Method for inventorying forest stands

Abstract

The invention relates to the forestry industry. The present inventory method comprises obtaining laser scanning data, stitching the data into a point cloud and dividing the cloud into regions, searching for tree tops and detecting the tops of main-storey trees using a local maxima algorithm on a Z-projection of a region while applying median filter filtering, and segmenting the trees on the basis of the detected tops. Segmentation includes detecting a second storey by monitoring reflection density waves, determining the cutoff of the second storey, checking the number of trees, calculating a Voronoi diagram, performing watershed segmentation, obtaining the result as an intersection of pixels, re-dividing the segmented trees, generating data for classifying the trees by species, performing classification by a support vector method using a trained machine model, generating, on the basis of the results, a list indicating height and species name, selecting the diameter of a tree on the basis of a height class table, the stand height class for each species and the forest inventory region, calculating the volume of the tree trunk, and establishing a forest inventory. The invention makes it possible to create a computer-implemented method for inventorying forest stands using airborne laser scanning data, which is carried out automatically and meets established accuracy criteria.

Description

[0001] METHOD OF FOREST STAND ASSESSMENT BASED ON AIRBORNE LASER SCANNING DATA FIELD OF TECHNOLOGY

[0002] The claimed technical solution generally relates to the field of computer technology, and in particular to a method for forest plantation taxation based on airborne laser scanning data.

[0003] LEVEL OF TECHNOLOGY

[0004] The prior art includes patent RU2828596C1 "METHOD FOR DETERMINING FOREST SCALE PARAMETERS", issued by DIGITAL RESEARCH Limited Liability Company, published on 14.10.2024.

[0005] This patent describes a solution related to a technology for determining forest parameters by conducting preliminary surveys using unmanned aerial vehicles. The method for determining forest parameters involves surveying a forest area from an unmanned aerial vehicle using a multispectral camera with at least four channels and a digital RGB camera. The method then captures images of the forest area from the multispectral camera and images of the forest area from the digital RGB camera, along with information about the forest area images. The images are then photogrammetrically processed, followed by the construction of an orthomosaic and a point cloud based on the images from the multispectral camera for each of their channels, and the production of the orthomosaic and point cloud based on the images from the digital RGB camera.Using a point cloud constructed from RGB images, a digital terrain model and a digital elevation model are constructed, the heights of the tree stand are adjusted relative to the terrain features, individual trees are searched for, and a set of features of a specific tree is fed to the input of a pre-trained machine learning model to predict the tree species, followed by obtaining a set of tree vertices and crowns with assigned class labels indicating the tree species.

[0006] A disadvantage of the known solution in this area of ​​technology is the strong dependence of the quality of aerial photography on different periods of the year and lighting conditions.

[0007] The prior art includes patent RU2773144C1, "Method for Determining Stemwood Stock Using Unmanned Aerial Survey Data," issued by Lomonosov Moscow State University (MSU), published May 31, 2022. This patent describes a solution related to forest inventory techniques, which is carried out using data obtained from unmanned aerial vehicles (UAVs). The essence of the claimed solution lies in UAV photography to create a database of images and their subsequent photogrammetric processing as part of the creation of a digital tree canopy model (DTCM). The DTCM is used to subsequently determine the stemwood stock separately for each tree species.

[0008] A disadvantage of the known solution in this area of ​​technology is the strong dependence of the quality of aerial photography on different periods of the year and lighting conditions, which significantly affect the results of interpretation.

[0009] In addition, the prior art includes patent RU2739436C1 "METHOD FOR UPDATING FOREST INVENTORY DATA BASED ON MACHINE LEARNING", issued by Ecomonitoring Limited Liability Company, published on 24.12.2020.

[0010] This patent describes a method for updating forest inventory data using machine learning. This known solution relates to forestry, specifically to inventorying forest cover and forest growth conditions. Remote sensing data and existing forest inventory data are prepared. The remote sensing data consists of multispectral satellite imagery and its derivatives, as well as a digital elevation model and its derivatives. The prepared data is processed using machine learning, implemented on the original forest inventory data. Remote sensing is conducted repeatedly, with observation periods covering various canopy conditions, and also takes into account spatial data contained in historical forest identification data and thematic maps. This improves the accuracy of forest inventory.

[0011] The disadvantage of this solution is the impossibility of using the technology for surveying small areas for timber harvesting, as well as the strong dependence of the quality of the obtained materials on weather conditions during surveying.

[0012] The prior art includes solution RU2728159C1, "METHOD FOR TAXING PLANTS," published by the Center for Space Technologies and Services, Limited Liability Company, on July 28, 2020. The method involves remote sensing of the area being taxed using satellite imagery or aerial photography data and conducting ground measurements at control points. Airborne laser scanning data is also used to construct a digital terrain model and a digital canopy model of the stands. Ultra-high-resolution images of at least 0.1 m / pixel are used as satellite imagery or aerial photography data. Networks of calibration test plots are constructed based on the classification of synthesized aerial photography, satellite imagery, and airborne laser scanning images by age groups, forest growth condition types, and dominant species.Based on ground measurements on calibration sample plots, statistical models are constructed to correlate forest inventory parameters with variables from aerial photography, satellite imagery, and airborne laser scanning. Based on the identified dependencies, forest inventory parameters are calculated for cells of a 16x16-meter calculation matrix covering the entire forest inventory area. Parameters are determined for forest inventory units, creating a forest inventory database, which is used to compile a forest inventory description.

[0013] The disadvantage of this known solution is the use of ground-based measuring instruments, which significantly reduces the productivity of the work and increases labor costs.

[0014] The traditional method for determining forest parameters is forest inventory, which can be accomplished by visual, visual and measurement, or interpretation. The main drawbacks of the above-mentioned approaches are the high labor intensity and subjectivity of parameter assessment. Furthermore, prior art methods for determining forest parameters use satellite imagery. However, the low spatial resolution of these images precludes tree-by-tree analysis and, as a result, does not provide comprehensive forest parameters.

[0015] The technical solution proposed for consideration is aimed at eliminating the shortcomings of the current level of technology.

[0016] ESSENCE OF THE INVENTION

[0017] The technical challenge addressed by the proposed technical solution is the creation of a computer-implemented method for forest stand taxation based on aerial laser scanning data, which is performed automatically and meets established accuracy criteria, in accordance with legislation and relevant regulatory legal acts of the federal executive bodies of the Russian Federation. APPROVED

[0018] by order of the Ministry of Natural Resources of Russia dated October 17, 2022 No. 688 Procedure for the allocation and taxation of logging areas (as amended on September 17, 2024):

[0019] 6. The error in the conducted allocation and inventory of logging areas for the total stock and stock of commercial timber in the logging area should not exceed 10% for the stock and species (species) composition.

[0020] The technical result of the solution is the ability to obtain taxation results using the continuous counting method in automatic mode, with an accuracy that meets the requirements of Russian forest legislation.

[0021] The claimed technical result is achieved through a method of forest stand taxation based on aerial laser scanning data, which comprises the following stages:

[0022] - receive data from ultra-dense aerial laser scanning of forest stands using a drone;

[0023] - stitch the obtained data into a point cloud and divide this cloud into sections of a certain size;

[0024] - search for tree tops in the obtained areas and detect the tops of the main tier trees using the local maximum algorithm on the Z-projection of the area with filtering by the median filter;

[0025] - after which they perform tree segmentation based on the detected vertices, during which:

[0026] • detection of the second tier is carried out by monitoring the reflection density waves for each segmented tree, which is a set of points, and the second tier is cut off;

[0027] • for each segmented tree, the number of trees belonging to it is checked, in part, based on the raster image of the heights, the Voronoi diagram is calculated and the segmentation is calculated using the watershed method, after which the final result is taken as the intersection of pixels belonging to the same class according to both methods, the re-segmented trees are separated;

[0028] • for the final segmented trees, data is generated for their classification by species and, based on the generated vector, classification is performed using the support vector method using a trained machine learning model;

[0029] • a list of trees is formed based on the results of segmentation and classification, indicating the height value and the name of the species, and based on tables of height categories, the height category of the stand for each species and forest inventory area, the diameter of the tree is selected and, using formulas for calculating the reserve, the volume of the tree trunk is calculated;

[0030] • based on the results obtained at the previous stage, the forest inventory is calculated for the area under study for the main and second tiers - total reserve, reserve by species, reserve by trunk thickness.

[0031] DESCRIPTION OF DRAWINGS

[0032] The invention will be further described in accordance with the accompanying drawing, which is provided to illustrate the essence of the invention and in no way limits its scope. The following drawing is attached to the application:

[0033] Fig. 1 illustrates a block diagram of the claimed method.

[0034] DETAILED DESCRIPTION OF THE INVENTION

[0035] The following detailed description of the invention includes numerous implementation details to provide a clear understanding of the present invention. However, it will be apparent to one skilled in the art how the present invention may be used with or without these implementation details. In other instances, well-known methods, procedures, and components have not been described in detail to avoid obscuring the features of the present invention.

[0036] Furthermore, it will be clear from the foregoing description that the invention is not limited to the embodiment described. Numerous possible modifications, changes, variations, and substitutions, while preserving the spirit and form of the present invention, will be apparent to those skilled in the art.

[0037] The claimed technical solution proposes a device for comprehensive monitoring of the technical condition of electromechanical equipment with a decision-making unit based on a neural network.

[0038] Fig. 1 shows a block diagram of the claimed method for forest stand taxation based on aerial laser scanning data.

[0039] To determine forest inventory, the first step involves acquiring ultra-high-density aerial laser scanning data from a drone. Inventory is conducted to assess forest condition and involves a series of measures to determine the average width, height, age, tree species, number of tree stands, and forest density. Before processing, the obtained data is then normalized using the knnidw algorithm.

[0040] The next step is to stitch the obtained data into a point cloud and divide this cloud into sections of a specified size. The resulting stands are then divided into 1x1 meter cells, with 1,000 points randomly selected from each cell. Tree tops are then searched for in these sections and mainstory tree tops are detected using a local maximum algorithm on the Z-projection of the section, filtered with a median filter.

[0041] Mainstory tree top detection is performed using a local maxima search algorithm on a raster representation of the plot's projection along the Z axis at a resolution of 10 cm per pixel. The pixel intensity value is taken to be equal to the height of the highest point corresponding to its cell. Gaps are filled using a median filter, shifting a 3x3 pixel window.

[0042] A local maximum is a pixel whose intensity is higher than all surrounding pixels (on eight sides). Local maxima are detected by dividing the raster into 3x3 cells, with the central pixel in each cell being considered the initial pixel. Next, an iterative algorithm is run, searching for the highest pixel in a 9x9 neighborhood around the initial pixel. This pixel becomes the initial pixel for the next iteration of the algorithm. This algorithm is run until the result of the nth run is consistent with the result of the nth run, after which it terminates, and the pixels found are considered local maxima.

[0043] Potential vertices are filtered. The Euclidean distance between pairs of adjacent vertices is controlled. For each vertex, its 16 nearest neighbors are selected and the number of such neighbors that are closer than 3 meters from this vertex is calculated. The algorithm then iteratively eliminates vertices in descending order of the number of nearby vertices and recalculates the number of nearby vertices for each vertex until all vertices have vertices closer than 3 meters.

[0044] Tree segmentation is then performed based on the detected vertices. Tree segmentation is performed using the algorithm described by Dalponte et al. (2016) based on the detected vertices.

[0045] Each local maximum is used as a starting point for tree crown segmentation. The algorithm expands the region around the local maximum, including neighboring pixels if their height values ​​meet certain uniformity criteria. This expansion process continues until the crown boundaries are reached, defined by a sharp drop in height or other thresholds set by the algorithm. Thus, each segment corresponds to a separate tree crown.

[0046] After segmentation of the tree crown, information about its boundaries is superimposed on the point cloud, assigning each point a label corresponding to a specific tree crown based on its horizontal position.

[0047] During segmentation, the second tier is detected using reflection density wave monitoring. For each segmented tree, which is represented by a set of points, the second tier is cut off. The tree is divided into segments by height at 1-meter intervals, the number of points in each segment is calculated, and a graph of the number of points versus height is plotted. The graph is smoothed with a Gaussian, sigma = 2, size = 6. The algorithm then analyzes the graph from right to left, searching for the second derivative to reach zero. Provided that such a value is reached no closer to the maximum height than 3 meters, this height is accepted as the second tier cutoff. Points below the cutoff are considered the second tier.

[0048] For each segmented tree, the number of trees belonging to it is checked. Local maxima are detected with a resolution of 0.07 m per pixel using a 3x3 moving window median filter. A Voronoi diagram is calculated based on the raster image of the heights, and segmentation is performed using the watershed method. The final result is then taken as the intersection of pixels belonging to the same class according to both methods. Repeatedly segmented trees are then separated.

[0049] For the final segmented trees, data is generated for their classification by species, and the generated vector is used for classification using the support vector machine learning model. The bucket height is calculated as the height of the highest point divided by 20. A bucket is defined as a group of trees with heights from X meters to X + Y meters, for example, from 20 to 20 + 2 = 22 meters. Thus, Y, or 2 in this example, is the bucket height. The division is performed relative to the height along the Z-axis (the axis perpendicular to the ground plane) of each point cloud describing the tree representation, regardless of the heights of all trees, in order to obtain a feature description from the tree's morphological features.

[0050] To classify tree species using the support vector machine, the center of the bounding PC17RU2025 / 000056 is first determined from the resulting segmented point cloud.

[0051] A frame along the X and Y axes (the ground plane) is created, through which a perpendicular axis of rotation is defined. Six planes are then constructed relative to this axis, in 30° increments, onto which all cloud points are projected. Next, X and Y coordinate axes are created relative to the lower left corner of each projection (with the Y axis corresponding to the vertical Z axis), and the Y-axis range from 0 to the maximum height is divided into 20 equal segments. In each segment, points are filtered and the minimum and maximum values ​​are calculated along the X and Y axes. The resulting boundary values ​​are connected as the height increases, forming polygons of the lateral projection of the crown, with the connections at the outer segments closing a polygon describing the shape of the crown.The rectangle that bounds this polygon can inscribe geometric figures such as a rectangle, an ellipse, isosceles triangles with bases on the lower and upper faces, and a semi-ellipse resting with its minor axis on the lower face, which allows for a detailed morphological description for further classification.

[0052] All points belonging to a segmented tree are distributed into buckets. For each bucket, points with the maximum X coordinate, minimum X coordinate, maximum Y coordinate, and minimum Y coordinate are selected. Then, projection polygons are formed separately for pairs of maximum and minimum X and Y coordinates. These projection polygons consist of successively connected minimum values ​​along one coordinate axis and maximum values ​​in ascending order of height. In the lowest and highest buckets, the maximum values ​​are connected with the minimum values, thus forming a polygon that describes the crown shape in side projection. The following shapes are inscribed within the rectangle circumscribed around the polygon: a rectangle, an ellipse, an isosceles triangle with its base on the lower face and on the upper face, and a semi-ellipse resting with its minor axis on the lower face.For each figure, a compression of 5% to 100% of the minor edge length of the circumscribed rectangle is applied, and the intersection-to-union ratio with the polygon is calculated. The maximum value of this ratio among all applied compressions is then calculated for each figure. The median value of these maxima is then calculated for all considered projections (along the X and Y axes). A vector of these values ​​is then formed, where each scalar represents the median value of the maximum intersection-to-union ratio for the figure type. In addition, the maximum, median, mean, 75th, and 95th percentiles of the red, green, and blue reflectance intensity attribute values, as well as the reflectance intensity, are added to the vector. The vector components are normalized.

[0053] Breed classification is performed using the support vector machine (SVM) based on a generated vector. The model is pretrained. The proposed solution uses a support vector machine (SVM) model. The SVM is described in Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20(3), 273–297. DOI: 10.1007 / BF00994018.

[0054] Tree species classification is performed using a support vector machine (SVM). This algorithm is trained to find a hyperplane separating data in a multidimensional feature space. The features fed to the model are already prepared through analysis of crown shape, point distribution, and reflectivity.

[0055] Tree species classification uses a variety of features, including both geometric and color characteristics. The analysis is based on figures describing the shape of a tree's profile in lateral projection. These include triangles, inverted triangles, rectangles, ellipses, trapezoids, extended semicircles, semicircles, and diamonds. These figures are generated for both the entire tree and its upper portion, allowing for consideration of tree structure at different elevations.

[0056] Additionally, statistical indicators of reflectance intensity and color channels are analyzed. For each channel—red, green, and blue—median values, as well as the 75th and 95th percentiles, are calculated. Similar indicators are determined for overall reflectance intensity.

[0057] To more accurately reflect color characteristics, the ratios between channels are calculated. For example, the ratios of the median values ​​of the red channel to the green and blue channels, as well as the green to blue channel, are analyzed. Furthermore, the ratios of the 75th and 95th percentiles of these same channels are studied, which helps identify patterns in the distribution of color characteristics.

[0058] All these features are combined into a multidimensional vector, which allows the support vector machine (SVM) algorithm to learn to recognize tree species. This is achieved by taking into account both the tree's profile shape and its color patterns at different heights.

[0059] Based on the segmentation and classification results, a list of trees is generated, with each row containing height and species values. Using height rank tables and the user-specified rank and forest inventory region, the tree diameter is determined. Knowing the tree diameter, the trunk volume is calculated using the stock calculation formulas.

[0060] Forest Inventory Table 1 contains the relationship between diameters, heights, species, and ranks. The rank is specified by the user, along with the species and height. PCI7RU2025 / 000056

[0061] The algorithm determines the height in the previous steps. Thus, knowing the breed and rank (categorical parameter), it is only necessary to find the closest height value in the table (which is contained in the table cells, with diameters in the rows and ranks in the columns) based on the minimum difference from the height actually detected by the algorithm, taken as the absolute value.

[0062] Based on the results obtained at the previous stage, the forest inventory is calculated for the study area for the main and second tiers - the total reserve, the reserve by species, and the reserve by trunk thickness.

[0063] Below are examples of specific implementations and experimental results: The experiments were conducted at six test sites in the Arkhangelsk Region. During the experiments, complete tree counts were conducted using in-kind methods in accordance with the requirements of Order No. 688 of the Ministry of Natural Resources of the Russian Federation dated October 17, 2022. Following this, aerial laser scanning was performed, and the results were processed using the described algorithms. The experimental results are presented in Forest Inventory Table 1.

[0064] Forest taxation table 1

[0065]

[0066] The results of the proposed method of forest stand inventory based on aerial laser scanning indicate the achievement of the main technical result, namely, an error in the reserve of less than 10% of the reference volume.

[0067] These application materials present a preferred implementation of the claimed technical solution, which should not be used as limiting other particular embodiments of its implementation that do not go beyond the scope of the requested scope of legal protection and are obvious to specialists in the relevant field of technology.

Claims

CLAUSES OF THE INVENTION L A method for forest stand assessment based on airborne laser scanning data, comprising the following stages: - receive data from ultra-dense aerial laser scanning of forest stands using a drone; - stitch the obtained data into a point cloud and divide this cloud into sections of a certain size; - search for tree tops in the obtained areas and detect the tops of the main tier trees using the local maximum algorithm on the Z-projection of the area with filtering by the median filter; - after which they perform tree segmentation based on the detected vertices, during which: • detection of the second tier is carried out by monitoring the reflection density waves for each segmented tree, which is a set of points, and the second tier is cut off; • for each segmented tree, the number of trees belonging to it is checked, in part, based on the raster image of the heights, the Voronoi diagram is calculated and the segmentation is calculated using the watershed method, after which the final result is taken as the intersection of pixels belonging to the same class according to both methods, the re-segmented trees are separated; • for the final segmented trees, data is generated for their classification by species and, based on the generated vector, classification is performed using the support vector method using a trained machine learning model; • a list of trees is formed based on the results of segmentation and classification, indicating the height value and the name of the species, and based on tables of height categories, the height category of the stand for each species and forest inventory area, the diameter of the tree is selected and, using formulas for calculating the reserve, the volume of the tree trunk is calculated; • based on the results obtained at the previous stage, the forest inventory is calculated for the area under study for the main and second tiers - total reserve, reserve by species, reserve by trunk thickness. ii

Images