quality assessment method based on point cloud data and machine learning in this application from the perspective of system implementation. Figure 1 , the method provided in the embodiment of the present application comprises the following steps:

[0078] S11. using a laser radar to obtain three-dimensional point cloud data of the forest land, and performing data preprocessing based on the obtained three-dimensional point cloud data;

[0079] In this embodiment, the three-dimensional point cloud data of the forest is obtained by using LiDAR technology, and data preprocessing is performed based on the collected three-dimensional point cloud data. Representative forest areas are selected for data collection, covering different forest stand types, topography, and different forest age structures, to ensure that the collected data can fully reflect the overall characteristics of the forest. Specifically, according to the forest conditions and research needs, a suitable scanning angle range is set, and the number of scans performed by the LiDAR per unit time, the wavelength of the laser, the power and other parameters are set. During the data collection process, the operating status of the equipment is monitored in real time to ensure that the LiDAR transmits and receives laser pulse signals normally, and the data can be accurately recorded in the storage device. The collected point cloud data is preprocessed, including noise reduction, coordinate transformation, etc., to generate a high-quality three-dimensional model. Then, the quantitative structural model constructed using TreeQSM software is used to extract tree parameters and create a 3D voxel information data frame with leaf area density (LAD) values. Please refer to Figure 2 , from left to right, respectively represent the picture of the forest point cloud data collected by lidar, the model diagram of building a quantitative structural model to extract tree parameters, and the schematic diagram of creating a 3D voxel data frame.

[0080] S12. extracting landscape 3D metric features based on the preprocessed data;

[0081] In this embodiment, step S12 includes the following:

[0082] 1. Use the PointNet model to normalize and voxelize the preprocessed data to obtain the global feature information of the point cloud;

[0083] Assume that the preprocessed point cloud data is P = {p 1 ,p 2 ,...,p n}, where p i =(x i ,y i ,z i ) represents the three-dimensional coordinates of the i-th point, calculates the minimum and maximum values ​​on each coordinate axis, and normalizes the coordinates of each point.

[0084] Determine the size of the voxel s according to the spatial range of the point cloud [x min ,x max ]×[y min ,y max ]×[z min ,z max ], divide the space into a three-dimensional voxel grid, for each point p′ i , calculate the voxel index (v x ,v y ,v z ),in Count the number of points N in each voxel v , average coordinates and average reflection intensity as the eigenvalues ​​of the voxels.

[0085] The normalized point cloud data P′={p 1 ′,p 2 ′,...,p n ′} is input into PointNet, and the initial feature vector f of each point i Can be set to f i =(x′ i ,y′ i ,z′ i ,r′ i ). After multiple MLP layers, each MLP layer consists of fully connected neurons. Let the input feature dimension of the j-th MLP layer be The output feature dimension is The weight matrix is ​​W j , the bias vector is b j For the eigenvector of the i-th point The output calculation formula of the MLP at the jth layer is: Where σ is the activation function. After passing through multiple MLP layers, the feature vector of each point is obtained The global feature vector G is obtained through the maximum pooling operation. For the kth dimension of the global feature vector G

[0086] 2. Input the voxelized data into the convolutional neural network to extract the local feature information of the voxel data;

[0087] Construct a CNN model containing multiple convolutional layers, pooling layers, and fully connected layers. For example, the convolution kernel size of the first convolutional layer is 3×3×3, the stride is 1, and the number of input channels is the dimension d of the voxel feature. v , such as d v =4, including the number of points, average coordinates and average reflection intensity, the number of output channels is c 1 , c 1 =16.

[0088] If there is labeled forest landscape point cloud data and its corresponding local feature labels of landscape 3D metric features, CNN is trained. Suppose there are m voxelized point cloud samples V = {V 1 ,V 2 ,…,V m}, the corresponding local feature label is Y = {y 1 ,y 2 ,…,y m}. Define the loss function, such as the mean square error loss function in is the predicted local feature of CNN for the i-th sample. Through the back propagation algorithm, the parameters of the CNN model are updated according to the loss function to minimize the loss function.

[0089] The voxelized point cloud data V = {V 1 ,V 2 ,…,V m} Input into the trained CNN model, and after processing by the convolution layer and the pooling layer, the feature map of the voxel data is obtained Where l is the number of feature maps, and the size of each feature map is h×w×c, where h and w are the height and width of the feature map, and c is the number of channels.

[0090] 3. Fuse the global feature information of the point cloud with the local feature information of the voxel data to obtain a fused feature vector;

[0091] Select the top d from the global feature vector G extracted by PointNet g Dimensions are used as key global features G s , and transform its dimension to d through a fully connected layer f, the transformed feature is recorded as G f , the calculation formula is: G f =σ(G s W f +b f ), where: W f is the weight matrix of the fully connected layer, b f is the bias vector and σ is the activation function.

[0092] For the feature maps extracted by CNN, a global average pooling operation is used to convert each feature map into a feature vector in: is the eigenvalue of the jth feature map at position (x, y); then these feature vectors are concatenated into a long vector And through a fully connected layer, its dimension is also transformed to d f , the transformed features are recorded as Using the splicing fusion strategy, the transformed PointNet feature G f and CNN features Splice in dimension to get dimension 2d f Fuse feature vectors:

[0093] 4. Calculate the landscape 3D metric features based on the fused feature vector.

[0094] The 3D metric features of the landscape include landscape diversity, landscape shape, landscape connectivity, color uniformity, openness, and topography; in terms of landscape diversity, the horizontal, vertical, and distance diversity (HVDD) index is used for quantification. The main components of HVDD include the horizontal, vertical, and distance profiles of the landscape visual quality. The horizontal profile shows the landscape structure based on different horizontal segments; the vertical profile reveals the vertical layered structure of the landscape; and the distance curve represents the distance of the landscape element to the central location. The spatial structure of the landscape visual quality is described by combining the horizontal, vertical, and distance profiles. The calculation of HVDD requires first dividing the landscape into groups based on the horizontal angle, height, and distance to the viewpoint. The number of groups depends on the voxel size and is not fixed. Then, the volumes of each group are compared to calculate the spatial diversity of the landscape visual quality. The larger the HVDD value, the higher the spatial diversity and landscape complexity. The calculation formula is as follows:

[0095]

[0096] Where: HVDD is the diversity index, N is the number of groups, V i is the volume of the i-th group, is the average volume of all groups.

[0097] In terms of landscape shape, morphological index and surface area to volume ratio are used to quantify shape diversity or regularity. The three-dimensional landscape morphological index is the ratio of the actual surface area of ​​the landscape to the minimum surface area of ​​a landscape of the same size. Compared with 2D shape indicators based on edge length or 2D area, 3D indicators use 3D surface area or volume for analysis. The value range of three-dimensional morphological indicators is 1 to infinity. The surface area to volume ratio is the ratio of the surface area of ​​the landscape to its volume. The lower the three-dimensional morphological index or surface area to volume ratio, the more standard and regular the landscape shape is, and vice versa, the more complex the shape is. Shape extraction is performed based on the landscape shape recognition algorithm based on unsupervised learning in machine learning. The calculation formula is as follows:

[0098]

[0099]

[0100] Where: MSI is the morphological index, SVR is the surface area to volume ratio, S is the actual surface area of ​​the landscape, S min is the minimum surface area of ​​landscapes of the same size, and V is the volume of the landscape.

[0101] Landscape connectivity includes four indicators, namely cluster number, cluster density, cluster size and cluster distance. The number of clusters refers to the number of independent clusters, and the independent clusters must have a volume and distance greater than a certain size and distance from the nearest cluster; cluster density is the ratio of the number of clusters to the volume of the landscape space; cluster size refers to the average volume of clusters; cluster distance refers to the average distance between the centroid voxel points of different clusters. The fewer the number of clusters, the lower the density, the larger the volume, and the shorter the distance, the higher the landscape connectivity. The calculation formula is as follows:

[0102] NC=M

[0103]

[0104] Where: NC is the number of clusters, CD is the cluster density, CS is the cluster size, CDist is the cluster distance, M is the number of clusters obtained after statistical clustering, H a is the horizontal area of ​​the landscape, V h is the maximum height of the landscape, |C i For each cluster C i The number of voxels included, d(C i ,C j ) is the distance between the centroid voxel points of the i-th cluster and the j-th cluster.

[0105] Color uniformity is quantified by the color variation between voxels. First, the color distance between the voxel and the baseline (black - RGB (0,0,0)) is calculated; then, the color distance is used to calculate the color variation index. A high color variation indicates that the landscape colors are significantly different, otherwise, the landscape colors are similar. The formula is as follows:

[0106]

[0107]

[0108] Among them: CD v is the color distance, Dave is the average color distance, C v is the color change index, v r 、v g and v b are the values ​​of the red, green and blue color channels of the voxel, respectively, v is the total number of voxels;

[0109] Openness is the quantification of visual scale. In terms of landscape openness, the visual distance is used to characterize the depth and range of the landscape, that is, the degree of openness is quantified based on the central viewpoint. The visual distance measures the distance between the viewpoint and the surrounding obstacles. The space is divided into groups according to the horizontal and vertical angles, and the visual distance of all groups (the minimum distance between the viewpoint and any voxel in each group) is measured. The average visual distance is calculated as the visual distance of the landscape unit. A longer visual distance indicates a high openness of the landscape, and vice versa, the landscape is closed and compact. The calculation formula is as follows:

[0110]

[0111] Where: VD is the visual distance, K is the number of groups, d min (g k ) is the minimum distance between the viewpoint of the kth group and any voxel in the group.

[0112] The terrain is quantified using average slope and gradient change. The grid slope is extracted based on the generated ground grid data, and the average of all slopes is the average slope of the ground steepness. The gradient change reflects the difference in grid gradients. A high gradient change index indicates that the terrain of the entire plot has a large change, otherwise it indicates that the ground is flat. The calculation formula is as follows:

[0113]

[0114]

[0115] Where: AS is the average slope, GV is the gradient change, N g is the number of ground grids, s i is the slope of the ith ground grid, g i is the gradient of the ith ground grid, g ave is the average value of all grid gradients.

[0116] S13. construct a landscape visual quality evaluation system based on the extracted landscape 3D metric features, determine the visual quality evaluation score or visual quality evaluation grade corresponding to each landscape 3D metric feature according to the landscape visual quality evaluation system, and determine a sample data set for machine learning based on the visual quality evaluation score or visual quality evaluation grade;

[0117] Based on the extracted landscape 3D metric features, a landscape visual quality evaluation index system is constructed. Figure 3 , where the index system includes the target layer, the criterion layer and the indicator layer. The target layer is the visual quality of the landscape, the criterion layer includes the 3D metric features of the landscape, namely landscape diversity, landscape shape, landscape connectivity, color unity, openness and terrain; the indicator layer includes a subset of key features obtained by analyzing and screening the indicators in each 3D metric feature of the landscape. Specifically, they are as follows:

[0118] 1. Analyze and screen the extracted landscape 3D metric features to determine the key feature subsets related to landscape visual quality in each landscape 3D metric feature;

[0119] As mentioned above, the 3D metric features of landscapes include landscape diversity, landscape shape, landscape connectivity, color uniformity, openness, terrain and other features. Each feature includes multiple indicators. For example, landscape diversity may include horizontal, vertical, distance and other indicators; landscape shape includes morphological index, volume and other indicators. The key feature subset is to select some indicators that best reflect the essence and core of landscape visual quality from these large numbers of indicators.

[0120] First, we collect experts’ visual quality ratings of various landscape 3D metric features.

[0121] The visual quality scores of the landscape 3D metric features of the experts were collected. Specifically, corresponding to the landscape 3D metric features extracted above, a forest landscape visual quality evaluation questionnaire was set based on all landscape 3D metric features including landscape diversity, landscape shape, landscape connectivity, color uniformity, openness and terrain. Forest landscape pictures were collected in each forestland where point cloud data was collected. The same person and one camera were used to take pictures during the bright and high visibility time period of the day (9:00-11:00, 15:00-17:00), and the consistency of pictures in various plots was ensured on the basis of fully reflecting the vegetation characteristics. The collected photos were first processed for background removal and pixel uniformity using Adobe Photoshop, and then the subjects were invited to score after observing the photos for 5-7 seconds, and the photo landscape was evaluated by first intuition. The questionnaire was conducted by combining field research with random distribution on the network platform. The questionnaire results were analyzed for validity and reliability using Cronbach test and Pearson correlation analysis, and then the rationality of the number of valid questionnaires was determined.

[0122] In order to eliminate the influence of individual differences in landscape evaluation, the original questionnaire results were standardized and the calculation formula is as follows:

[0123]

[0124] in: is the normalized value of the landscape visual quality score of the pth subject for the landscape 3D metric feature X, is the landscape visual quality score of the p-th subject for the landscape 3D metric feature X, is the average of the landscape visual quality scores of the pth subject on the landscape 3D metric feature X, is the standard deviation of the landscape visual quality score of the pth subject for the landscape 3D metric feature X, N p is the total number of subjects, LVD X is the final normalized landscape visual quality value of the landscape 3D metric feature X.

[0125] Then, the Pearson correlation coefficient between each indicator in each extracted landscape 3D metric feature and the above-mentioned known landscape visual quality score is calculated to screen out the key feature subset related to landscape visual quality in each landscape 3D metric feature. Let the landscape diversity vector in the landscape 3D metric feature be X 1 ={x 1 ,x 2 ,…,x n}, where n is the number of indicators, and the corresponding landscape visual quality score is Y = {y 1 ,y 2 ,…,y m}, where m is the number of samples. For the i-th indicator x i and score y, Pearson correlation coefficient The calculation formula is:

[0126]

[0127] Where: x ij is the i-th index value of the j-th sample, is the average value of the ith indicator, is the average of the scores. According to the absolute value of the correlation coefficient, select the indicators with higher correlation, such as Indicators > 0.5 were used as a preliminary subset of key features.

[0128] 2. Collect expert scores for the selected subset of key features and determine the visual quality level or visual quality score corresponding to each landscape 3D metric feature;

[0129] Collect expert scores on landscape visual quality for key feature subsets corresponding to landscape 3D metric features: for example, the key feature subsets obtained based on the above landscape diversity screening include horizontal, vertical and distance diversity. Similarly, determine the key feature subsets of other landscape 3D metric features, and then set up a forest landscape visual quality evaluation questionnaire based on the key feature subsets corresponding to all landscape 3D metric features. The questionnaire image collection, post-processing and inspection are the same as above.

[0130] Similarly, in order to eliminate the influence of individual differences in landscape evaluation, the original questionnaire results were standardized and the calculation formula was as follows:

[0131]

[0132] in: is the normalized value of the landscape visual quality score of the pth subject for the i-th key feature subset in the landscape 3D metric feature X, The landscape visual quality score of the p-th subject for the i-th key feature subset of the landscape 3D metric feature X, is the average of the landscape visual quality scores of the pth subject for all subsets of the landscape 3D metric features X, is the standard deviation of the landscape visual quality ratings of the pth subject for all subsets of the landscape 3D metric features X, N p is the total number of subjects, is the final normalized landscape visual quality value of the i-th key feature subset in the landscape 3D metric feature X.

[0133] After determining the scores of the key feature subsets corresponding to each landscape 3D metric feature, the visual quality assessment results of the landscape 3D metric feature can be obtained in two ways. One is to determine the assessment result by calculating the quality assessment score by setting the weight distribution of each subset, and the other is to determine the assessment result by calculating the quality assessment grade based on the assessment score by the isometric method.

[0134] In terms of evaluation scores, the calculation formula for the landscape visual quality score of each landscape 3D metric feature is as follows:

[0135]

[0136] Where: S X is the visual quality score of the landscape 3D metric feature X, k is the number of key feature subsets in the landscape 3D metric feature X, and w Xi is the weight of the i-th key feature subset in the landscape 3D metric feature X, and the weight is determined by the proportion of the scores of each key feature subset in the landscape 3D metric feature X to the total scores of all subsets.

[0137] In terms of evaluation level, the landscape visual quality is divided into five levels according to the equidistance method, namely very good, good, average, poor, and very poor.

[0138] Very good: [(LVD max ―LVD min )×0.8+LVD min ,LVD max ]

[0139] Good: [(LVD max ―LVD min )×0.6,(LVD max ―LVD min )×0.8+LVD min ]

[0140] generally:

[0141] [(LVD max ―LVD min )×0.4+LVD min ,(LVD max ―LVD min )×0.6+LVD min ]

[0142] Difference:

[0143] [(LVD max ―LVD min )×0.2+LVD min ,(LVD max ―LVD min )×0.4+LVD min ]

[0144] Very poor: [LVD min ,(LVD max ―LVD min )×0.2+LVD min ]

[0145] Among them, LVD max Indicates the maximum landscape visual quality value, LVD min Indicates the minimum landscape visual quality value.

[0146] 3. The landscape 3D metric feature vector extracted from each sample is used as the feature data of the dataset, and the visual quality score or rating of each landscape 3D metric feature is used as the label data of the dataset to obtain a dataset for machine learning.

[0147] Specifically, the obtained machine learning dataset is divided into a training set, a validation set, and a test set, wherein the training set is used as a sample dataset for training the model, the validation set is used to adjust the model parameters during the model training process, and the test set is used to evaluate the performance of the model after training. The landscape 3D metric feature vector corresponding to the selected key feature subset extracted from each sample is used as the feature data of the dataset, and the visual quality score or rating S of each landscape 3D metric feature is used as the label data to construct a dataset for machine learning {(X 1 ,S 1 ),(X 2 ,S 2 ),…,(X m ,S m )}, where X j is the feature vector value of the jth sample, S j Give it a corresponding visual quality score or rating.

[0148] The label data of the machine learning sample dataset here can be either the visual quality rating of each landscape 3D metric feature or the visual quality score. Providing two types of evaluation results can meet the training needs of different models. The machine learning models involved in visual quality evaluation include: decision tree (DT), random forest (RF), Gaussian process regression (GPR), support vector machine regression (SVR) and neural network (NN).

[0149] S14. Use the sample data set to train the preset landscape visual quality assessment model, and use the trained landscape visual quality assessment model to perform forest landscape visual quality assessment on any new three-dimensional point cloud data.

[0150] In this embodiment, step S14 also includes the following:

[0151] The model learns the mapping relationship between the key feature subsets in each landscape 3D metric feature and the corresponding landscape 3D metric feature, as well as the mapping relationship between each landscape 3D metric feature and the landscape visual quality, and determines the trained landscape visual quality assessment model based on the learned mapping relationship.

[0152] Assume that the landscape 3D metric feature vector corresponding to the sample dataset is X target =[x t1 ,x t2 ,...,x tn ], where n is the number of features, x ti is the i-th eigenvalue. The landscape 3D metric eigenvector X targetand its corresponding visual quality level or visual quality score are input into the constructed landscape visual quality assessment model to learn the mapping relationship between each landscape 3D metric feature and landscape visual quality of different degrees, where the different degrees include the mapping relationship between the key feature subsets in each landscape 3D metric feature and the corresponding landscape 3D metric feature of different visual quality levels or different visual quality scores; and the mapping relationship between each landscape 3D metric feature and the above-determined different levels of landscape visual quality or the mapping relationship between different ranges of landscape visual quality score values. In addition, different types of models have different input layer structures, but all receive this feature vector as the initial input data.

[0153] The specific visual quality score or grade of the forest landscape can be determined based on the mapping relationship between the learned 3D metric features of each landscape and the visual quality of the landscape. Here, whether the model outputs the visual quality score or the visual quality grade is determined based on the selected training model. If the selected model is trained based on a regression task, a specific value will be output; if the selected model is trained based on a classification task, the grade of the landscape visual quality will be output. During the training process, a partial data validation set with landscape 3D metric features and scores is used to monitor the training effect of the model and prevent overfitting. By comparing the prediction results of the model on the validation set with the performance of the actual scoring evaluation model, the trained visual quality evaluation model is finally determined.

[0154] When the 3D metric feature data of a new forest landscape is input into the trained model, the model calculates and judges based on the learned mapping relationship, and outputs the corresponding visual quality quantification value or the quality level (such as very good, good, average, poor, very poor). For example, for a forest landscape with a specific combination of landscape diversity, shape, connectivity and other features, the model can give an accurate score, such as 75 points, or classify it into the "good" grade category, thereby achieving an objective and accurate assessment of the visual quality of the forest landscape, overcoming the subjectivity and ambiguity of traditional qualitative assessment methods.

[0155] Finally, the test set is used to evaluate and optimize the trained model. Here, the root mean square error, R 2 , accuracy, recall and other evaluation indicators are used to comprehensively evaluate the model performance. According to the evaluation results, the model is optimized, including but not limited to adjusting hyperparameters, selecting different feature combinations, and increasing training data. The selection of hyperparameters is crucial in the training process of machine learning models. Hyperparameter tuning can effectively improve the accuracy and stability of the model. The tuning methods include grid search, random search or Bayesian optimization.

[0156] The above embodiment constructs a landscape visual quality evaluation system based on landscape 3D metric features, and determines a sample data set suitable for machine learning, laying the foundation for the subsequent training of an accurate landscape visual quality prediction model. Finally, any new three-dimensional data acquired in real time is input into the trained landscape visual quality evaluation model to obtain the corresponding visual quality evaluation results, whether it is a quantitative score or a classification level, which can provide a reference for related applications such as forest landscape management and ecotourism planning.

[0157] Embodiment 2

[0158] An embodiment of a forest landscape visual quality assessment system based on point cloud data and machine learning in the present invention comprises the following steps:

[0159] A data acquisition and preprocessing unit, used for acquiring three-dimensional point cloud data of the forest land by using a laser radar, and performing data preprocessing based on the acquired three-dimensional point cloud data;

[0160] A landscape 3D metric feature extraction unit, used for extracting landscape 3D metric features based on the preprocessed data;

[0161] A sample data set determination unit, used to construct a landscape visual quality evaluation system based on the extracted landscape 3D metric features, and determine a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D metric feature according to the landscape visual quality evaluation system, and determine a sample data set for machine learning based on the visual quality evaluation score or the visual quality evaluation grade;

[0162] The landscape visual quality assessment model training unit is used to train a preset landscape visual quality assessment model using a sample data set, and use the trained landscape visual quality assessment model to perform forest landscape visual quality assessment on any new three-dimensional point cloud data.

[0163] For the specific definition of the system, please refer to the definition of the method above, which will not be repeated here. Each module in the above system can be implemented in whole or in part by software, hardware and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0164] Those of ordinary skill in the art will appreciate that the units of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0165] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored. In addition, each functional unit in each embodiment of the present invention can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units.

[0166] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nlyMemory), random access memory (RAM, RandomAccessMemory), mobile hard disk, magnetic disk or optical disk, etc., which can store program code.

[0167] It can be understood that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A method for assessing forest landscape visual quality based on point cloud data and machine learning, characterized in that: include: Use laser radar to obtain three-dimensional point cloud data of the forest land, and perform data preprocessing based on the obtained three-dimensional point cloud data; Extracting 3D metric features of the landscape based on the preprocessed data; Constructing a landscape visual quality evaluation system based on the extracted landscape 3D metric features, determining a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D metric feature according to the landscape visual quality evaluation system, and determining a sample data set for machine learning based on the visual quality evaluation score or the visual quality evaluation grade; The sample data set is used to train a preset landscape visual quality assessment model, and the trained landscape visual quality assessment model is used to perform forest landscape visual quality assessment on any new three-dimensional point cloud data.

2. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 1, characterized in that: The extracting of landscape 3D metric features based on the preprocessed data includes: Use the PointNet model to normalize and voxelize the preprocessed data to obtain the global feature information of the point cloud; The voxelized data is input into the convolutional neural network to extract the local feature information of the voxel data; Fusing the global feature information of the point cloud with the local feature information of the voxel data to obtain a fused feature vector; The landscape 3D metric features are calculated based on the fused feature vector.

3. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 2 is characterized in that: The global feature information of the point cloud is fused with the local feature information of the voxel data to obtain a fused feature vector, and the expressions include: G f =σ(G s W f +b f ) Among them: F fusion is the fusion feature vector, G f is the transformed PointNet feature, is the feature after the convolutional neural network transformation, σ is the activation function, G s is the key global feature, W f is the weight matrix of the fully connected layer, b f is the bias vector, is the feature vector converted by global average pooling, h and w are the height and width of the feature map, is the eigenvalue of the jth feature map at the (x, y) position.

4. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 2 is characterized in that: The calculating of the landscape 3D metric feature based on the fused feature vector comprises: The landscape 3D metric features include landscape diversity, landscape shape, landscape connectivity, color uniformity, openness, and topography; Landscape diversity: Where: HVDD is the diversity index, N is the number of groups, V i is the volume of the i-th group, is the average volume of all groups; Landscape shape: Where: MSI is the morphological index, SVR is the surface area to volume ratio, S is the actual surface area of ​​the landscape, S min is the minimum surface area of ​​landscapes of the same size, and V is the volume of the landscape; Landscape Connectivity: NC=M Where: NC is the number of clusters, CD is the cluster density, CS is the cluster size, CDist is the cluster distance, M is the number of clusters obtained after statistical clustering, H a is the horizontal area of ​​the landscape, V h is the maximum height of the landscape, |C i For each cluster C i The number of voxels included, d(C i ,C j ) is the distance between the centroid voxel points of the i-th cluster and the j-th cluster; Color uniformity: Among them: CD v is the color distance, Dave is the average color distance, C v is the color change index, v r 、v g and v b are the values ​​of the red, green and blue color channels of the voxel, respectively, v is the total number of voxels; Openness: Where: VD is the visual distance, K is the number of groups, d min (g k ) is the minimum distance between the viewpoint of the kth group and any voxel in the group; terrain: Where: AS is the average slope, GV is the gradient change, N g is the number of ground grids, s i is the slope of the ith ground grid, g i is the gradient of the ith ground grid, g ave is the average value of all grid gradients.

5. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 1, characterized in that: The method comprises: constructing a landscape visual quality evaluation system based on the extracted landscape 3D metric features, determining a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D metric feature according to the landscape visual quality evaluation system, and determining a sample data set for machine learning based on the visual quality evaluation score or the visual quality evaluation grade, including: Analyze and screen the extracted landscape 3D metric features to determine the key feature subsets related to landscape visual quality in each landscape 3D metric feature; Collect expert scores for the selected subset of key features and determine the visual quality level or visual quality score corresponding to each landscape 3D metric feature; The landscape 3D metric feature vector corresponding to the key feature subset screened out from each sample is used as the feature data of the data set, and the score or rating of the landscape visual quality of each landscape 3D metric feature is used as the label data of the data set to obtain a sample data set for machine learning.

6. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 5, characterized in that: The expression for collecting and filtering out the expert scores of the key feature subsets includes: in: is the normalized value of the landscape visual quality score of the pth subject for the i-th key feature subset in the landscape 3D metric feature X, The landscape visual quality score of the p-th subject for the i-th key feature subset of the landscape 3D metric feature X, is the average of the landscape visual quality scores of the pth subject for all subsets of the landscape 3D metric features X, is the standard deviation of the landscape visual quality ratings of the pth subject for all subsets of the landscape 3D metric features X, N p is the total number of subjects, is the final normalized landscape visual quality value of the i-th key feature subset in the landscape 3D metric feature X.

7. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 6, characterized in that: The collecting of expert scores of the selected key feature subsets and determining the visual quality level or visual quality score corresponding to each landscape 3D metric feature includes: According to the isometric method, the landscape visual quality is divided into five levels, namely very good, good, average, poor, and very poor: Very good: [(LVD max ―LVD min )×0.8+LVD min ,LVD max ] Good: [(LVD max ―LVD min )×0.6, (LVD max ―LVD min )×0.8 + LVD min ​ generally: [(LVD max ―LVD min )×0.4+LVD min ,(LVD max ―LVD min )×0.6+LVD min ] Difference: [(LVD max ―LVD min )×0.2+LVD min ,(LVD max ―LVD min )×0.4+LVD min ] Very poor: [LVD min ,(LVD max ―LVD min )×0.2+LVD min ] Among them, LVD max Indicates the maximum landscape visual quality value, LVD min Indicates the minimum landscape visual quality value.

8. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 7, characterized in that: The collecting of expert scores of the selected key feature subsets and determining the visual quality level or visual quality score corresponding to each landscape 3D metric feature includes: Where: S is the visual quality score, k is the number of key feature subsets, w i is the weight of the i-th key feature subset.

9. The method for forest landscape visual quality assessment based on point cloud data and machine learning according to claim 1, characterized in that: The method of using the sample data set to train a preset landscape visual quality assessment model, and using the trained landscape visual quality assessment model to perform forest landscape visual quality assessment on any new three-dimensional point cloud data, includes: The model learns the mapping relationship between the key feature subsets in each landscape 3D metric feature and the corresponding landscape 3D metric feature, as well as the mapping relationship between each landscape 3D metric feature and the landscape visual quality, and determines the trained landscape visual quality assessment model based on the learned mapping relationship.

10. A forest landscape visual quality assessment system based on point cloud data and machine learning, characterized in that: include: A data acquisition and preprocessing unit, used to acquire three-dimensional point cloud data of the forest land using a laser radar, and perform data preprocessing based on the acquired three-dimensional point cloud data; A landscape 3D metric feature extraction unit, used for extracting landscape 3D metric features based on the preprocessed data; A sample data set determination unit, configured to construct a landscape visual quality evaluation system based on the extracted landscape 3D metric features, determine a visual quality evaluation score or a visual quality evaluation grade corresponding to each landscape 3D metric feature according to the landscape visual quality evaluation system, and determine a sample data set for machine learning based on the visual quality evaluation score or the visual quality evaluation grade; The landscape visual quality assessment model training unit is used to train a preset landscape visual quality assessment model using the sample data set, and use the trained landscape visual quality assessment model to perform forest landscape visual quality assessment on any new three-dimensional point cloud data.

Images