Forestry remote sensing image recognition method and system based on image recognition

Abstract

The invention discloses a forestry remote sensing image recognition method and system based on image recognition, and relates to the technical field of image processing, and the method comprises the steps: collecting and preprocessing a LiDAR point cloud and a multispectral optical image, and obtaining a three-dimensional forest point cloud; forest feature segmentation is carried out through spectral difference analysis, forest feature segmentation boundary points are redistributed based on spectral dissimilarity, a core spectral signature is calculated, and a single-tree point cloud entity set is output; and carrying out topological analysis on the single tree point cloud entity set, and outputting a vector result. According to the method, the three-dimensional forest point cloud is directly operated, each tree is identified as a single-tree point cloud entity containing complete space structure and spectrum information, the terrain influence is fundamentally eliminated, the three-dimensional structure parameters of the single tree can be accurately obtained, a two-stage strategy of geometric rough division and spectrum refinement is creatively provided, and the method is suitable for the three-dimensional tree point cloud entity of the single tree. The deep fusion perfectly combines the macroscopic accuracy of geometric segmentation and the microscopic identification of spectrum segmentation, and greatly improves the segmentation precision of the adhered tree crowns.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a forestry remote sensing image recognition method and system based on image recognition. Background Technology

[0002] As a core component of the ecosystem, forestry plays a crucial role in carbon sequestration, biodiversity conservation, soil and water conservation, and timber resource supply. Its dynamic changes directly impact global ecological security and sustainable development. With the intensification of global climate change and the increasing disturbance of forests by human activities, 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 on changes in forest cover, vegetation growth, and the distribution of endangered species habitats. From a resource management perspective, it is necessary to accurately count timber reserves, classify forest types (such as coniferous forests and broad-leaved forests), and monitor illegal logging. From a disaster prevention and control perspective, it is necessary to promptly identify forest fire hotspots and areas infected by pests and diseases to provide a basis for emergency response.

[0003] Currently, Chinese invention patent application number CN202411903952.4 discloses a forestry surveying method based on remote sensing big data, including the following steps: integrating remote sensing imagery, UAV images, temperature sensor data, and humidity sensor data for spatiotemporal alignment processing, filling in missing data through time interpolation, detecting outliers, calculating ecological indicators of the forest area based on the processed data, and outputting forest health ecological assessment results. This method improves the accuracy and completeness of data by integrating remote sensing imagery, UAV images, and temperature and humidity sensor data. Species identification and ecological function contribution analysis can accurately identify key species and their contributions to the ecosystem. By analyzing the dynamic changes in forest resources and climate impacts, it can effectively optimize resource allocation, improve management efficiency and risk control capabilities, enhance the accuracy and timeliness of forest resource monitoring, assessment, and management, and strengthen sustainable utilization and ecological protection capabilities. However, most existing technologies are based on two-dimensional images for canopy segmentation, which is easily affected by shadows, canopy overlap and terrain undulations, resulting in low segmentation accuracy and missing tree height information. They either rely solely on the geometric shape of point clouds or solely on the spectral information of images, with insufficient combination of the two. This leads to poor results in areas with complex forest stands and overlapping canopies, often stopping at extracting parameters of individual trees and lacking structural analysis of the forest as a whole ecosystem. Summary of the Invention

[0004] The technical problem solved by this invention is that most existing technologies are based on two-dimensional images for canopy segmentation, which is easily affected by shadows, canopy overlap and terrain undulations, resulting in low segmentation accuracy and missing tree height information. They either rely solely on the geometric shape of point clouds or solely on the spectral information of images, and the combination of the two is insufficient. This results in poor performance in areas with complex forest stands and overlapping canopies, and often stops at extracting parameters of individual trees, lacking structural analysis of the forest as a whole ecosystem.

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

[0006] Step S1: Collect LiDAR point cloud and multispectral optical images by drone within a preset time window, perform data preprocessing on the LiDAR point cloud and multispectral optical images to obtain three-dimensional forest point cloud;

[0007] Step S2: Perform forest feature segmentation through spectral difference analysis, redistribute the forest feature segmentation boundary points based on spectral dissimilarity, calculate the core spectral signature of each initial cluster ID of the three-dimensional forest point cloud after redistribution, and output the single tree point cloud entity set.

[0008] Step S3: Based on the single-tree point cloud entity set, output the vector results and back to the central database.

[0009] Preferably, step S1 includes:

[0010] The UAV completes the acquisition of LiDAR point cloud and multispectral optical images within a preset time window, converts the digital quantization values ​​of the LiDAR point cloud and multispectral optical images into radiance values, and converts the radiance values ​​into surface reflectance through QUAC.

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

[0012] The original point cloud data is separated into ground points and non-ground points by a progressive TIN encryption filtering algorithm. A digital terrain model is generated based on the separated ground points through interpolation. The elevation values ​​of the digital terrain model corresponding to the non-ground points are subtracted from the elevation values ​​of all non-ground points to obtain a three-dimensional forest point cloud. The Z value of the three-dimensional forest point cloud represents the actual height of the trees above the ground.

[0013] Preferably, step S2 includes:

[0014] After the data conversion is completed, the multispectral optical image is projected onto the three-dimensional forest point cloud, and the three-dimensional forest point cloud is given the characteristic attribute of spectral band values.

[0015] A horizontal resolution is set for the three-dimensional forest point cloud. The area involved in the three-dimensional forest point cloud is divided into a two-dimensional grid on the XY plane. The maximum Z coordinate value of all three-dimensional forest point cloud points in each cell of the two-dimensional grid is calculated and the maximum Z coordinate value is assigned to the cell.

[0016] For all empty cells without point cloud data, the inverse distance weighted interpolation method is used to fill the empty cells with the values ​​of the cells adjacent to them. After the two-dimensional grid data is filled, the canopy height raster data grid map is obtained.

[0017] The canopy height raster data grid map is inverted to generate a new raster data grid map. The inverted raster data grid map is used as input, and a standard watershed segmentation algorithm is used to generate preliminary segmentation area raster data. Each cell in the preliminary segmentation area raster data has a segmentation area ID with an integer value.

[0018] Traverse each point in the 3D forest point cloud, locate the corresponding cell in the raster data of the initially segmented region based on the two-dimensional plane coordinates of the point, and assign the segmentation region ID of the cell as the initial cluster ID to the point. After completing the initial cluster ID assignment operation, output the forest point cloud with the initial cluster ID.

[0019] Preferably, the process of redistributing forest feature segmentation boundary points based on spectral dissimilarity includes:

[0020] 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, and search for points that meet the conditions including having an initial cluster ID different from the current traversed point and a horizontal distance of less than the horizontal distance threshold. When a point that meets the conditions is found, add a boundary point type marker to the current traversed point.

[0021] When no point that meets the conditions is found, add a core point type marker to the current traversed point;

[0022] Output a forest point cloud with initial cluster ID and boundary point type markers or core point type markers.

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

[0024] All points in the 3D forest point cloud are grouped according to the initial cluster ID. All points in each group are marked as core points. The average values ​​of the spectral indices NDVI, PSRI, and MCARI of all extracted core point type marked points are calculated. The 3D vector [mean(NDVI), mean(PSRI), mean(MCARI)] is output as the core spectral signature vector, where mean() represents the average value. The core spectral signature vectors of all core point type marked 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.

[0025] Extract all points marked as boundary points in each group, and perform re-determination on the ownership of each point. The re-determination includes:

[0026] Step S21: Obtain the spectral vector of the point marked as a boundary point of the specified type. The mathematical expression of the spectral vector is:

[0027] ;

[0028] in, Let p be the spectral vector. Let p be the NDVI spectral index. For point p, The MCARI spectral index of point p is used to identify the initial cluster ID to which point p currently belongs and the IDs of all spatially adjacent clusters to point p;

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

[0030] Step S23: Calculate the multidimensional 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 multidimensional Euclidean distance, and output the updated three-dimensional forest point cloud.

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

[0032] Create a hash table where the key represents a unique initial cluster ID and the value represents a list of points. Iterate through each point in the updated 3D forest point cloud, read the initial cluster ID of the point, and add the read initial cluster ID to the point list of the corresponding cluster ID in the hash table. After the iteration is complete, each key-value pair in the hash table corresponds to an independent single tree point cloud entity.

[0033] Each individual tree point cloud entity is a point cloud set, representing an independent tree that is homogeneous in both space and spectrum;

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

[0035] ;

[0036] in, Let p be the multi-attribute feature vector. , and Let p be the three-dimensional coordinates. Let be spectral bands 1 to spectral band n, where n is a natural number greater than 1.

[0037] Preferably, the convex hull of each individual tree point cloud entity on the two-dimensional plane is calculated to generate a two-dimensional polygon corresponding to the canopy shape of the individual tree point cloud entity. The two-dimensional polygon includes a vector file describing an independent tree canopy. The vector file is subjected to contour tracking and hierarchical analysis to derive forest stand structure indicators, identify forest gaps, measure canopy closure, and analyze the spatial arrangement and adjacency relationship of trees. A forest gap is an area not covered by any canopy polygon.

[0038] Two-dimensional projection is performed on each single point cloud entity to obtain a two-dimensional point set. The convex hull is calculated on the two-dimensional point set. The calculated convex hull boundary vertices are connected in sequence to generate a closed two-dimensional polygon.

[0039] The two-dimensional polygons are associated with the corresponding initial cluster IDs and other feature attributes calculated from the corresponding points. These other feature attributes include NDVI spectral index, PSRI spectral index, MCARI spectral index, three-dimensional coordinate values, and spectral bands. A tree canopy vector file containing all independent tree canopy polygons is obtained. Each element in the file represents the canopy of a tree and includes the tree's attribute information.

[0040] Preferably, the single-tree point cloud entity set is analyzed topologically to output vector results, which include negative space regions, canopy closure, tree distribution patterns, and single-tree adjacency graphs.

[0041] The topology analysis process includes:

[0042] Perform a union operation on all two-dimensional polygons to obtain a large two-dimensional polygon that includes multiple unconnected parts. The large two-dimensional polygon represents all areas covered by tree canopies within the entire three-dimensional forest point cloud.

[0043] Subtract the two-dimensional large polygon from the boundary polygon of the entire three-dimensional forest point cloud to obtain the negative space region not covered by any tree canopy.

[0044] The total area of ​​the two-dimensional large polygon and the total area of ​​the entire three-dimensional forest point cloud are calculated. The canopy closure is obtained 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.

[0045] The distance between nearest-neighbor trees is obtained by calculating the distance between the centroid of the 2D polygon corresponding to a single tree and the centroid of its nearest neighbor tree. The expected average distance between nearest-neighbor trees under a random distribution is then calculated to determine the tree distribution pattern. Distribution patterns include clustered distribution, uniform distribution, and random distribution. The determination logic includes:

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

[0047] When the expected average value is greater than the uniform distribution threshold and less than the random distribution threshold, the tree distribution pattern is uniform.

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

[0049] Each tree's two-dimensional polygon is considered as a node in the tree's adjacency graph. The intersection test of GIS is used to determine whether the two-dimensional polygons of adjacent trees are in contact or overlap in space. If they are in contact or overlap, an edge is drawn between the corresponding adjacent tree nodes.

[0050] Preferably, attribute groups are calculated and linked for each vectorized canopy using the canopy vector file. The attribute groups include geometric attributes, spectral attributes, and derived attributes. The geometric attributes include canopy area, perimeter, tree height, and geographic location.

[0051] Spectral properties include spectral indices, PSRI spectral indices, MCARI spectral indices, and spectral bands;

[0052] Derived attributes include individual tree health status and tree species predicted by CNN based on the tree canopy vector file;

[0053] The tree canopy vector file, vector results, and the final output 3D forest point cloud and single tree point cloud entity set are stored back into the central database.

[0054] A forestry remote sensing image recognition system based on image recognition, comprising an image acquisition module, a feature segmentation and optimization module, and an analysis and application module, for implementing a forestry remote sensing image recognition method based on image recognition.

[0055] The image acquisition module is used to acquire LiDAR point cloud and multispectral optical image by UAV within a preset time window, and to perform data preprocessing on the LiDAR point cloud and multispectral optical image to obtain three-dimensional forest point cloud.

[0056] The feature segmentation optimization module is used to segment forest features through spectral difference analysis, redistribute the forest feature segmentation boundary points based on spectral dissimilarity, calculate the core spectral signature of each initial cluster ID of the three-dimensional forest point cloud after redistribution, and output the single tree point cloud entity set.

[0057] The analysis application module is used to output vector results based on the single-tree point cloud entity set and back to the central database.

[0058] The beneficial effects of this invention are as follows: This application directly manipulates three-dimensional forest point clouds, identifying each tree as a single-tree point cloud entity containing complete spatial structure and spectral information. This not only fundamentally eliminates the influence of terrain but also accurately obtains the three-dimensional structural parameters of individual trees, laying a solid foundation for more accurate biomass estimation and carbon storage calculation. It innovatively proposes a two-stage strategy of first geometric coarse segmentation and then spectral refinement. First, it utilizes the three-dimensional geometric advantages of point clouds for rapid and robust preliminary segmentation, defining the approximate range. Then, it uses multispectral information to re-determine the most difficult-to-handle boundary points point by point. This deep fusion perfectly combines the macroscopic accuracy of geometric segmentation with the microscopic identification power of spectral segmentation, greatly improving the segmentation accuracy of adherent canopies. Based on the accurate segmentation of individual trees, GIS topology analysis is introduced to systematically integrate discrete individual tree information, automatically interpreting key forest stand structure indicators such as forest gaps, canopy closure, spatial distribution patterns, and adjacency relationships. This elevates the analysis from tree identification to forest understanding, providing unprecedented quantitative and spatial decision-making basis for forest management, ecological health assessment, and biodiversity research. A fully automated, closed-loop system was built, encompassing data acquisition, feature segmentation, analysis and application, and final archiving to the central database. This design not only improves processing efficiency but, more importantly, ensures data standardization, process traceability, and result reusability, providing a solid technical platform for the time-series dynamic monitoring and big data analysis of forestry resources. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the basic process of a forestry remote sensing image recognition method provided in one embodiment of the present invention. Detailed Implementation

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0061] Reference Figure 1 As an embodiment of the present invention, a forestry remote sensing image recognition method based on image recognition is provided, comprising the following steps:

[0062] Step S1: Collect LiDAR point cloud and multispectral optical images using a drone within a preset time window, perform data preprocessing on the LiDAR point cloud and multispectral optical images, and obtain a three-dimensional forest point cloud.

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

[0064] Step S3: Based on the single-tree point cloud entity set, output the vector results and back to the central database.

[0065] This application identifies each tree as a single-tree point cloud entity containing complete spatial structure and spectral information through direct 3D individual tree entity analysis. This not only fundamentally eliminates the influence of terrain but also accurately obtains the 3D structural parameters of individual trees. It innovatively proposes a two-stage strategy of first geometric coarse segmentation and then spectral refinement. Deep fusion perfectly combines the macroscopic accuracy of geometric segmentation with the microscopic discernibility of spectral segmentation, greatly improving the segmentation accuracy of adhered tree canopies. Furthermore, it generates vector results that are easily applied by users through topological analysis, and finally archives them back to the central database in a fully automated, closed-loop process.

[0066] Step S1 includes:

[0067] The drone completes the acquisition of LiDAR point cloud and multispectral optical images within a preset time window, converts the digital quantization values ​​of LiDAR point cloud and multispectral optical images into radiance values, and converts the radiance values ​​into surface reflectance through QUAC.

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

[0069] The original point cloud data is separated into ground points and non-ground points by a progressive TIN encryption filtering algorithm. A digital terrain model is generated based on the separated ground points through interpolation. The elevation values ​​of the digital terrain model corresponding to the non-ground points are subtracted from the elevation values ​​of all non-ground points to obtain a three-dimensional forest point cloud. The Z-value of the three-dimensional forest point cloud represents the actual height of the trees above the ground.

[0070] In this embodiment, the time window is set to be no more than one hour apart within the same day to minimize intraday fluctuations in light and shadow differences and tree physiological states caused by changes in solar altitude angle. Optical images need to ensure a high overlap rate, with the overlap rate set to 80% to support the generation of high-quality orthophotos. LiDAR needs to ensure sufficient lateral overlap, with the lateral overlap rate set to 50% to ensure the integrity of data coverage on the sides and bottom of the canopy and avoid data gaps.

[0071] The matching and adjustment between UAV flight strips eliminates systematic elevation or plane errors between flight strips through the ICP algorithm, ensuring the internal consistency of the point cloud in the entire study area;

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

[0073] Step S2 includes:

[0074] After the data conversion is completed, the multispectral optical image is projected onto the 3D forest point cloud, and the 3D forest point cloud is given the characteristic attribute of spectral band values.

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

[0076] For all empty cells without point cloud data, the inverse distance weighted interpolation method is used to fill the empty cells with the values ​​of the cells adjacent to them. After the two-dimensional grid data is filled, the canopy height raster data grid map is obtained.

[0077] The canopy height raster data grid map is inverted to generate a new raster data grid map. The inverted raster data grid map is used as input to generate preliminary segmented area raster data through a standard watershed segmentation algorithm. Each cell in the preliminary segmented area raster data has a segmentation area ID with an integer value.

[0078] Traverse each point in the 3D forest point cloud, locate the corresponding cell in the raster data of the initially segmented region based on the point's 2D planar coordinates, and assign the segmentation region ID of the cell as the initial cluster ID to the point. After completing the initial cluster ID assignment operation, output the forest point cloud with the initial cluster ID.

[0079] Canopy height raster data grid map is used to represent the height of tree canopies;

[0080] This stage is the geometric coarse segmentation stage, which can intuitively reflect the advantages of the tree canopy top morphology. The watershed algorithm quickly finds the vertex of each tree canopy and uses it as the center for initial segmentation. It is computationally fast and robust, quickly assigning initial tree-specific IDs to massive point cloud data, providing an efficient starting point for subsequent fine-grained processing and greatly reducing computational complexity.

[0081] The process of redistributing forest feature segmentation boundary points based on spectral dissimilarity includes:

[0082] Assign a preset horizontal distance threshold to the forest point cloud with initial cluster ID. In this embodiment, the horizontal distance threshold is set to 1 meter. Traverse each point in the forest point cloud with initial cluster ID and search for points that meet the conditions including having an initial cluster ID different from the current traversed point and a horizontal distance of less than the horizontal distance threshold. When a point that meets the conditions is found, add a boundary point type marker to the current traversed point.

[0083] When no point that meets the conditions is found, add a core point type marker to the current traversed point;

[0084] Output a forest point cloud with initial cluster ID and boundary point type markers or core point type markers.

[0085] This is a key prerequisite for refined processing. It cleverly simplifies the problem, transforming the processing of all points into processing only the most ambiguous and error-prone boundary points, which greatly improves computational efficiency and targeting. The core points are stable and reliable and do not require processing; only a few boundary points are re-determined, avoiding unnecessary computational overhead.

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

[0087] Group all points in the 3D forest point cloud according to the initial cluster ID, extract all points in each group that are labeled as core points, calculate the average values ​​of the NDVI, PSRI, and MCARI spectral indices of all extracted core point type labeled points, and output a 3D vector [mean(NDVI), mean(PSRI), mean(MCARI)] as the core spectral signature vector, where mean() represents the average value. The core spectral signature vectors of all core point type labeled points are collected into a core spectral signature vector table, and each initial cluster ID in the core spectral signature vector table is mapped to a corresponding core spectral signature vector.

[0088] By calculating the spectral average of each cluster using only the core point, a pure and stable spectral representation of each potential canopy is obtained. The core point is far from the mixing region, and its spectral characteristics best represent the true situation of the tree, avoiding the contamination of the spectrum of neighboring trees, and providing an extremely reliable reference benchmark for subsequent boundary point decisions.

[0089] Extract all points marked as boundary points in each group, and perform re-determination on the ownership of each point. The re-determination includes:

[0090] Step S21: Obtain the spectral vector of the point marked as a boundary point. The mathematical expression for the spectral vector is:

[0091] ;

[0092] in, Let p be the spectral vector. Let p be the NDVI spectral index. The PSRI spectral index at point p. For the MCARI spectral index of point p, identify the initial cluster ID to which point p currently belongs and the IDs of all spatially adjacent clusters of point p;

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

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

[0095] Step S23: Calculate the multidimensional 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 multidimensional Euclidean distance, and output the updated three-dimensional forest point cloud.

[0096] By comparing the Euclidean distance between the self-spectrum of each boundary point and the core spectral signature of the surrounding clusters, pixel-level accurate correction of the boundaries of the adhering tree canopy is achieved. The spectral characteristics of a boundary point located between two trees are necessarily closer to the core spectrum of the tree to which it truly belongs. By voting based on spectral similarity, boundary points that have been incorrectly classified by the geometric algorithm can be pulled back into the correct tree canopy entity, greatly improving the segmentation accuracy.

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

[0098] Create a hash table where the key represents a unique initial cluster ID and the value represents a list of points. Iterate through each point in the updated 3D forest point cloud, read the initial cluster ID of the point, and add the read initial cluster ID to the point list of the corresponding cluster ID in the hash table. After the iteration is complete, each key-value pair in the hash table corresponds to an independent single tree point cloud entity.

[0099] Each individual tree point cloud entity is a point cloud set, representing an independent tree that is homogeneous in both space and spectrum;

[0100] Each point in the point cloud set included in a single tree point cloud entity contains a multi-attribute feature vector. The mathematical expression for the multi-attribute feature vector is:

[0101] ;

[0102] in, Let p be the multi-attribute feature vector. , and Let p be the three-dimensional coordinates. Let be spectral bands 1 to spectral band n, where n is a natural number greater than 1.

[0103] By aggregating point clouds into entities with multi-attribute feature vectors, the transformation from discrete points to structured objects is completed. The output is no longer scattered points, but an organized, attribute-rich digital twin tree, with each point carrying rich information, providing complete data preparation for subsequent vectorization and analysis.

[0104] The convex hull of each individual tree point cloud entity on the two-dimensional plane is calculated to generate a two-dimensional polygon corresponding to the canopy shape of the individual tree point cloud entity. The two-dimensional polygon includes a vector file describing the independent canopy. The vector file is subjected to contour tracking and hierarchical analysis to derive forest stand structure indicators, identify forest gaps, measure canopy closure, and analyze the spatial arrangement and adjacency relationship of trees. Forest gaps are areas not covered by any canopy polygons.

[0105] Perform a two-dimensional projection on each individual point cloud entity to obtain a two-dimensional point set. Calculate the convex hull on the two-dimensional point set and connect the calculated convex hull boundary vertices in sequence to generate a closed two-dimensional polygon.

[0106] The two-dimensional polygons are associated with the corresponding initial cluster IDs and other feature attributes calculated from the corresponding points. These other feature attributes include NDVI spectral index, PSRI spectral index, MCARI spectral index, three-dimensional coordinate values, and spectral bands. This yields a canopy vector file containing all independent canopy polygons. Each element in the file represents the canopy of a tree and includes the tree's attribute information.

[0107] This step involves projecting the 3D point cloud entities and abstracting them into 2D tree canopy polygons, which are considered as a feature in GIS.

[0108] A convex hull is a minimal convex polygon that contains all points. The convex hull algorithm is fast, provides stable results, and can accurately summarize the general outline of a tree canopy.

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

[0110] Convex hull vectorization projects each individual entity into a two-dimensional polygon, achieving seamless integration with standard GIS platforms. Vector polygons are the basic unit of GIS analysis, facilitating area and perimeter calculations and providing an object basis for subsequent topology analysis. Using the convex hull algorithm balances computational efficiency and the accuracy of contour generalization.

[0111] Step S3 includes:

[0112] The topological analysis of the single-tree point cloud entity set outputs vector results, which include negative space region, canopy closure, tree distribution pattern and single-tree adjacency graph;

[0113] The topology analysis process includes:

[0114] Perform a union operation on all two-dimensional polygons to obtain a large two-dimensional polygon that includes multiple unconnected parts. This large two-dimensional polygon represents all areas covered by tree canopies within the entire three-dimensional forest point cloud.

[0115] Subtract the two-dimensional large polygon from the boundary polygon of the entire 3D forest point cloud to obtain the negative space region not covered by any tree canopy.

[0116] The total area of ​​the two-dimensional large polygon and the total area of ​​the entire three-dimensional forest point cloud are calculated. The canopy closure is obtained 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. The canopy closure index macroscopically reflects the density of the forest stand.

[0117] The distance between nearest-neighbor trees is obtained by calculating the distance between the centroid of the 2D polygon corresponding to a single tree and the centroid of its nearest neighbor tree. The expected average distance between nearest-neighbor trees under a random distribution is then calculated to determine the tree distribution pattern. Distribution patterns include clustered distribution, uniform distribution, and random distribution. The determination logic includes:

[0118] When the expected average value is greater than the random distribution threshold, the distribution pattern of trees is random.

[0119] When the expected average value is greater than the uniform distribution threshold and less than the random distribution threshold, the tree distribution pattern is uniform.

[0120] When the expected average value is less than the uniform distribution threshold, the distribution pattern of trees is clustered.

[0121] Each tree's two-dimensional polygon is considered as a node in the tree's adjacency graph. The intersection test of GIS is used to determine whether the two-dimensional polygons of adjacent trees are in contact or overlap in space. If they are in contact or overlap, an edge is drawn between the corresponding adjacent tree nodes.

[0122] This adjacency graph visually shows which trees are adjacent to each other. Based on this graph, the degree of each node can be calculated, that is, how many neighbors each tree has, so as to analyze the competitive relationship between trees.

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

[0124] Forest gap, canopy closure, spatial pattern, and adjacency analysis provide systematic insights into forest stands beyond the individual tree scale. These indicators are core parameters in forest ecology and management. Forest gap identification helps assess forest land use and habitat fragmentation, canopy closure is a key indicator for evaluating stand density and light competition, spatial distribution patterns reveal intraspecific and interspecific relationships and environmental adaptation strategies of trees, and individual tree adjacency maps quantify direct competition between trees and can be used to guide management activities such as tending and thinning.

[0125] This step elevates the entire process, shifting from analyzing individual trees to analyzing the structure of a forest composed of these trees. Here, we will apply the concepts of contour tracing and hierarchical analysis, but the object of our work will no longer be pixels, but rather a collection of tree canopy polygons.

[0126] For each vectorized canopy, attribute groups are calculated and linked using the canopy vector file. These attribute groups include geometric attributes, spectral attributes, and derived attributes. Geometric attributes include canopy area, perimeter, tree height, and geographic location.

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

[0128] Derived attributes include individual tree health status and tree species predicted by CNN based on canopy vector files;

[0129] Store the tree canopy vector files, vector results, and the final output 3D forest point cloud and single tree point cloud entity set back to the central database.

[0130] This final stage compiles all derived information and archives it back into the central database. This workflow provides a robust, end-to-end solution that is both fast, thanks to efficient watershed segmentation and refinement algorithms, and highly accurate. By combining 3D geometry with adaptive spectral refinement and advanced feature prediction, it constructs a cyclical, iterative, and traceable intelligent forestry information system. All data and results are traceable, facilitating long-term monitoring of dynamic changes (such as analyzing changes in forest gaps and tree growth over several years) and providing high-quality, structured datasets for training more powerful AI models (such as combining CNNs to predict tree species and health status).

[0131] A forestry remote sensing image recognition system based on image recognition, comprising an image acquisition module, a feature segmentation and optimization module, and an analysis and application module, for implementing a forestry remote sensing image recognition method based on image recognition.

[0132] The image acquisition module is used to acquire LiDAR point clouds and multispectral optical images by drone within a preset time window, perform data preprocessing on the LiDAR point clouds and multispectral optical images, and obtain three-dimensional forest point clouds.

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

[0134] The analysis application module is used to output vector results based on the single-tree point cloud entity set and back to the central database.

[0135] This application directly manipulates 3D forest point clouds, identifying each tree as a single-tree point cloud entity containing complete spatial structure and spectral information. This not only fundamentally eliminates the influence of topography but also accurately obtains the 3D structural parameters of individual trees, laying a solid foundation for more accurate biomass estimation and carbon storage calculation. It innovatively proposes a two-stage strategy: first, geometric coarse segmentation, then spectral refinement. First, the 3D geometric advantages of the point cloud are utilized for rapid and robust preliminary segmentation, defining the approximate range. Then, multispectral information is used to re-determine the most difficult-to-handle boundary points point by point. This deep fusion perfectly combines the macroscopic accuracy of geometric segmentation with the microscopic identification power of spectral segmentation, greatly improving the segmentation accuracy of adherent canopies. Based on the accurate segmentation of individual trees, GIS topology analysis is introduced to systematically integrate discrete individual tree information, automatically interpreting key forest stand structure indicators such as forest gaps, canopy closure, spatial distribution patterns, and adjacency relationships. This elevates the analysis from tree identification to forest understanding, providing unprecedented quantitative and spatial decision-making basis for forest management, ecological health assessment, and biodiversity research. A fully automated, closed-loop system was built, encompassing data acquisition, feature segmentation, analysis and application, and final archiving to the central database. This design not only improves processing efficiency but, more importantly, ensures data standardization, process traceability, and result reusability, providing a solid technical platform for the time-series dynamic monitoring and big data analysis of forestry resources.

[0136] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEpROM), Erasable Programmable Read Only Memory (EpROM), Programmable Red-Only Memory (pROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A forestry remote sensing image recognition method based on image recognition, characterized in that, Includes the following steps: Step S1: Collect LiDAR point cloud and multispectral optical images by drone within a preset time window, perform data preprocessing on the LiDAR point cloud and multispectral optical images to obtain three-dimensional forest point cloud; Step S2: Perform forest feature segmentation through spectral difference analysis, redistribute the forest feature segmentation boundary points based on spectral dissimilarity, calculate the core spectral signature of each initial cluster ID of the three-dimensional forest point cloud after redistribution, and output the single tree point cloud entity set. Step S3: Based on the single-tree point cloud entity set, output the vector results and back to the central database.

2. The forestry remote sensing image recognition method based on image recognition as described in claim 1, characterized in that, Step S1 includes: The digital quantization values ​​of the LiDAR point cloud and multispectral optical images are converted into radiance values, and the radiance values ​​are converted into surface reflectance using QUAC. The three-dimensional coordinates of the return point of each laser pulse are calculated using POS data, and the three-dimensional coordinates are assigned to geographic coordinates to generate raw point cloud data. The original point cloud data is separated into ground points and non-ground points by a progressive TIN encryption filtering algorithm. A digital terrain model is generated based on the separated ground points through interpolation. The elevation values ​​of the digital terrain model corresponding to the non-ground points are subtracted from the elevation values ​​of all non-ground points to obtain a three-dimensional forest point cloud. The Z value of the three-dimensional forest point cloud represents the actual height of the trees above the ground.

3. The forestry remote sensing image recognition method based on image recognition as described in claim 2, characterized in that, Step S2 includes: After the data conversion is completed, the multispectral optical image is projected onto the three-dimensional forest point cloud, and the three-dimensional forest point cloud is given the characteristic attribute of spectral band values. A horizontal resolution is set for the three-dimensional forest point cloud. The area involved in the three-dimensional forest point cloud is divided into a two-dimensional grid on the XY plane. The maximum Z coordinate value of all three-dimensional forest point cloud points in each cell of the two-dimensional grid is calculated and the maximum Z coordinate value is assigned to the cell. For all empty cells without point cloud data, the inverse distance weighted interpolation method is used to fill the empty cells with the values ​​of the cells adjacent to them. After the two-dimensional grid data is filled, the canopy height raster data grid map is obtained. The canopy height raster data grid map is inverted to generate a new raster data grid map. The inverted raster data grid map is used as input, and a standard watershed segmentation algorithm is used to generate preliminary segmentation area raster data. Each cell in the preliminary segmentation area raster data has a segmentation area ID with an integer value. Traverse each point in the 3D forest point cloud, locate the corresponding cell in the raster data of the initially segmented region based on the two-dimensional plane coordinates of the point, and assign the segmentation region ID of the cell as the initial cluster ID to the point. After completing the initial cluster ID assignment operation, output the forest point cloud with the initial cluster ID.

4. The forestry remote sensing image recognition method based on image recognition as described in claim 3, characterized in that, The process of redistributing forest feature segmentation boundary points based on spectral dissimilarity includes: 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, and search for points that meet the conditions including having an initial cluster ID different from the current traversed point and a horizontal distance of less than the horizontal distance threshold. When a point that meets the conditions is found, add a boundary point type marker to the current traversed point. When no point that meets the conditions is found, add a core point type marker to the current traversed point; Output a forest point cloud with initial cluster ID and boundary point type markers or core point type markers.

5. The forestry remote sensing image recognition method based on image recognition as described in claim 4, characterized in that, The calculation process for the core spectral signature of each initial cluster ID includes: All points in the 3D forest point cloud are grouped according to the initial cluster ID. All points in each group are marked as core points. The average values ​​of the spectral indices NDVI, PSRI, and MCARI of all extracted core point type marked points are calculated. The 3D vector [mean(NDVI), mean(PSRI), mean(MCARI)] is output as the core spectral signature vector, where mean() represents the average value. The core spectral signature vectors of all core point type marked 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 all points marked as boundary points in each group, and perform re-determination on the ownership of each point. The re-determination includes: Step S21: Obtain the spectral vector of the point marked as a boundary point of the specified type. The mathematical expression of the spectral vector is: ； in, Let p be the spectral vector. Let p be the NDVI spectral index. For point p, The MCARI spectral index of point p is used to identify the initial cluster ID to which point p currently belongs and the IDs of all spatially adjacent clusters to point p; Step S22: Query the core spectral signature vectors corresponding to the initial cluster ID to which point p currently belongs and the IDs of all spatially adjacent clusters of point p from the core spectral signature vector table; Step S23: Calculate the multidimensional 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 multidimensional Euclidean distance, and output the updated three-dimensional forest point cloud.

6. The forestry remote sensing image recognition method based on image recognition as described in claim 5, characterized in that, The output single-tree point cloud entity set includes: Create a hash table where the key represents a unique initial cluster ID and the value represents a list of points. Iterate through each point in the updated 3D forest point cloud, read the initial cluster ID of the point, and add the read initial cluster ID to the point list of the corresponding cluster ID in the hash table. After the iteration is complete, each key-value pair in the hash table corresponds to an independent single tree point cloud entity. Each individual tree point cloud entity is a point cloud set, representing an independent tree that is homogeneous in both space and spectrum; Each point in the point cloud set included in the single-tree point cloud entity comprises a multi-attribute feature vector, and the mathematical expression of the multi-attribute feature vector is: ； in, Let p be the multi-attribute feature vector. , and Let p be the three-dimensional coordinates. Let be spectral bands 1 to spectral band n, where n is a natural number greater than 1.

7. The forestry remote sensing image recognition method based on image recognition as described in claim 6, characterized in that: The convex hull of each individual tree point cloud entity on the two-dimensional plane is calculated to generate a two-dimensional polygon corresponding to the canopy shape of the individual tree point cloud entity. The two-dimensional polygon includes a vector file describing an independent tree canopy. The vector file is subjected to contour tracking and hierarchical analysis to derive forest stand structure indicators, identify forest gaps, measure canopy closure, and analyze the spatial arrangement and adjacency relationship of trees. A forest gap is an area not covered by any canopy polygon. Two-dimensional projection is performed on each single point cloud entity to obtain a two-dimensional point set. The convex hull is calculated on the two-dimensional point set. The calculated convex hull boundary vertices are connected in sequence to generate a closed two-dimensional polygon. The two-dimensional polygons are associated with the corresponding initial cluster IDs and other feature attributes calculated from the corresponding points. These other feature attributes include NDVI spectral index, PSRI spectral index, MCARI spectral index, three-dimensional coordinate values, and spectral bands. A tree canopy vector file containing all independent tree canopy polygons is obtained. Each element in the file represents the canopy of a tree and includes the tree's attribute information.

8. The forestry remote sensing image recognition method based on image recognition as described in claim 7, characterized in that: The topological analysis of the single-tree point cloud entity set outputs vector results, which include negative space region, canopy closure, tree distribution pattern and single-tree adjacency graph; The topology analysis process includes: Perform a union operation on all two-dimensional polygons to obtain a large two-dimensional polygon that includes multiple unconnected parts. The large two-dimensional polygon represents all areas covered by tree canopies within the entire three-dimensional forest point cloud. Subtract the two-dimensional large polygon from the boundary polygon of the entire three-dimensional forest point cloud to obtain the negative space region not covered by any tree canopy. The total area of ​​the two-dimensional large polygon and the total area of ​​the entire three-dimensional forest point cloud are calculated. The canopy closure is obtained 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. The distance between nearest-neighbor trees is obtained by calculating the distance between the centroid of the 2D polygon corresponding to a single tree and the centroid of its nearest neighbor tree. The expected average distance between nearest-neighbor trees under a random distribution is then calculated to determine the tree distribution pattern. Distribution patterns include clustered distribution, uniform distribution, and random distribution. The determination logic includes: When the expected average value is greater than the random distribution threshold, the distribution pattern of trees is random. When the expected average value is greater than the uniform distribution threshold and less than the random distribution threshold, the tree distribution pattern is uniform. When the expected average value is less than the uniform distribution threshold, the distribution pattern of trees is clustered. Each tree's two-dimensional polygon is considered as a node in the tree's adjacency graph. The intersection test of GIS is used to determine whether the two-dimensional polygons of adjacent trees are in contact or overlap in space. If they are in contact or overlap, an edge is drawn between the corresponding adjacent tree nodes.

9. The forestry remote sensing image recognition method based on image recognition as described in claim 8, characterized in that: For each vectorized canopy, attribute groups are calculated and linked using the canopy vector file. These attribute groups include geometric attributes, spectral attributes, and derived attributes. The geometric attributes include canopy area, perimeter, tree height, and geographic location. Spectral properties include spectral indices, PSRI spectral indices, MCARI spectral indices, and spectral bands; Derived attributes include individual tree health status and tree species predicted by CNN based on the tree canopy vector file; The tree canopy vector file, vector results, and the final output 3D forest point cloud and single tree point cloud entity set are stored back into the central database.

10. A forestry remote sensing image recognition system based on image recognition, the system being used to execute a forestry remote sensing image recognition method based on image recognition, characterized in that, It includes an image acquisition module, a feature segmentation optimization module, and an analysis and application module: The image acquisition module is used to acquire LiDAR point cloud and multispectral optical image by UAV within a preset time window, and to perform data preprocessing on the LiDAR point cloud and multispectral optical image to obtain three-dimensional forest point cloud. The feature segmentation optimization module is used to segment forest features through spectral difference analysis, redistribute the forest feature segmentation boundary points based on spectral dissimilarity, calculate the core spectral signature of each initial cluster ID of the three-dimensional forest point cloud after redistribution, and output the single tree point cloud entity set. The analysis application module is used to output vector results based on the single-tree point cloud entity set and back to the central database.

Images