A forestry remote sensing image recognition method and system based on image recognition

Abstract

The application discloses a forestry remote sensing image recognition method and system based on image recognition, relates to the technical field of image processing, and comprises the following steps: collecting and pre-processing LiDAR point clouds and multispectral optical images, and obtaining three-dimensional tree point clouds; performing tree feature segmentation through spectral difference analysis, re-distributing tree feature segmentation boundary points based on spectral dissimilarity, calculating core spectral signatures, and outputting a single-tree point cloud entity set; and performing topological analysis on the single-tree point cloud entity set and outputting a vector result. The three-dimensional tree point clouds are directly operated, and each tree is recognized as a single-tree point cloud entity containing complete spatial structure and spectral information, which not only fundamentally eliminates the influence of terrain, but also accurately obtains three-dimensional structure parameters of single trees, and a two-stage strategy of geometric rough segmentation first and spectral refinement later is creatively proposed, the deep fusion perfectly combines the macro-accuracy of geometric segmentation and the micro-identification of spectral segmentation, and the segmentation accuracy of adhered tree crowns is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a forestry remote sensing image recognition method and system based on image recognition. BACKGROUND

[0002] Forestry, as a core component of the ecological system, undertakes key functions such as carbon sink regulation, biodiversity protection, soil and water conservation, and timber resource supply. Its dynamic changes directly affect global ecological security and sustainable development. With the intensification of global climate change and the increasing human activities that interfere with forests, real-time, accurate, and large-scale forestry monitoring has become an urgent need. From an ecological protection perspective, it is necessary to quickly grasp information such as forest coverage changes, vegetation growth conditions, and endangered species habitat distribution. From a resource management perspective, it is necessary to accurately count timber reserves, classify forest types (such as coniferous and broadleaf forests), and monitor illegal logging. From a disaster prevention and control perspective, it is necessary to identify forest fire hotspots and pest infection areas in a timely manner to provide a basis for emergency response.

[0003] At present, the Chinese invention patent with application number CN202411903952.4 discloses a forestry surveying and mapping method based on remote sensing big data, which includes the following steps: integrating remote sensing images, unmanned aerial vehicle images, temperature sensor and humidity sensor data for spatio-temporal alignment processing, and filling in missing data through time interpolation while detecting outliers, combining the processed data to calculate the ecological indicators of the forest area, and outputting the forest health ecological assessment results. This method integrates remote sensing images, unmanned aerial vehicle images, and temperature and humidity sensor data to improve the accuracy and completeness of the data, species identification and ecological function contribution analysis, which can accurately identify key species and their contribution to the ecological system. By analyzing the dynamic change trend of forest resources and the influence of climate, it can effectively optimize resource allocation, improve management efficiency and risk control ability, and improve the precision and timeliness of forest resource monitoring, evaluation and management, and enhance the sustainable utilization and ecological protection ability. However, most of the existing technologies are based on two-dimensional images for tree crown segmentation, which is easily affected by shadows, tree crown overlapping and terrain undulations, resulting in low segmentation accuracy and missing tree height information. Or it only relies on the geometric shape of point cloud or only relies on the spectral information of image, and the combination of the two is not sufficient. In complex forest stands and tree crown adhesion areas, the effect is not good, and it often stops at extracting single tree parameters, lacking structural analysis of the forest as a whole ecological system. SUMMARY

[0004] The technical problem solved by the present application is that most of the prior art is based on two-dimensional images for crown segmentation, which is easily affected by shadows, overlapping tree crowns and terrain undulations, resulting in low segmentation accuracy and missing tree height information, or only relying on the geometric morphology of point clouds or only relying on the spectral information of images, and the combination of the two is insufficient, which is not effective in complex forest stands and tree crown adhesion areas, and often stops at extracting single-tree parameters, lacking structural analysis of forests as a whole ecosystem.

[0005] To solve the above technical problems, the present application provides the following technical solutions: a forestry remote sensing image recognition method based on image recognition, comprising the following steps:

[0006] Step S1: collecting LiDAR point clouds and multispectral optical images within a preset time window by a UAV, pre-processing the LiDAR point clouds and multispectral optical images, and obtaining three-dimensional tree point clouds;

[0007] Step S2: performing tree feature segmentation through spectral difference analysis, reassigning tree feature segmentation boundary points based on spectral dissimilarity, calculating the core spectral signature of each initial cluster ID of the three-dimensional tree point clouds after reassignment, and outputting a single-tree point cloud entity set;

[0008] Step S3: outputting a vector result according to the single-tree point cloud entity set and archiving the center database.

[0009] Preferably, the step S1 comprises:

[0010] The collection of LiDAR point clouds and multispectral optical images is completed by a UAV within a preset time window, the digital quantization values of the LiDAR point clouds and multispectral optical images are converted into radiance values, and the radiance values are converted into surface reflectivity through QUAC;

[0011] The three-dimensional coordinates of each laser pulse return point are calculated from the POS data, and the three-dimensional coordinates are assigned to the geographic coordinates to generate original point cloud data;

[0012] The original point cloud data is separated into ground points and non-ground points through a progressive TIN encryption filtering algorithm, a digital terrain model is generated based on the separated ground points through interpolation, the digital terrain model elevation value at the corresponding position of each non-ground point is subtracted from the elevation value of the non-ground point, and three-dimensional tree point clouds are obtained, wherein the Z value of the three-dimensional tree point clouds represents the true height of the tree from the ground.

[0013] Preferably, the step S2 comprises:

[0014] The multispectral optical image after data conversion is projected onto the three-dimensional tree point clouds, and the three-dimensional tree point clouds are given the characteristic attribute of spectral band value;

[0015] Setting a horizontal resolution for the three-dimensional forest point cloud, dividing the involved area of the three-dimensional forest point cloud into a two-dimensional grid on the X-Y plane, respectively calculating the maximum Z coordinate value of all three-dimensional forest point cloud points in each cell of the two-dimensional grid, and assigning the maximum Z coordinate value to the cell;

[0016] For all null cells without point cloud data, using inverse distance weighted interpolation method to fill the values of adjacent cells of the null cells, filling the two-dimensional grid data to obtain a crown height grid data grid map;

[0017] Performing a value inversion operation on the crown height grid data grid map to generate a new grid data grid map, taking the grid data grid map after value inversion as input, and generating a preliminary segmentation region grid data through a standard watershed segmentation algorithm, each pixel in the preliminary segmentation region grid data being provided with an integer segmentation region ID;

[0018] Traversing each point in the three-dimensional forest point cloud, locating the corresponding pixel of the point in the preliminary segmentation region grid data according to the two-dimensional plane coordinates of the point, and assigning the segmentation region ID of the pixel as an initial cluster ID to the point, and outputting the forest point cloud with the initial cluster ID after completing the initial cluster ID assignment operation.

[0019] Preferably, the reassignment process of the forest feature segmentation boundary point based on the spectral dissimilarity degree comprises:

[0020] Assigning a preset horizontal distance threshold to the forest point cloud with the initial cluster ID, traversing each point in the forest point cloud with the initial cluster ID, searching for points satisfying the condition including different initial cluster ID carried by the current traversed point and horizontal distance less than the horizontal distance threshold, and adding a boundary point type label to the current traversed point when the point satisfying the condition is searched;

[0021] When the point satisfying the condition cannot be searched, a core point type label is added to the current traversed point;

[0022] Outputting the forest point cloud with the initial cluster ID and the boundary point type label or the core point type label.

[0023] Preferably, the calculation process of the core spectral signature of each initial cluster ID comprises:

[0024] Group all points of the three-dimensional forest point cloud according to initial cluster IDs, extract all points of each group marked as core points of all types, respectively calculate the average values of the spectral indices of NDVI, PSRI and MCARI of all core points marked as points of all types, and output the three-dimensional vector [mean(NDVI), mean(PSRI), mean(MCARI)] as the core spectral signature vector, wherein mean() represents the average value, and the core spectral signature vector of all core points marked as points of all types is collected as a core spectral signature vector table, each initial cluster ID of the core spectral signature vector table respectively maps a corresponding core spectral signature vector;

[0025] Extract all points of each group marked as boundary points of all types, and respectively perform re-assignment of the points, and the re-assignment includes:

[0026] Step S21: Obtain the spectral vector of the point marked as the boundary point of the type, and the mathematical expression of the spectral vector is:

[0027] ;

[0028] Wherein, is the spectral vector of the point p, is the NDVI spectral index of the point p, is the of the point p, is the MCARI spectral index of the point p, and the current initial cluster ID to which the point p belongs and the IDs of all spatially adjacent clusters of the point p are identified;

[0029] Step S22: Query the core spectral signature vector corresponding to the current initial cluster ID to which the point p belongs and the IDs of all spatially adjacent clusters of the point p from the core spectral signature vector table;

[0030] Step S23: Calculate the multi-dimensional Euclidean distance between the spectral vector of the point p and each core spectral signature vector queried in step S22, update the initial cluster ID of the point p to the initial cluster ID corresponding to the core spectral signature vector with the smallest multi-dimensional Euclidean distance, and output the updated three-dimensional forest point cloud.

[0031] Preferably, the output single tree point cloud entity set comprises:

[0032] A hash table is created, the key of the hash table represents a unique initial cluster ID, and the value represents a point list, each point in the updated three-dimensional forest point cloud is traversed, the initial cluster ID of the point is read, and the traversed and read initial cluster ID is added to the point list corresponding to the cluster ID in the hash table, after the traversal is completed, each key-value pair in the hash table respectively corresponds to an independent single tree point cloud entity;

[0033] Each single-tree point cloud entity is a point cloud set, representing an independent tree with homogeneity in space and spectrum;

[0034] Each point in the point cloud set included in the single-tree point cloud entity includes a multi-attribute feature vector, and a mathematical expression of the multi-attribute feature vector is:

[0035] ;

[0036] Wherein, is a multi-attribute feature vector of point p, 、 and is a three-dimensional coordinate value of point p, is a spectral band 1 to a spectral band n, wherein n is a natural number greater than 1.

[0037] Preferably, the convex hull of each single-tree point cloud entity in a two-dimensional plane is calculated respectively to generate a two-dimensional polygon corresponding to the crown shape of the single-tree point cloud entity, the two-dimensional polygon includes a vector file containing an independent crown, the vector file is subjected to contour tracking and hierarchical analysis, forest structure indicators are derived, forest gaps are identified, canopy density is measured, and spatial arrangement and adjacency relationship of trees are analyzed, and the forest gap is a region not covered by any crown polygon;

[0038] Each single-tree point cloud entity is projected in two dimensions to obtain a two-dimensional point set, and a convex hull is calculated on the two-dimensional point set, and the calculated convex hull boundary vertices are sequentially connected to generate a closed two-dimensional polygon;

[0039] The two-dimensional polygon is associated with the corresponding initial cluster ID and other calculated feature attributes in the corresponding point, and the other calculated feature attributes include NDVI spectral index, PSRI spectral index, MCARI spectral index, three-dimensional coordinate value and spectral band, a crown vector file including all independent crown polygons is obtained, each element in the file represents a crown of a tree, and attribute information of the tree is attached.

[0040] Preferably, the single-tree point cloud entity set is topologically analyzed, and a vector result is output, the vector result including a negative space region, a canopy density, a distribution pattern of trees and a single-tree adjacency graph;

[0041] The topological analysis process includes:

[0042] All two-dimensional polygons are subjected to set operation to obtain a two-dimensional large polygon including multiple unconnected parts, and the two-dimensional large polygon represents all regions covered by crowns in a region involved in the entire three-dimensional forest point cloud;

[0043] Subtracting the two-dimensional large polygon from the boundary polygon of the involved area of the entire three-dimensional forest point cloud, a negative space area not covered by any tree crown is obtained;

[0044] Calculating the total area of the two-dimensional large polygon and the total area of the involved area of the entire three-dimensional forest point cloud, and obtaining the canopy density by calculating the ratio of the total area of the two-dimensional large polygon to the total area of the involved area of the entire three-dimensional forest point cloud;

[0045] Obtaining the distance between the nearest neighbor trees by calculating the distance between the centroid of the two-dimensional polygon corresponding to the single tree and the centroid of the nearest neighbor tree, calculating the expected average value of the distance between the nearest neighbor trees under random distribution, and judging the distribution pattern of the trees, the distribution pattern including clustered distribution, uniform distribution and random distribution, and the judgment logic including:

[0046] When the expected average value is greater than the random distribution threshold value, the distribution pattern of the trees is random distribution;

[0047] When the expected average value is greater than the uniform distribution threshold value and less than the random distribution threshold value, the distribution pattern of the trees is uniform distribution;

[0048] When the expected average value is less than the uniform distribution threshold value, the distribution pattern of the trees is clustered distribution;

[0049] Regarding the two-dimensional polygon of each single tree as a node in the single tree adjacency graph, judging whether the two-dimensional polygons of the adjacent single trees have contact or overlap in space through the intersection test of GIS, and drawing an edge between the corresponding adjacent single tree nodes if there is contact or overlap.

[0050] Preferably, attribute groups are calculated and linked for each vectorized tree crown through the tree crown vector file, the attribute groups including geometric attributes, spectral attributes and derived attributes, the geometric attributes including crown area, perimeter, tree height and geographic position;

[0051] The spectral attributes include spectral indices, PSRI spectral indices, MCARI spectral indices and spectral bands;

[0052] The derived attributes include single tree health and tree species predicted based on the tree crown vector file through CNN;

[0053] The tree crown vector file, the vector result, the final output three-dimensional forest point cloud and the single tree point cloud entity set are stored back to the central database.

[0054] A forestry remote sensing image recognition system based on image recognition, which is used to execute a forestry remote sensing image recognition method based on image recognition, and includes an image acquisition module, a feature segmentation optimization module and an analysis application module:

[0055] The image acquisition module is used for acquiring LiDAR point cloud and multispectral optical image within a preset time window by a UAV, performing data preprocessing on the LiDAR point cloud and the multispectral optical image, and obtaining three-dimensional forest point cloud;

[0056] The feature segmentation optimization module is used for performing forest feature segmentation through spectral difference analysis, reassigning forest feature segmentation boundary points based on spectral dissimilarity, calculating core spectral signatures of each initial cluster ID of the three-dimensional forest point cloud after reassignment, and outputting a single tree point cloud entity set;

[0057] The analysis application module is used for outputting a vector result and archiving a central database according to the single tree point cloud entity set.

[0058] The present application directly operates three-dimensional forest point cloud, identifies each tree as a single tree point cloud entity containing complete spatial structure and spectral information, fundamentally eliminates the influence of terrain, and accurately obtains three-dimensional structure parameters of single trees, thereby laying a solid foundation for more accurate biomass estimation and carbon storage calculation. A two-stage strategy of preliminary geometric segmentation and then spectral refinement is creatively proposed. The three-dimensional geometric advantage of point cloud is used to perform rapid and robust preliminary segmentation and frame a general range. Then, multispectral information is used to perform point-by-point reassignment of the most difficult boundary points. This deep fusion perfectly combines the macro-accuracy of geometric segmentation and the micro-identification of spectral segmentation, greatly improving the segmentation accuracy of adhered crowns. On the basis of accurate single tree segmentation, GIS topological analysis is introduced to systematically integrate discrete single tree information and automatically interpret key stand structure indexes such as forest gap, canopy density, spatial distribution pattern and adjacency relationship, which enables analysis to rise from tree recognition to forest understanding, and provides unprecedented quantitative and spatial decision basis for forest management, ecological health evaluation and biodiversity research. A full-process automatic and closed-loop system from data acquisition, feature segmentation to analysis application and finally archiving a central database is constructed. This design not only improves processing efficiency, but more importantly, ensures data standardization, process traceability and result reusability, and provides a solid technical platform for time-series dynamic monitoring and big data analysis of forestry resources. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A basic flowchart of a forestry remote sensing image recognition method for image recognition is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0060] To make the above objectives, features and advantages of the present application more apparent and understandable, specific embodiments of the present application are described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the present application, rather than all embodiments.

[0061] Referring to Figure 1 For an embodiment of the present application, an image recognition-based forestry remote sensing image recognition method is provided, comprising the following steps:

[0062] Step S1: Collect LiDAR point cloud and multispectral optical image within a preset time window by a UAV, perform data preprocessing on the LiDAR point cloud and multispectral optical image, and obtain three-dimensional tree point cloud.

[0063] Step S2: Perform tree feature segmentation through spectral difference analysis, reassign tree feature segmentation boundary points based on spectral dissimilarity, calculate the core spectral signature of each initial cluster ID of the three-dimensional tree point cloud after reassignment, and output a single tree point cloud entity set.

[0064] Step S3: Output a vector result according to the single tree point cloud entity set, and back up the central database.

[0065] The present application identifies each tree as a single tree point cloud entity containing complete spatial structure and spectral information through direct three-dimensional single tree entity analysis, which not only fundamentally eliminates the influence of terrain, but also accurately obtains three-dimensional structure parameters of single trees, initiatively proposes a two-stage strategy of geometric rough segmentation followed by spectral refinement, and perfectly combines the macro-accuracy of geometric segmentation and the micro-identification of spectral segmentation through deep fusion, greatly improving the segmentation accuracy of adhered tree crowns. The vector result generated through topological analysis can be simply applied by users, and the central database is finally backed up in a full-process automatic and closed-loop manner.

[0066] Step S1 comprises:

[0067] The collection of LiDAR point cloud and multispectral optical image within a preset time window by a UAV converts the digital quantization values of LiDAR point cloud and multispectral optical image into radiance values, and converts the radiance values into surface reflectivity through QUAC.

[0068] The three-dimensional coordinates of each laser pulse return point are calculated from the POS data, and the three-dimensional coordinates are assigned to geographic coordinates to generate original point cloud data.

[0069] The original point cloud data is separated into ground points and non-ground points through a progressive TIN encryption filtering algorithm, a digital terrain model is generated based on the separated ground points through interpolation, the digital terrain model elevation value at the corresponding position of the non-ground point is subtracted from the elevation value of all non-ground points, and three-dimensional tree point cloud is obtained, with the Z value of the three-dimensional tree point cloud representing the true height of the tree from the ground.

[0070] In this embodiment, the time window is set to be no more than one hour apart within the same day, which minimizes the illumination difference, shadow difference and the intra-day fluctuation of forest physiological state caused by the change of solar elevation angle. The optical image needs to ensure high overlap rate, and the overlap rate is set to 80% to support the generation of high-quality orthographic image. The LiDAR needs to ensure sufficient lateral overlap, and the lateral overlap rate is set to 50% to ensure the completeness of the data coverage of the crown side and bottom, and avoid data voids.

[0071] The matching and adjustment between the flight strips of the unmanned aerial vehicle eliminate the systematic height or plane error between the strips through the ICP algorithm, and ensure the internal consistency of the point cloud in the whole study area.

[0072] The height value is the z coordinate in the three-dimensional coordinate, and the influence of the terrain undulation on the tree height is eliminated by subtracting the height value.

[0073] Step S2 comprises:

[0074] The multispectral optical image after data conversion is projected onto the three-dimensional forest point cloud, and the three-dimensional forest point cloud is given the characteristic attribute of the spectral band value;

[0075] The horizontal resolution of the three-dimensional forest point cloud is set, and in this embodiment, the horizontal resolution is set to 0.5 meters. The involved area of the three-dimensional forest point cloud is divided into a two-dimensional grid in the X-Y plane, the maximum Z coordinate value of all three-dimensional forest point cloud points in each cell in the two-dimensional grid is calculated respectively, and the maximum Z coordinate value is given to the cell.

[0076] For all null cells without point cloud data, the values of the adjacent cells of the null cells are used to fill the null cells by using the inverse distance weighted interpolation method. After the two-dimensional grid data is filled, the crown height grid data grid map is obtained.

[0077] The crown height grid data grid map is subjected to value inversion operation to generate a new grid data grid map. The grid data grid map after value inversion is taken as input, and a preliminary segmentation area grid data is generated by using a standard watershed segmentation algorithm. Each pixel in the preliminary segmentation area grid data has a segmentation area ID with an integer value.

[0078] Each point in the three-dimensional forest point cloud is traversed, and the two-dimensional plane coordinates of the point are located to the corresponding pixel of the point in the preliminary segmentation area grid data. The segmentation area ID of the pixel is taken as the initial cluster ID of the point. After the initial cluster ID assignment operation is completed, the forest point cloud with the initial cluster ID is output.

[0079] The crown height grid data grid map is used to represent the crown height of the forest;

[0080] This stage is a geometric rough division stage, which can intuitively reflect the advantages of the top shape of the crown, quickly find the top of each crown through the watershed algorithm, and preliminarily divide the center. It has fast calculation speed, strong robustness, can quickly assign an initial single tree attribution ID to massive point cloud data, provides an efficient starting point for subsequent refinement processing, and greatly reduces the calculation complexity.

[0081] The process of reassigning the boundary points of the tree feature based on spectral dissimilarity includes:

[0082] Assign a preset horizontal distance threshold to the tree point cloud with an initial cluster ID, and set the horizontal distance threshold to 1 meter in this embodiment. Traverse each point in the tree point cloud with an initial cluster ID, search for points that meet the conditions including different initial cluster ID and horizontal distance less than the horizontal distance threshold, and add a boundary point type label to the current traversal point when the points that meet the conditions are searched.

[0083] When the points that meet the conditions cannot be searched, a core point type label is added to the current traversal point.

[0084] Output the tree point cloud with initial cluster ID and boundary point type label or core point type label.

[0085] This is the key premise of refinement processing, which ingeniously simplifies the problem from processing all points to only processing the most ambiguous and most error-prone boundary points, greatly improving the calculation efficiency and pertinence, and the core point is stable and reliable without processing; only a few boundary points are re-decided, avoiding unnecessary computational overhead.

[0086] The calculation process of the core spectral signature of each initial cluster ID includes:

[0087] Group all points of the three-dimensional tree point cloud according to the initial cluster ID, extract points of all types labeled as core points in each group, calculate the average values of the spectral indices of NDVI, PSRI and MCARI of the extracted points of all core point type labeled points, and output the three-dimensional vector [mean(NDVI), mean(PSRI), mean(MCARI)] as the core spectral signature vector. Wherein mean() represents the average value, the core spectral signature vector of all core point type labeled points is collected as a core spectral signature vector table, and each initial cluster ID of the core spectral signature vector table is mapped to a corresponding core spectral signature vector.

[0088] By calculating the spectral average of each cluster only using the core points, a pure and stable spectral representation of each potential tree crown is obtained, the core points are far away from the mixed area, and their spectral characteristics best represent the true situation of the tree, avoiding being contaminated by the spectra of adjacent trees, providing an extremely reliable reference benchmark for subsequent boundary point judgment.

[0089] Extracting all types of points marked as boundary points in each group, respectively, and performing re-assignment of points, including:

[0090] Step S21: Obtain the spectral vector of the point marked as the boundary point, and the mathematical expression of the spectral vector is:

[0091] ;

[0092] Wherein, is the spectral vector of point p, is the NDVI spectral index of point p, is the PSRI spectral index of point p, is the MCARI spectral index of point p, and the ID of the initial cluster to which point p currently belongs and the ID of all spatially adjacent clusters of point p are identified.

[0093] Wherein, the NDVI spectral index is the normalized vegetation index, the PSRI spectral index is the vegetation senescence reflectance index, and the MCARI spectral index is the modified chlorophyll absorption reflectance index.

[0094] Step S22: Query the core spectral signature vector corresponding to the ID of the initial cluster to which point p currently belongs and the ID of all spatially adjacent clusters of point p from the core spectral signature vector table;

[0095] Step S23: Calculate the multi-dimensional Euclidean distance between the spectral vector of point p and each core spectral signature vector queried in step S22, update the initial cluster ID of point p to the initial cluster ID corresponding to the core spectral signature vector with the smallest multi-dimensional Euclidean distance, and output the updated three-dimensional forest point cloud.

[0096] Comparing the Euclidean distance between the boundary point's own spectrum and the core spectrum signature of the surrounding cluster one by one, the pixel-level accurate correction of the adhesion tree crown boundary is realized, the spectral characteristics of a boundary point located between two trees must be closer to the core spectrum of the tree it really belongs to, and through this spectral similarity voting method, the boundary point incorrectly divided by the geometric algorithm can be pulled back to the correct tree crown entity, greatly improving the segmentation accuracy.

[0097] The output single-tree point cloud entity set includes:

[0098] A hash table is created, the key of the hash table represents a unique initial cluster ID, and the value represents a point list, each point in the updated three-dimensional forest point cloud is traversed, the initial cluster ID of the point is read, the read initial cluster ID is added to the point list corresponding to the cluster ID in the hash table, and after the traversal is completed, each key-value pair in the hash table corresponds to an independent single tree point cloud entity respectively;

[0099] Each single tree point cloud entity is a point cloud set, representing an independent tree with homogeneity in space and spectrum;

[0100] Each point in the point cloud set included in the single tree point cloud entity includes a multi-attribute feature vector, and the mathematical expression of the multi-attribute feature vector is:

[0101] ;

[0102] Wherein, is the multi-attribute feature vector of point p, , and is the three-dimensional coordinate value of point p, is the spectral band 1 to the spectral band n, wherein n is a natural number greater than 1.

[0103] The point cloud is aggregated into an entity with a multi-attribute feature vector, completing the conversion from discrete points to structured objects. The output is no longer scattered points, but organized and attribute digital twin trees, each point carrying rich information, making complete data preparation for subsequent vectorization and analysis.

[0104] The convex hull of each single tree point cloud entity in the two-dimensional plane is calculated respectively to generate a two-dimensional polygon corresponding to the crown shape of the single tree point cloud entity. The two-dimensional polygon includes a vector file describing an independent crown. The vector file is subjected to contour tracking and hierarchical analysis to derive stand structure indicators, identify forest windows, measure canopy density, and analyze the spatial arrangement and adjacency relationship of trees. The forest window is an area not covered by any crown polygon;

[0105] Each single tree point cloud entity is projected in two dimensions to obtain a two-dimensional point set, and the convex hull is calculated on the two-dimensional point set. The calculated convex hull boundary vertices are sequentially connected to generate a closed two-dimensional polygon.

[0106] The two-dimensional polygon is associated with the corresponding initial cluster ID and other calculated feature attributes in the corresponding point, and the other feature attributes include the NDVI spectral index, the PSRI spectral index, the MCARI spectral index, the three-dimensional coordinate value and the spectral band. A crown vector file including all independent crown polygons is obtained, and each element in the file represents the crown of a tree and is attached with attribute information of the tree.

[0107] This step is to project and abstract the three-dimensional point cloud entity into a two-dimensional crown polygon, which is regarded as a feature in GIS.

[0108] The convex hull is the smallest convex polygon that contains all the points. The convex hull algorithm is fast in calculation and stable in result, and can well outline the general outline of the crown.

[0109] This stage converts the refined point cloud cluster into a standard GIS format and performs higher-level spatial analysis. The core task of this stage is to convert the discrete and accurately classified single tree point cloud cluster output from the previous stage into structured vector data that can be analyzed by a geographic information system, and on this basis, extract higher-level forest structure information.

[0110] Convex hull vectorization projects each single tree entity into a two-dimensional polygon, realizing seamless connection with the standard GIS platform. The vector polygon is the basic unit of GIS analysis, facilitating area and perimeter calculation, and providing an object basis for subsequent topological analysis. The convex hull algorithm takes into account both calculation efficiency and outline accuracy.

[0111] Step S3 includes:

[0112] Topological analysis of single tree point cloud entity set, output vector result, vector result includes negative space area, canopy density, tree distribution pattern and single tree adjacency graph;

[0113] The topological analysis process includes:

[0114] Perform a union operation on all two-dimensional polygons to obtain a two-dimensional large polygon that includes multiple unconnected parts, which represents all areas covered by tree crowns within the entire three-dimensional forest point cloud involved area;

[0115] Subtract the two-dimensional large polygon from the boundary polygon of the entire three-dimensional forest point cloud involved area to obtain the negative space area not covered by any tree crown;

[0116] Calculate the total area of the two-dimensional large polygon and the total area of the entire three-dimensional forest point cloud involved area, and obtain the canopy density by calculating the ratio of the total area of the two-dimensional large polygon to the total area of the entire three-dimensional forest point cloud involved area. The canopy density index macroscopically reflects the density of the forest stand;

[0117] Calculate the distance between the centroids of the two-dimensional polygons corresponding to the single trees and the centroids of the nearest neighbor trees to obtain the distance between the nearest neighbor single trees, calculate the expected average value of the distance between the nearest neighbor single trees under random distribution, and judge the tree distribution pattern. The distribution pattern includes clustered distribution, uniform distribution and random distribution, and the judgment logic includes:

[0118] When the expected mean is greater than the random distribution threshold, the distribution pattern of the trees is random distribution;

[0119] When the expected mean is greater than the uniform distribution threshold and less than the random distribution threshold, the distribution pattern of the trees is uniform distribution;

[0120] When the expected mean is less than the uniform distribution threshold, the distribution pattern of the trees is aggregated distribution;

[0121] The two-dimensional polygon of each individual tree is regarded as a node in the individual tree adjacency graph, and the intersection test of GIS is used to determine whether the two-dimensional polygons of adjacent individual trees have contact or overlap in space. If there is contact or overlap, a line is drawn between the corresponding adjacent individual tree nodes.

[0122] This adjacency graph intuitively shows which trees are next to each other. Based on this graph, the degree of each node, i.e. how many neighbors each tree has, can be calculated, thereby analyzing the competition relationship between trees;

[0123] Through the above detailed steps, we successfully convert point cloud data into GIS data with rich spatial and attribute information, and on this basis, we extract structural indicators that are crucial to ecology and forestry management from the individual level (individual tree) to the group level (forest stand).

[0124] Forest gap, canopy density, spatial pattern, and adjacency relationship analysis provide systematic forest insights beyond the individual tree scale. These indicators are core parameters in forest ecology and management. Gap identification helps to assess land use efficiency and habitat fragmentation. Canopy density is a key indicator for evaluating stand density and light competition. Spatial distribution pattern reveals intra- and interspecific relationships and environmental adaptation strategies of trees. The individual tree adjacency graph quantifies the direct competition relationship between trees and can be used to guide tending and thinning operations.

[0125] This step is the culmination of the entire process, shifting from analyzing individual trees to analyzing the structure of the forest composed of these trees. Here we will apply the ideas of contour tracing and hierarchical analysis, but the objects of action are no longer pixels, but a collection of crown polygons.

[0126] For each vectorized crown, calculate and link attribute groups through the crown vector file. Attribute groups include geometric attributes, spectral attributes, and derived attributes. Geometric attributes include crown area, perimeter, tree height, and geographic location;

[0127] Spectral attributes include spectral indices, PSRI spectral indices, MCARI spectral indices, and spectral bands;

[0128] Derived attributes include individual tree health and tree species predicted based on the crown vector file through CNN;

[0129] The crown vector file, the vector result, the final output three-dimensional forest point cloud and the single tree point cloud entity set are stored back to the central database.

[0130] This final stage assembles all the derived information and archives it back to the central database. This workflow provides a robust, end-to-end solution that is both fast, thanks to the efficient watershed segmentation and refinement algorithms, and highly accurate, by combining three-dimensional geometry with adaptive spectral refinement and advanced feature prediction, building a circular, iterative, and traceable intelligent forestry information system. All data and results are source-traceable, facilitating long-term dynamic change monitoring (such as analyzing the changes of forest gaps and the growth of trees over several years), and providing high-quality, structured data sets for training more powerful AI models (such as predicting tree species and health status using CNN).

[0131] An image recognition-based forestry remote sensing image recognition system, which is used to execute an image recognition-based forestry remote sensing image recognition method, comprising an image acquisition module, a feature segmentation optimization module and an analysis application module:

[0132] The image acquisition module is used to acquire LiDAR point cloud and multispectral optical image within a preset time window through a UAV, to perform data preprocessing on the LiDAR point cloud and the multispectral optical image, and to obtain three-dimensional forest point cloud;

[0133] The feature segmentation optimization module is used to perform forest feature segmentation through spectral difference analysis, to redistribute forest feature segmentation boundary points based on spectral dissimilarity, to calculate core spectral signatures of each initial cluster ID of the three-dimensional forest point cloud after redistribution, and to output a single tree point cloud entity set;

[0134] The analysis application module is used to output a vector result according to the single tree point cloud entity set, and to archive the central database.

[0135] The present application directly operates three-dimensional forest point cloud, identifies each tree as a single tree point cloud entity containing complete spatial structure and spectral information, which not only fundamentally eliminates the influence of terrain, but also accurately obtains three-dimensional structure parameters of single trees, lays a solid foundation for more accurate biomass estimation and carbon storage calculation. A two-stage strategy of geometric rough division first and spectral refinement later is proposed, which firstly uses the three-dimensional geometric advantage of point cloud for fast and robust preliminary segmentation to frame the approximate range, and then uses multispectral information to make point-by-point re-decision for the most difficult boundary points. This deep fusion perfectly combines the macro-accuracy of geometric segmentation and the micro-recognition of spectral segmentation, greatly improving the segmentation accuracy of the adhered crown. On the basis of accurate segmentation of single trees, GIS topological analysis is introduced to systematically integrate the discrete single tree information and automatically interpret key forest structure indicators such as forest gap, canopy density, spatial distribution pattern and adjacency relationship, which makes the analysis rise from tree recognition to forest understanding and provides unprecedented quantitative and spatial decision basis for forest management, ecological health evaluation and biodiversity research. A full-process automatic and closed-loop system from data acquisition, feature segmentation to analysis application and finally returning to the central database is constructed. This design not only improves the processing efficiency, but more importantly, ensures the standardization of data, the traceability of process and the reusability of results, and provides a solid technical platform for the time-series dynamic monitoring and big data analysis of forestry resources.

[0136] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (or computer- readable storage media) having computer-usable program code embodied in the medium. The medium can be any available storage media that can be accessed by a computer. By way of example, and not limitation, such computer-usable storage media can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other storage medium(s) that can be used to carry or store desired computer program code in the form of instructions or data structures and that can be accessed by a computer. Also, the present application can be embodied in a computer program product that can be traded as goods or merchandise, through the storage medium described above or any other suitable medium. Accordingly, the computer medium can be any entity or device containing, or Figure 1 the functions specified in the flow or flows and / or blocks Figure 1 the functions specified in the flow or flows and / or blocks

[0137] It should be noted that the above-mentioned embodiments are only used to illustrate but not to limit the technical solutions of the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. An image recognition-based forestry remote sensing image recognition method, characterized in that, The method comprises the following steps: Step S1: collecting LiDAR point cloud and multi-spectral optical image within a preset time window by a UAV, pre-processing the LiDAR point cloud and multi-spectral optical image, and obtaining three-dimensional forest point cloud; Step S2: performing forest feature segmentation by spectral difference analysis, re-distributing forest feature segmentation boundary points based on spectral dissimilarity, calculating core spectral signatures of each initial cluster ID of the three-dimensional forest point cloud after re-distribution, and outputting a single tree point cloud entity set; Step S3: outputting a vector result according to the single tree point cloud entity set, and archiving a central database; The step S1 comprises: Converting digital quantization values of the LiDAR point cloud and multi-spectral optical image into radiance values, and converting the radiance values into ground reflectivity by QUAC; Solving three-dimensional coordinates of each laser pulse return point by POS data, and generating original point cloud data by assigning the three-dimensional coordinates to geographic coordinates; Separating the original point cloud data into ground points and non-ground points by a progressive TIN encryption filtering algorithm, generating a digital terrain model based on the separated ground points, subtracting the digital terrain model elevation value of the corresponding position of all non-ground points from the elevation value of the non-ground points, and obtaining three-dimensional forest point cloud, wherein the Z value of the three-dimensional forest point cloud represents the true height of the forest from the ground.

2. The image recognition-based forestry remote sensing image recognition method of claim 1, wherein, The step S2 comprises: Projecting the multi-spectral optical image after data conversion onto the three-dimensional forest point cloud, and giving the three-dimensional forest point cloud a characteristic attribute of spectral band value; Setting a horizontal resolution for the three-dimensional forest point cloud, dividing the involved area of the three-dimensional forest point cloud into a two-dimensional grid on the X-Y plane, calculating the maximum Z coordinate value of all three-dimensional forest point cloud points in each cell in the two-dimensional grid, and assigning the maximum Z coordinate value to the cell; Filling the value of the adjacent cell using the inverse distance weighted interpolation method for all null value cells without point cloud data, and obtaining a crown height grid data mesh graph after filling the two-dimensional grid data; Performing a value inversion operation on the crown height grid data mesh graph to generate a new grid data mesh graph, taking the grid data mesh graph after value inversion as input, and generating preliminary segmentation area grid data by a standard watershed segmentation algorithm, wherein each pixel in the preliminary segmentation area grid data has an integer value of a segmentation area ID; Traversing each point in the three-dimensional forest point cloud, locating the corresponding pixel of the point in the preliminary segmentation area grid data according to the two-dimensional plane coordinates of the point, and assigning the segmentation area ID of the pixel as an initial cluster ID to the point, and outputting the forest point cloud with the initial cluster ID after completing the initial cluster ID assignment operation.

3. The image recognition-based forestry remote sensing image recognition method of claim 2, wherein, The process of re-distributing forest feature segmentation boundary points based on spectral dissimilarity comprises: Assign a preset horizontal distance threshold to the forest point cloud with initial cluster ID, traverse each point in the forest point cloud with initial cluster ID, search for points that meet the conditions including being different from the initial cluster ID carried by the current traversed point and the horizontal distance of the straight line being less than the horizontal distance threshold, when the points meeting the conditions are searched, a boundary point type label is added to the current traversed point; When the points meeting the conditions cannot be searched, a core point type label is added to the current traversed point; Output the forest point cloud with initial cluster ID and boundary point type label or core point type label.

4. The image recognition-based forestry remote sensing image recognition method of claim 3, wherein, The calculation process of the core spectral signature of each initial cluster ID includes: Group all points of the three-dimensional forest point cloud according to the initial cluster ID, extract points of all types labeled as core points in each group, calculate the average values of the spectral indices of NDVI, PSRI and MCARI of the extracted points of all types labeled as core points respectively, and output the three-dimensional vector [mean(NDVI), mean(PSRI), mean(MCARI)] as the core spectral signature vector, wherein mean() represents the average value, and the core spectral signature vectors of all points of all types labeled as core points are collected into a core spectral signature vector table, each initial cluster ID of the core spectral signature vector table is mapped to a corresponding core spectral signature vector; Extract points of all types labeled as boundary points in each group, and perform rejudgment of attribution on the points respectively, which includes: Step S21: Obtain the spectral vector of the point of the type labeled as a boundary point, and the mathematical expression of the spectral vector is: ； wherein, is the spectral vector of point p, is the NDVI spectral index of point p, is the PSRI spectral index of point p, is the MCARI spectral index of point p, identifies the initial cluster ID to which the point p currently belongs and the IDs of all spatially neighboring clusters of the point p; Step S22: Query the core spectral signature vectors corresponding to the initial cluster ID to which the point p currently belongs and the IDs of all spatially adjacent clusters of the point p from the core spectral signature vector table; Step S23: Calculate the multidimensional Euclidean distance between the spectral vector of the point p and each core spectral signature vector queried in step S22, update the initial cluster ID of the point p to the initial cluster ID corresponding to the core spectral signature vector with the smallest multidimensional Euclidean distance, and output the updated three-dimensional forest point cloud.

5. The image recognition-based forestry remote sensing image recognition method of claim 4, wherein, The output single tree point cloud entity set includes: Create a hash table, the key of the hash table represents a unique initial cluster ID, and the value represents a point list, traverse each point in the updated three-dimensional forest point cloud, read the initial cluster ID of the point, add the traversed and read initial cluster ID to the point list corresponding to the cluster ID in the hash table, and after the traversal is completed, each key-value pair in the hash table corresponds to an independent single tree point cloud entity; Each single tree point cloud entity is a point cloud set, representing an independent tree with homogeneity in space and spectrum; Each point in the point cloud set included in the single tree point cloud entity includes a multi-attribute feature vector, and the mathematical expression of the multi-attribute feature vector is: ； wherein, is a multi-attribute feature vector for point p, , and is a three-dimensional coordinate value for point p, is a spectral band 1 to a spectral band n, where n is a natural number greater than 1.

6. The forestry remote sensing image recognition method based on image recognition according to claim 5, characterized in that: respectively, and the two-dimensional polygons include a vector file containing individual tree crowns, the vector file is subjected to contour tracing and hierarchical analysis to derive stand structure indicators, identify forest gaps, measure canopy density, and analyze spatial arrangement and adjacency of trees, and the forest gaps are regions not covered by any tree crown polygon; each individual tree point cloud entity is projected onto a two-dimensional plane to obtain a two-dimensional point set, a convex hull is calculated on the two-dimensional point set, and the calculated convex hull boundary vertices are sequentially connected to generate a closed two-dimensional polygon; the two-dimensional polygon is associated with the corresponding initial cluster ID and other calculated feature attributes in the corresponding point, and the other calculated feature attributes include NDVI spectral index, PSRI spectral index, MCARI spectral index, three-dimensional coordinate value, and spectral band, a tree crown vector file containing all individual tree crown polygons is obtained, each element in the file represents a tree crown of a tree and is accompanied by attribute information of the tree.

7. The image recognition-based forestry remote sensing image recognition method according to claim 6, characterized in that: topological analysis is performed on the individual tree point cloud entity set to output a vector result, and the vector result includes a negative space region, canopy density, distribution pattern of trees, and a tree adjacency graph; the topological analysis process includes: all two-dimensional polygons are subjected to a set operation to obtain a two-dimensional large polygon including multiple unconnected parts, and the two-dimensional large polygon represents all regions covered by tree crowns in a region of interest of the three-dimensional forest point cloud; the two-dimensional large polygon is subtracted from a boundary polygon of the region of interest of the three-dimensional forest point cloud to obtain a negative space region not covered by any tree crown; total areas of the two-dimensional large polygon and the region of interest of the three-dimensional forest point cloud are calculated, and the canopy density is obtained by calculating a ratio of the total area of the two-dimensional large polygon to the total area of the region of interest of the three-dimensional forest point cloud; the distance between nearest-neighbor trees is obtained by calculating a distance between a centroid of a two-dimensional polygon corresponding to an individual tree and a centroid of a nearest-neighbor tree, an expected average value of the distance between nearest-neighbor trees under random distribution is calculated, and a distribution pattern of trees is determined, and the distribution pattern includes clustered distribution, uniform distribution, and random distribution, and the determination logic includes: when the expected average value is greater than a random distribution threshold value, the distribution pattern of trees is random distribution; when the expected average value is greater than a uniform distribution threshold value and less than the random distribution threshold value, the distribution pattern of trees is uniform distribution; when the expected average value is less than the uniform distribution threshold value, the distribution pattern of trees is clustered distribution; a two-dimensional polygon of each individual tree is regarded as a node in a tree adjacency graph, and whether two-dimensional polygons of adjacent individual trees have contact or overlap in space is determined through GIS intersection testing, and if the two-dimensional polygons have contact or overlap, an edge is drawn between nodes of the corresponding adjacent individual trees.

8. The image recognition-based forestry remote sensing image recognition method according to claim 7, characterized in that: calculating and linking attribute groups for each vectorized tree crown through the tree crown vector file, the attribute groups including geometric attributes, spectral attributes and derived attributes, the geometric attributes including tree crown area, perimeter, tree height and geographic location; the spectral attributes including spectral indices, PSRI spectral indices, MCARI spectral indices and spectral bands; the derived attributes including single tree health and tree species predicted through CNN based on the tree crown vector file; storing the tree crown vector file, vector results, final output three-dimensional forest point cloud and single tree point cloud entity set back to the central database.

9. An image recognition-based forestry remote sensing image recognition system for performing an image recognition-based forestry remote sensing image recognition method, characterized by, comprising an image acquisition module, a feature segmentation optimization module and an analysis application module: the image acquisition module is configured to acquire LiDAR point cloud and multispectral optical image within a preset time window through a UAV, perform data preprocessing on the LiDAR point cloud and multispectral optical image, and obtain a three-dimensional forest point cloud; the feature segmentation optimization module is configured to perform forest feature segmentation through spectral difference analysis, reassign forest feature segmentation boundary points based on spectral dissimilarity, calculate core spectral signatures of each initial cluster ID of the three-dimensional forest point cloud after reassignment, and output a single tree point cloud entity set; the analysis application module is configured to output vector results according to the single tree point cloud entity set and store the vector results back to the central database; the image acquisition module comprises: converting digital quantization values of the LiDAR point cloud and multispectral optical image into radiance values, and converting the radiance values into surface reflectance through QUAC; calculating three-dimensional coordinates of each laser pulse return point through POS data, and generating original point cloud data by assigning the three-dimensional coordinates to geographic coordinates; separating the original point cloud data into ground points and non-ground points through a progressive TIN encryption filtering algorithm, generating a digital terrain model based on the separated ground points, subtracting the digital terrain model elevation value of the corresponding position of all non-ground points from the elevation value of the non-ground points, and obtaining a three-dimensional forest point cloud, wherein the Z value of the three-dimensional forest point cloud represents the true height of the forest from the ground.

Images