Forest stand volume remote sensing image estimation and inversion method and system

Abstract

This invention discloses a method and system for estimating and inverting forest stand volume using remote sensing images, relating to the fields of remote sensing image processing and forest resource surveys. The method includes: acquiring three-dimensional point clouds using UAV lidar and generating a canopy height model through progressive morphological filtering; extracting three-dimensional structural parameters; extracting texture features from near-infrared and red-edge bands of multispectral satellite imagery based on the gray-level co-occurrence matrix; fusing the three-dimensional structural parameters and texture features and adaptively filtering them through a closed-loop feature importance feedback mechanism; and constructing a stochastic forest regression model with measured ground volume as the dependent variable to perform pixel-by-pixel volume inversion and generate a spatial distribution map. This invention overcomes spectral saturation limitations and improves the estimation accuracy of high-volume forest stands.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and forest resource survey technology, and in particular to a method and system for estimating and inverting forest stand volume from remote sensing images. Background Technology

[0002] Forest stock volume is a core indicator for measuring the quantity and quality of forest resources, and a crucial foundational parameter for forestry carbon sequestration measurement, timber resource assessment, and ecosystem service function quantification. Accurately obtaining regional-scale forest stock volume information is of great significance for sustainable forest management, carbon cycle research, and forest carbon sequestration accounting in the context of global climate change. Traditional forest stock volume survey methods primarily rely on ground-based plot-by-plot measurements, specifically measuring the diameter at breast height (DBH) and height of each tree in designated standard plots. The stock volume of each plot is then estimated using binary volume tables or allometric growth equations, and finally extrapolated to the regional scale using sampling theory. However, this method is limited by factors such as the number of plots, spatial distribution, and insufficient representativeness, resulting in sampling errors exceeding 15% in large-area forest stock volume estimations. Furthermore, the survey cycle is long and costly, making it difficult to meet the needs of high-frequency dynamic monitoring of stock volume over large areas.

[0003] While satellite optical remote sensing technology offers wide coverage and short revisit cycles, providing macroscopic information for large-area forest monitoring, it can only acquire spectral reflectance information from the canopy surface and cannot directly measure key parameters closely related to individual timber volume, such as trunk diameter and tree height. Regression models for timber volume using vegetation indices like the Normalized Difference Vegetation Index (NDVI) as independent variables can achieve acceptable estimation accuracy in stands with low canopy closure. However, when canopy closure exceeds 0.7, the spectral signal tends to saturate—the vegetation index no longer increases significantly with increasing timber volume, and the model's predictions severely underestimate the actual timber volume, with biases exceeding 30% in high-volume mature forests. This spectral saturation problem has become a core bottleneck restricting the application of optical remote sensing in estimating timber volume in medium- and high-density forest stands. In recent years, although some studies have attempted to introduce synthetic aperture radar backscattering coefficients or optical image texture features as auxiliary variables to alleviate the saturation effect, the improvement in prediction accuracy for high-volume forest stands is limited in the absence of three-dimensional structural information.

[0004] LiDAR technology, due to the penetrating power of its laser pulses through the canopy, can directly acquire three-dimensional structural information from the top of the canopy to the forest floor, and is considered an effective means to overcome the spectral saturation bottleneck. However, existing LiDAR-based forest parameter inversion studies have mostly focused on aboveground biomass estimation, with insufficient attention paid to dedicated inversion methods and system architectures for volumetric volume, a classic forestry survey parameter. Furthermore, existing methods still have significant room for improvement in terms of feature extraction dimensions and model optimization strategies.

[0005] Patent CN108921885A discloses a method for jointly inverting forest aboveground biomass by integrating high-resolution CCD data, hyperspectral imagery, and lidar point cloud data. This method requires simultaneously acquiring three data sources: high-resolution CCD images, hyperspectral images, and lidar point clouds. Texture features, spectral features, and point cloud structural features are extracted respectively, and a random forest model is then established for prediction. However, this approach has the following shortcomings: First, this method relies on an aircraft platform to simultaneously acquire the three data sources, resulting in high data acquisition costs and poor operational flexibility, making it difficult to apply to repeated monitoring in forest areas with complex terrain or small areas. Second, this method estimates aboveground biomass rather than volume, and does not establish a dedicated inversion model for volume, a classic forestry survey parameter. Third, the extraction of point cloud structural features in this method only uses canopy height percentiles and Weibull profile fitting, failing to construct a canopy volume density profile that can finely characterize the differences in the vertical structure of forest stands. Fourth, this method lacks a feature adaptive selection mechanism based on model feedback, and cannot dynamically optimize feature combinations based on model training results. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as the severe underestimation of stand volume due to spectral saturation in optical remote sensing and the lack of adaptive feedback mechanisms in feature selection and model optimization in multi-source data fusion and inversion methods, this invention provides a method and system for estimating and inverting stand volume using remote sensing images. This method fuses the three-dimensional canopy structure parameters of the stand extracted from UAV-borne lidar point clouds with texture features from multispectral satellite images, and achieves high-precision spatial inversion of stand volume through random forest regression modeling and feature importance feedback loop.

[0007] This invention provides a method for estimating and inverting forest stand volume using remote sensing images, comprising the following steps: a three-dimensional point cloud data acquisition and preprocessing step, in which a UAV equipped with a lidar scanner flies over the target forest stand to acquire three-dimensional point cloud data, separates ground points through progressive morphological filtering and generates a digital elevation model, and constructs a canopy height model after normalizing the point cloud; a three-dimensional canopy structure parameter extraction step, in which three-dimensional structure parameters, including average canopy height, canopy height standard deviation, canopy height percentile, canopy cover, and canopy volume density profile, are extracted based on the canopy height model; and multispectral satellite imagery. The textural feature extraction step extracts uniformity, contrast, and correlation texture features based on the gray-level co-occurrence matrix for the near-infrared and red-edge bands. The multi-source feature fusion and adaptive screening step concatenates the 3D structural parameters and texture features into a comprehensive feature vector and filters them based on feature importance scores. The random forest regression modeling and volume spatial inversion step constructs a random forest regression model with the measured volume of ground sample plots as the dependent variable, predicts the volume pixel by pixel after cross-validation, and generates a spatial distribution map. The feature importance ranking output by the random forest regression model is fed back to the feature screening step for iterative optimization.

[0008] Another aspect of this invention provides a remote sensing image estimation and inversion system for forest stand volume, including a point cloud acquisition and preprocessing module, a three-dimensional structural parameter extraction module, a texture feature extraction module, a feature fusion and filtering module, and a modeling and inversion module. Specifically, the point cloud acquisition and preprocessing module acquires three-dimensional point cloud data using a lidar scanner mounted on a UAV and outputs a normalized point cloud and a canopy height model after preprocessing; the three-dimensional structural parameter extraction module extracts multi-dimensional structural parameters, including canopy height percentiles and canopy volume density profiles, based on the canopy height model; the texture feature extraction module extracts gray-level co-occurrence matrix texture features from the near-infrared and red-edge bands of multispectral satellite imagery; the feature fusion and filtering module stitches and fuses multi-source features and outputs an optimized feature subset through adaptive filtering; and the modeling and inversion module constructs a random forest regression model and performs spatial inversion of the entire forest stand volume. Each module corresponds one-to-one with the corresponding steps in the above method, and the feature importance ranking output by the modeling and inversion module is fed back to the feature fusion and filtering module to achieve closed-loop optimization.

[0009] The beneficial effects of this invention are as follows: Extracting multi-dimensional structural parameters, including canopy volume density profiles, from 3D point clouds acquired by UAV lidar effectively overcomes the spectral saturation limitations of optical remote sensing in high-canopy-density forest stands; the fusion of 3D structural parameters and spectral texture features allows the model to simultaneously consider vertical structural information and horizontal spatial heterogeneity information, achieving a synergistic effect greater than the sum of its parts; the adaptive selection mechanism based on feature importance feedback enables dynamic optimization of feature combinations during model training, further improving the accuracy of volume inversion. Furthermore, this invention uses a UAV platform instead of a traditional manned aircraft platform for lidar data acquisition, significantly reducing data acquisition costs and operational barriers, making repeated monitoring of small to medium-sized forest areas possible. The canopy volume density profile parameters in this invention can precisely distinguish the differences in vertical structure between uniformly layered pure forests and multi-layered mixed forests, compensating for the shortcomings of traditional point cloud statistical parameters, which only provide general descriptions. The reverse feedback connection between the modeling and inversion module and the feature fusion and selection module in this invention's system achieves an end-to-end automated optimization closed loop, reducing manual intervention and improving the system's practicality and operability. Attached Figure Description

[0010] Figure 1 This is a flowchart of the forest stand volume remote sensing image estimation and inversion method provided in the embodiments of the present invention.

[0011] Figure 2 This is an architecture diagram of the forest stand volume remote sensing image estimation and inversion system provided in this embodiment of the invention. Detailed Implementation

[0012] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0013] Reference Figure 1 As shown, this embodiment of the invention provides a method for estimating and inverting forest stand volume using remote sensing images. This method uses three-dimensional canopy structure parameters extracted from UAV-borne lidar point clouds and texture features from multispectral satellite imagery as core data sources. It achieves high-precision spatial inversion of forest stand volume through random forest regression modeling and a feature importance feedback loop. In this embodiment, the method specifically includes the following steps.

[0014] Step S1: Acquisition and preprocessing of 3D point cloud data. The goal of this step is to acquire high-density 3D point cloud data of the target forest stand area and convert the original point cloud into a normalized canopy height model that can directly reflect the 3D structure of the forest stand canopy through a series of preprocessing procedures.

[0015] In one embodiment of the present invention, a drone equipped with a lightweight LiDAR scanner is used for data acquisition. Preferably, the drone's flight altitude is set to 100-150m, the flight speed is controlled at 35m / s, and the flight path overlap rate is not less than 30%, to ensure that the point density of the acquired 3D point cloud data is not less than 10pts / m. 2 This invention chooses an unmanned aerial vehicle (UAV) platform instead of a traditional manned aircraft platform because: UAVs can fly at low altitudes, resulting in shorter laser pulse propagation distances and smaller divergence angles, leading to a more concentrated laser footprint and thus a higher ground point density at the same laser emission frequency; simultaneously, UAVs offer flexible takeoff and landing, lower operating costs, and are particularly suitable for repeated monitoring in forest areas with complex terrain or limited areas. In this embodiment, the lidar scanner operates at a wavelength of 905nm or 1550nm, with a scanning frequency of no less than 100kHz, supporting a maximum of three echo recordings, and a field of view of 60°~70°. Each laser pulse hitting the canopy may generate multiple echoes—the first echo typically originates from the top of the canopy, the middle echoes come from different layers of branches and leaves within the canopy, and the final echo penetrates the canopy to reach the ground or the understory shrub layer. This multiple echo information provides crucial information for the subsequent accurate separation of ground points from vegetation points.

[0016] After acquiring the original 3D point cloud data, noise point removal preprocessing is performed first. During the flight of the drone above the forest stand, laser pulses may scatter onto birds, suspended leaf debris, or atmospheric particles, generating false echoes. These echo points appear as outliers in 3D space, far removed from the main point cloud distribution. Failure to remove these outliers will severely interfere with the accurate construction of the subsequent canopy height model. In this embodiment, a statistical outlier filtering method is used. For each point, the average distance to its K nearest neighbors is calculated. Points with an average distance exceeding the global average distance plus twice the standard deviation are marked as noise points and removed, where K is set to 20. This step effectively removes outlier noise points caused by laser pulse scattering onto birds, insects, or suspended particles, avoiding the introduction of false canopy height extrema in subsequent processing. In one embodiment of this invention, the noise point removal ratio is typically 0.5% to 2% of the total original point cloud data, and the remaining effective point cloud data completely covers the forest stand canopy and ground area.

[0017] After noise removal, the crucial ground point separation step is performed. This invention employs a progressive morphological filtering algorithm to classify ground points in the point cloud. The core idea of ​​this algorithm is to perform morphological opening operations on the point cloud elevation surface using progressively increasing structuring element sizes. Specifically, let the initial radius of the structuring element be... Maximum radius Increasing step size at each level Simultaneously, set the elevation difference threshold function and the initial elevation difference threshold. Final height difference threshold The elevation difference threshold increases linearly with the size of the structuring element. In each level of opening operation, points whose difference between the point cloud elevation and the opening operation result exceeds the current elevation difference threshold are marked as non-ground points, while points whose difference is less than the elevation difference threshold are retained as ground point candidates. After iterative processing through all levels, a complete set of ground points is finally obtained. The reason why this invention chooses progressive morphological filtering instead of other common methods such as cloth simulation or triangular irregular network encryption is that: progressive morphological filtering processes multi-scale structuring elements step by step, which can simultaneously adapt to the ground point recognition needs in both flat and undulating terrains, and performs particularly robustly in mountainous forest areas with complex understory terrain. In one embodiment of this invention, the specific form of the elevation difference threshold function is: ,in: For the first The elevation difference threshold for each level, in meters; The initial elevation difference threshold is set to 0.3m, which corresponds to the upper limit of the elevation difference between a ground point on a flat surface and low-lying grass. The final elevation difference threshold is set at 5m, which corresponds to the maximum reasonable elevation difference between adjacent ground points on a steep hillside. The current iteration level, with a value ranging from 0 to... ; Let be the total number of iterations, given by Calculations show that in this embodiment... The technical advantage of this linear incremental strategy is that a stricter elevation difference threshold is used in the small-scale structural element stage, which can accurately identify ground points in flat areas; as the size of the structural element increases, the elevation difference threshold is gradually relaxed, so that real ground points in steep slope areas will not be misidentified as non-ground points.

[0018] Based on the separated set of ground points, a digital elevation model (DEM) is generated using inverse distance weighted interpolation (IRW) or kriging interpolation, with a spatial resolution of 0.5m. In this embodiment, IWW is preferred, with a search radius of 3m and a weighting exponent of 2. After the DEM is generated, the original elevation value of each point in the filtered point cloud is subtracted from the ground elevation value corresponding to the point's projection on the DEM, thus completing the point cloud normalization process. The normalized point cloud elevation value represents the true height of that point above the ground, eliminating the interference of terrain undulations on subsequent canopy structure analysis.

[0019] Based on the normalization process, this step further constructs a canopy height model. The canopy height model represents the height distribution of the canopy's upper surface in the form of a regular grid; in this embodiment, the grid resolution is set to 1m. For each grid cell, the maximum height value among all normalized point clouds falling within that cell's range is taken as the canopy height of that grid. Preferably, the generated canopy height model is subjected to a 3×3 window Gaussian smoothing filter to eliminate local spike noise caused by uneven point cloud sampling, while maintaining the overall morphological characteristics of the canopy's upper surface. The canopy height model, as the core output of step S1, provides a spatially continuous foundation of canopy height information for the subsequent extraction of three-dimensional structural parameters in step S2. It is worth noting that the accuracy of the digital elevation model in step S1 directly affects the accuracy of point cloud normalization, which in turn affects the accuracy of the canopy height model and the extraction accuracy of all subsequent structural parameters. In this embodiment, the digital elevation model was validated by selecting several known elevation control points within the study area. The root mean square error of the elevation was controlled within 0.15m, meeting the accuracy requirements for canopy height in stand volume inversion. Furthermore, the spatial resolution of the canopy height model was set to 1m instead of a higher resolution (such as 0.5m) to match the point cloud density. This was because while a higher resolution could retain more canopy surface details, it would also introduce more local noise due to the randomness of point cloud sampling, and there would be a significant scale mismatch with the subsequent parameter statistics at the 30m×30m plot scale. Therefore, the 1m resolution canopy height model achieved a suitable balance between information preservation and noise suppression.

[0020] Step S2: Extraction of 3D Canopy Structure Parameters. This step, based on the canopy height model and normalized point cloud output from Step S1, uses a pre-defined sample plot grid as the statistical unit to extract a set of multidimensional parameters reflecting the 3D structural characteristics of the forest stand. These parameters characterize the vertical distribution and horizontal cover features of the forest canopy from different perspectives and are key independent variables for subsequent random forest regression modeling.

[0021] In this embodiment, the grid size of the sample plot is set to 30m × 30m, consistent with the size of the actual measured sample plot on the ground, ensuring that the remote sensing extracted parameters and the measured ground accumulation volume strictly correspond in spatial scale. Within each grid cell, the following five types of three-dimensional structural parameters are extracted.

[0022] The first type of parameters are the mean canopy height and the standard deviation of canopy height. Mean canopy height Defined as the arithmetic mean of all valid grid cells (grid cells with height greater than 0) within the canopy height model, reflecting the overall height level of the stand. Canopy height standard deviation. This quantifies the dispersion of canopy height within the grid; a larger value indicates more significant differences in tree height within the stand, typically indicating a multi-layered forest structure. The combination of these two parameters can distinguish between uniform forests (…). high, Low) and multi-layered forest ( medium, Two typical forest stand structure types (high) and high.

[0023] The second type of parameter is the canopy height percentile. In this embodiment, nine height values ​​are extracted from the 10th to the 90th percentile, with an interval of 10, and denoted as... , , , , , , , and The percentile is calculated as follows: All normalized point clouds within the grid are arranged in ascending order of height value, and the percentile is... Percentile is the number of digits after arranging. The height values ​​of the points, among which This represents the total number of point clouds within the grid. The canopy height percentile series constitutes a segmented description of the canopy's vertical height distribution, with lower percentiles (such as...) , ) Reflects the height level of understory shrubs and saplings in the stand, median percentile (e.g. Approximates the median canopy height, high percentile (e.g.) , The height of the dominant tree in the stand is close to that of the subspecies. Preferably, the percentile combination can accurately distinguish the vertical structural differences between contour-isophyte forests and multi-layered mixed forests: in contour-isophyte forests, the differences between percentiles are small, and the distribution curve shows a steep upward trend; while in multi-layered forests, the differences between low and high percentiles are significant, and the distribution curve shows a gentle upward trend. This distinguishing ability is completely unattainable by pure spectral vegetation indices, because under the same canopy closure conditions, the canopy surface reflectance spectra of contour-isophyte forests and multi-layered forests are almost indistinguishable.

[0024] The third type of parameter is canopy coverage. This invention defines canopy coverage. For grid cells with heights exceeding a preset threshold The proportion of echo points to total echo points. In this embodiment, a preset height threshold is used. The selection of this threshold is based on the fact that in most forest stands, echo points below 2m mainly originate from the ground, herbaceous layer, and low shrub layer, while echo points above 2m mainly originate from the tree canopy. The formula for calculating canopy coverage is: ,in: The canopy coverage is a dimensionless quantity, ranging from 0 to 1. For the normalized height within the grid greater than The number of echo points; This represents the total number of echo points within the grid. Canopy cover describes the density of the forest canopy in the horizontal direction, similar to the concept of canopy closure in traditional forestry surveys, but with a more objective and accurate calculation. In volume inversion models, canopy cover can effectively supplement the insufficient information of canopy height parameters. For stands with similar tree heights but significant differences in density, canopy cover can provide crucial distinguishing information.

[0025] The fourth type of parameter is the canopy volume density profile. The canopy volume density profile is an innovative parameter in this invention used to finely characterize the vertical structure of a forest stand. Its construction method is as follows: the canopy height range within the grid (from 0 to the maximum canopy height) is... Divided into equal parts Each layer segment, in this embodiment The value is 10. For the... Each layer has a height range of [number] segments. Count the number of normalized point cloud echo points falling within this segment. Calculate the normalized density ratio of this layer segment: ,in: For the first The volume density ratio of the layers, dimensionless. Values ​​range from 1 to ; For the first The number of echo points within a segment; This is the sum of the number of echo points across all segments. to Composition The canopy volume density profile, represented by the dimensional vector, describes the distribution of echo points along the vertical direction of the canopy. In pure coniferous forests, the canopy volume density profile exhibits a top-concentrated distribution, with the density proportions of the upper layers significantly higher than those of the lower layers. In broadleaf mixed forests, the density distribution is relatively uniform. In multi-layered, uneven-aged forests, the density profile may show a bimodal or even multimodal distribution. The technical advantage of the canopy volume density profile lies in its expansion of the canopy vertical structure information, which is traditionally simplified using only a few statistical quantities (such as average height and maximum height), into a refined, layered description. This allows random forest models to capture the complex nonlinear relationships between different vertical structure types and volume, thereby improving their ability to distinguish stands with significant differences in volume but similar average heights.

[0026] The five types of parameters mentioned above collectively form a high-dimensional three-dimensional structural parameter vector. Taking this embodiment as an example, the dimensions of the parameter vector are: 1 dimension of average canopy height + 1 dimension of canopy height standard deviation + 9 dimensions of canopy height percentiles + 1 dimension of canopy cover + 10 dimensions of canopy volume density profile = 22 dimensions. This parameter vector comprehensively characterizes the three-dimensional structural features of the forest stand from three aspects: height statistics, vertical distribution, and horizontal cover.

[0027] Step S3: Multispectral satellite image texture feature extraction. This step acquires multispectral satellite images spatially matching the point cloud data acquisition area in Step S1, and extracts texture features based on the gray-level co-occurrence matrix to supplement the horizontal spatial heterogeneity information of the canopy. Three-dimensional structural parameters mainly describe the vertical distribution characteristics of the canopy, but they are insufficient in depicting the spatial texture variations of the canopy in the horizontal direction (such as the distribution pattern of canopy gaps, the spectral heterogeneity of canopies of different tree species, etc.). The introduction of texture features is precisely to fill this information gap, enabling the comprehensive feature vector to simultaneously take into account both vertical structural information and horizontal spatial heterogeneity information, achieving complementary synergy between the two types of data sources.

[0028] In one embodiment of the present invention, the multispectral satellite imagery is preferably obtained from the Sentinel-2 satellite or a similar multispectral satellite platform with a spatial resolution of 10-20m. The present invention specifically selects the near-infrared band (wavelength range approximately 785-900nm) and the red-edge band (wavelength range approximately 705-745nm) as the input bands for texture feature extraction because: the near-infrared band is highly sensitive to the cell structure and water content of canopy leaves, reflecting spatial variations in canopy health and density; the red-edge band, located in the steep spectral slope region between visible red light and near-infrared, is extremely sensitive to changes in chlorophyll content and canopy phenological status, possessing unique advantages in distinguishing canopies of different tree species and canopies at different growth stages. Compared to traditional research methods that only use panchromatic or visible light bands to extract texture features, the present invention's combination of near-infrared and red-edge bands embeds richer vegetation physiological information into the texture information.

[0029] Texture features are extracted based on the gray-level co-occurrence matrix (GLCM). The GLCM describes the frequency distribution of pixel gray values ​​co-occurring at specific directions and distances in an image. In this embodiment, the calculation parameters of the GLCM are set as follows: the calculation window size is 3×3 pixels. This size is selected to ensure that the spatial range of texture feature extraction strictly matches the statistical units of the three-dimensional structural parameters in scale. For the Sentinel-2 near-infrared band (spatial resolution 10m), the 3×3 pixel window corresponds to a 30m×30m spatial range on the ground, which is exactly consistent with the grid size of the sample plot in step S2. For the red-edge band (spatial resolution 20m), the 3×3 pixel window corresponds to a 60m×60m ground. After texture extraction, it is resampled to a 30m×30m grid unit through bilinear interpolation for spatial alignment, thereby ensuring strict correspondence of features from different sources in spatial scale. The step size is set to 1 pixel. The calculation directions include four directions: 0°, 45°, 90°, and 135°. The final texture feature is the average of the four directions to eliminate directional bias.

[0030] For each band's gray-level co-occurrence matrix, three texture features are extracted: homogeneity, contrast, and correlation. Homogeneity (also known as inverse difference moment) reflects the consistency of image gray values ​​within a local area. Its calculation involves the inverse relationship between the value of each element in the gray-level co-occurrence matrix and its distance from the main diagonal. Homogeneity approaches 1 when gray-level changes are gradual and approaches 0 when gray-level changes are drastic. At the stand scale, higher homogeneity values ​​correspond to pure forests with uniform canopy structure, while lower values ​​indicate mixed forests with strong canopy spatial heterogeneity or stands with obvious canopy gaps. Contrast quantifies the intensity of gray-level differences between neighboring pixels in the image. A higher contrast value indicates more drastic changes in lightness and darkness on the canopy surface in space. Such drastic changes usually correspond to significant canopy gaps or staggered canopies of different heights within the stand. Correlation describes the degree of linear correlation between row and column elements in the gray-level co-occurrence matrix, reflecting the spatial regularity of canopy texture in a specific direction. Preferably, areas with high correlation values ​​typically correspond to plantations or regularly arranged stands that have undergone tending and management, while areas with low correlation values ​​are mostly natural forests or mixed forests of different ages. The three texture features characterize the horizontal spatial structure of the canopy from a complementary perspective: uniformity focuses on describing local consistency, contrast focuses on describing local differences, and correlation focuses on describing spatial relationships.

[0031] In this embodiment, the above three texture features are extracted from the near-infrared band and the red-edge band respectively, resulting in a total of [number missing] texture features. These six texture features supplement the horizontal spatial information that was not covered by the three-dimensional structural parameters in step S2 from different dimensions. Uniformity describes the spatial homogeneity of the canopy, contrast describes the intensity of spatial variation in the canopy, and correlation describes the spatial structural regularity of the canopy. Together, they constitute a complete characterization of the horizontal spatial heterogeneity of the canopy.

[0032] Step S4: Multi-source feature fusion and adaptive selection. This step concatenates the 22-dimensional structural parameter vector extracted in Step S2 with the 6-dimensional texture feature vector extracted in Step S3 to form a 28-dimensional comprehensive feature vector. An adaptive selection mechanism based on feature importance scoring is then used to optimize the feature combination. The purpose of feature fusion is to unify features from different sensors and with different physical meanings into the same feature space, enabling subsequent models to simultaneously utilize vertical structural information and horizontal texture information for accumulation prediction.

[0033] Before feature fusion, this embodiment first standardizes all features. Since the dimensions and numerical ranges of 3D structural parameters (e.g., average canopy height is typically between 5 and 30 meters, and canopy coverage is between 0 and 1) may differ significantly from the numerical range of texture features, features with larger numerical ranges may have excessive weight in subsequent models without standardization. This embodiment uses the Z-score standardization method, subtracting the mean of each feature across all plots and dividing by the standard deviation, transforming all features to a uniform scale with a mean of 0 and a standard deviation of 1. It is worth noting that the random forest algorithm itself is not sensitive to feature scale (because its splitting criterion based on decision trees depends only on the order of feature values ​​rather than their absolute size), but standardization is beneficial for the comparability analysis of subsequent feature importance scores.

[0034] Feature selection is the core of this step. This invention employs an adaptive selection mechanism based on a feature importance scoring method that reduces accuracy through permutation of out-of-bag data within a random forest. Specifically, after the random forest model is trained, the following operations are performed on each feature: the value of that feature in the validation set (out-of-bag data) is randomly permuted (i.e., the correspondence between the feature and the dependent variable is shuffled), and the decrease in model prediction accuracy after the permutation relative to before the permutation is calculated. This decrease is defined as the importance score of the feature. The larger the decrease, the greater the contribution of the feature to the model's predictive ability, and the higher its importance. In this embodiment, a preset threshold is set to 50% of the average importance score of all features. Features with importance scores below this threshold are considered redundant features that contribute little to the accumulation prediction and are discarded.

[0035] A key innovation of this step lies in the introduction of an adaptive feedback mechanism. Specifically, the feature importance ranking results output after the random forest regression model training in step S5 are fed back to step S4 as the basis for updating the feature selection in the next round. In the first round of modeling, all 28-dimensional features participate in model training, and the model outputs an initial feature importance ranking. Based on this, a feature subset (usually 15-20 dimensions) is obtained after selection. The random forest model is then retrained with this feature subset to obtain an updated feature importance ranking. This process is iterated for 2-3 rounds until the feature subset stabilizes and no longer changes. In one embodiment of this invention, the iteration termination condition is that the feature subsets after two consecutive rounds of selection are completely consistent, or the number of iterations reaches a preset upper limit of 3. The technical effect of this adaptive feedback mechanism is that it eliminates the importance score bias caused by the interference of redundant features in the initial feature set, enabling the final selected feature subset to more accurately reflect the true contribution of each feature to the accumulation prediction, thereby improving the generalization performance of the model. In one embodiment of the present invention, after three rounds of iterative optimization, the original 28-dimensional features were reduced to 15-18 dimensions, and the out-of-bag prediction error of the model was reduced by approximately 8%-12% compared to the first round of full-feature modeling. In practical applications, this feedback mechanism can also help researchers identify the feature combinations that contribute most to the prediction of stock volume for specific forest stand types. For example, in pure coniferous forests, the high percentile of canopy height and canopy cover rank highly in importance; while in multi-layered broadleaf mixed forests, the importance of the mid-layer features of the canopy volume density profile increases significantly. This information has important scientific reference value for understanding the main controlling factors of stock volume in different forest stand types.

[0036] Step S5: Random forest regression modeling and spatial inversion of stock volume. In this step, the measured stock volume of the ground sample plots is used as the dependent variable, and the feature subset selected in step S4 is used as the independent variable to construct a random forest regression model. After evaluating the accuracy of the model through cross-validation, it is applied to the whole forest area image to predict the stock volume pixel by pixel, and finally generates a spatial distribution map of stock volume.

[0037] The method for obtaining ground plot volume is as follows: Within the target forest stand area, 30m × 30m square plots are set up according to the principles of systematic sampling or stratified random sampling. Within each plot, all trees with a diameter at breast height (DBH) greater than 5cm are measured individually, and the tree species, DBH, and tree height are recorded. Individual timber volume is obtained using the local binary volume table or allometric growth equation. The calculation yielded, where For the first The volume of a tree, in meters. 3 ; Diameter at breast height, in cm; Tree height, in meters; , , These are the parameters for the volume equation corresponding to the tree species. The plot volume is the sum of the volumes of all individual trees within the plot, converted to hectares, and expressed in m³. 3 / hm 2 .

[0038] The construction process of the random forest regression model is as follows: The random forest algorithm is an ensemble learning method that constructs a large number of decision trees and averages their predictions to obtain the final regression prediction value. In this embodiment, the key hyperparameter of the model is set as: the number of decision trees. The value was determined through preliminary experiments. When the number of decision trees gradually increased from 100 to 500, the out-of-bag error of the model tended to stabilize around 300 trees. Further increasing the number of trees only increased the computational cost with a slight improvement in accuracy. Considering both prediction accuracy and computational efficiency, the optimal range for the number of decision trees was 200 to 500. The number of candidate features randomly selected when each decision tree splits at a node... ,in The total dimension of the filtered feature subset. This indicates rounding down, a classic empirical value in regression problems. The minimum number of leaf node samples is set to 3-5, with 5 being the preferred value in this embodiment to prevent overfitting caused by excessive growth of the decision tree. The rationale for setting the minimum number of leaf node samples to 3-5 is that a value that is too small (e.g., 1-2) will cause the decision tree to overfit noisy samples, while a value that is too large (e.g., above 10) will cause the model to underfit and lose the local nonlinear mapping relationship between forest stand structure parameters and stock volume. 3-5 is the preferred range that balances fitting accuracy and generalization ability.

[0039] During model training, the Random Forest algorithm performs bootstrap sampling on the training samples, from a total of Samples were drawn with replacement from each sample site. Each decision tree has 100 samples forming its training set, while the unselected samples (approximately 36.8%) are used as out-of-bag data for independent validation of that tree. During the growth of each decision tree, for each node to be split, starting from the current... Select the feature and split point that maximizes the reduction in impurity (measured by mean squared error) of the child nodes after splitting from the candidate features, and perform the splitting. Continue growing the tree until the minimum number of leaf node samples is reached. The arithmetic mean of the predictions from each decision tree is the final prediction value of the random forest. ,in: This is a predicted value for forest stand volume, in meters. 3 / hm 2 ; This represents the total number of decision trees in the random forest; in this embodiment, it is set to 300. For the first Each decision tree is used to process the input feature vector. The prediction results; This is the filtered feature subset output from step S4. The technical advantage of ensemble voting across multiple decision trees is that the prediction of a single decision tree is easily affected by fluctuations in the training samples, leading to overfitting. However, the average result of a large number of decision trees can effectively eliminate the random errors of individual trees, thereby obtaining stable prediction results with good generalization ability.

[0040] After model training, 10-fold cross-validation was used to evaluate the model's generalization performance. All sample data were randomly divided into 10 equal parts. In each iteration, 9 parts were used for model training and 1 part for validation, repeating this process 10 times so that each part served as the validation set once. The model accuracy evaluation metric for 10-fold cross-validation included the coefficient of determination. Root mean square error and relative root mean square error They are defined as follows: ,in: The coefficient of determination is dimensionless and ranges from 0 to 1. The closer the value is to 1, the higher the model fit. For the first The measured volume of each verification plot, in cubic meters. 3 / hm 2 ; For the first The model predicts the volume of sediment in the validation plots, in cubic meters. 3 / hm 2 ; This is the arithmetic mean of the measured volume of all verification plots, in meters. 3 / hm 2 ; To verify the number of sample plots. ,in: The root mean square error is expressed in meters. 3 / hm 2 This reflects the absolute amount of the average deviation between the predicted and measured values. ,in: The relative root mean square error, expressed as a percentage, eliminates the influence of the absolute magnitude of the accumulation on error evaluation, facilitating accuracy comparisons between different studies. In one embodiment of the invention, when cross-validation... Not less than 0.80 and If the accuracy is no more than 20%, the model is considered to meet the application requirements.

[0041] After successful cross-validation, the trained random forest regression model was applied to predict volume per pixel in the entire forest imagery. Specifically, the prediction range was defined as the common coverage area of ​​the canopy height model and the satellite imagery. A 30m × 30m moving window was used to slide pixel-by-pixel across the entire forest area. Three-dimensional structural parameters and texture features were extracted for each window location and input into the model to obtain the predicted volume value for that location. All predicted values ​​constituted a volume raster map with the same spatial resolution as the input imagery.

[0042] Finally, the accumulation raster map is rendered using a color gradient to generate a spatial distribution map of the accumulation. In this embodiment, the color gradient is set from light green (corresponding to low accumulation areas, such as 0~50m) to light green. 3 / hm 2 (Through dark green areas (medium-volume areas, such as 50-150m)) 3 / hm 2 ) to dark brown (high accumulation areas, such as greater than 150m) 3 / hm 2 The spatial distribution map of forest stock volume visually displays the spatial distribution pattern of forest stock volume throughout the entire forest area, providing direct spatial decision support for forestry managers to formulate regional-scale forest management plans and identify mature stands with high stock volume and young stands with low stock volume that urgently need tending. Preferably, the spatial distribution map of forest stock volume can also be overlaid with vector information such as administrative division boundaries and forest compartment boundaries to realize the function of calculating the total stock volume by management unit, thereby providing quantitative spatial data support for forest resource inventory and forest land asset assessment. In one embodiment of the present invention, the spatial distribution maps of forest stock volume obtained at different times can also be interpolated to generate a stock volume change detection map, which is used to monitor the spatiotemporal dynamic changes of forest stock volume and provide time-series comparative data for carbon sink increment accounting and forest management effect assessment.

[0043] Reference Figure 2 As shown, this embodiment of the invention also provides a remote sensing image estimation and inversion system for forest stand volume, which corresponds one-to-one with each step in the aforementioned method embodiments. In this embodiment, the system includes a point cloud acquisition and preprocessing module, a three-dimensional structural parameter extraction module, a texture feature extraction module, a feature fusion and filtering module, and a modeling and inversion module.

[0044] Step S1 in the point cloud acquisition and preprocessing module method is used to acquire three-dimensional point cloud data of the target forest stand using a UAV equipped with a LiDAR scanner. The three-dimensional point cloud data is then subjected to noise removal, progressive morphological filtering to separate ground points, generation of a digital elevation model based on the ground points, normalization of the point cloud, and construction of a canopy height model. In one embodiment of the invention, the point cloud acquisition and preprocessing module includes a data acquisition submodule and a preprocessing submodule. The data acquisition submodule controls the UAV to fly along a preset route and receives the three-dimensional point cloud data stream transmitted back by the LiDAR scanner in real time, simultaneously recording GNSS positioning information and inertial navigation attitude data to achieve accurate spatial positioning of the point cloud data. The preprocessing submodule receives the raw point cloud data output by the data acquisition submodule and sequentially performs statistical outlier filtering for noise reduction, progressive morphological filtering for ground point separation, inverse distance weighted interpolation to generate a digital elevation model, point cloud normalization, and canopy height model construction. Preferably, the key parameters such as the progressive morphological filtering parameters, digital elevation model resolution, and canopy height model resolution in the preprocessing submodule are consistent with the settings in the method embodiment, and will not be repeated here. The output of the preprocessing submodule includes normalized point cloud dataset and canopy height model raster data, both of which serve as inputs to the 3D structure parameter extraction module.

[0045] The 3D structural parameter extraction module corresponds to step S2 in the method, which is used to extract the 3D structural parameters of the forest stand area based on the canopy height model and normalized point cloud. This module uses a 30m × 30m plot grid as the statistical unit, and calculates the average canopy height, standard deviation of canopy height, and nine canopy height percentiles for each grid. to The module outputs a 22-dimensional three-dimensional structural parameter vector, including canopy coverage and a 10-dimensional canopy volume density profile. The parameter calculation logic, formulas, and thresholds within this module are completely consistent with the description in step S2 of the method embodiment.

[0046] The texture feature extraction module corresponds to step S3 in the method, and is used to acquire multispectral satellite imagery and extract texture features based on the gray-level co-occurrence matrix. This module first performs radiometric correction and geometric registration preprocessing on the input multispectral satellite imagery to ensure precise spatial alignment with the point cloud data. Then, it calculates the gray-level co-occurrence matrix for the near-infrared band and the red-edge band respectively, and extracts three texture indices: uniformity, contrast, and correlation, outputting a total of 6-dimensional texture feature vectors. The calculation parameters of the gray-level co-occurrence matrix (window size, step size, orientation, etc.) are consistent with the settings in step S3 of the method embodiment.

[0047] Step S4 in the feature fusion and filtering module corresponds to concatenating the 22-dimensional parameter vector output by the 3D structural parameter extraction module with the 6-dimensional feature vector output by the texture feature extraction module to form a 28-dimensional comprehensive feature vector, and then performing Z-score normalization and adaptive filtering based on feature importance scoring. This module has an interface for receiving feature importance feedback signals from the modeling and inversion module, which is used to achieve iterative optimization of feature filtering. When an updated feature importance ranking is received, the filtering logic is automatically re-executed and the optimized feature subset is output.

[0048] The modeling and inversion module corresponds to step S5 in the method, used to construct a random forest regression model, perform cross-validation accuracy assessment, and predict the volume of the entire forest area pixel by pixel. This module includes a model training submodule, an accuracy verification submodule, and a spatial inversion submodule. The model training submodule uses the measured volume of ground sample plots as the dependent variable and the selected feature subset as the independent variable, constructs a regression model according to the random forest algorithm parameters described in the method embodiment, and outputs the feature importance ranking results to the feature fusion and screening module. The accuracy verification submodule performs 10-fold cross-validation, calculates evaluation indicators such as the coefficient of determination, root mean square error, and relative root mean square error, and only allows the model to enter the inversion stage when a preset accuracy threshold is met. The spatial inversion submodule applies the validated model to the entire forest area image, extracts features pixel by pixel using a moving window method, predicts the volume, and finally renders a spatial distribution map of the volume using a color gradient method.

[0049] In the system embodiments of this invention, the data flow between modules reflects the architectural features of deep coupling and closed-loop collaboration. The output of the point cloud acquisition and preprocessing module (normalized point cloud and canopy height model) flows simultaneously to the 3D structure parameter extraction module; the outputs of the 3D structure parameter extraction module and the texture feature extraction module converge in the feature fusion and filtering module; the filtered feature subset is input to the modeling and inversion module; and the feature importance ranking output by the modeling and inversion module is then passed back to the feature fusion and filtering module, forming a complete closed loop of extraction-fusion-modeling-feedback-re-filtering, ensuring that the system automatically optimizes feature combinations in multiple iterations and continuously improves the accuracy of accumulation inversion.

[0050] Preferably, the system also includes a data management and visualization module, which is used to uniformly manage the intermediate data and final output results transferred between modules. In terms of data management, the data management and visualization module uniformly stores point cloud data, canopy height models, structural parameter matrices, texture feature matrices, random forest model files, and volume prediction results in a hierarchical file system, supporting indexing and retrieval by time version and spatial range. In terms of visualization, this module provides various visualization functions such as 3D rendering of the canopy height model, binding of feature importance bar charts, generation of cross-validation scatter plots, and color-gradient rendering of volume spatial distribution maps, enabling forestry managers to intuitively understand the model operation process and the spatial distribution characteristics of the prediction results. In one embodiment of the present invention, the system can be deployed on a workstation equipped with a graphics processing unit accelerator card, wherein the point cloud preprocessing and random forest model training processes support parallel computing to improve processing efficiency, for an area of ​​approximately 1500 hm². 2 In typical forest areas, the entire processing time can be controlled within 4 to 6 hours.

[0051] To verify the technical effectiveness of the method of this invention, an experiment was conducted in a subtropical natural secondary mixed forest area. The experimental area covered approximately 1500 hectares. 2 The elevation range is 50~350m ， The average annual precipitation is approximately 1200 mm, belonging to a typical subtropical monsoon climate zone. The forest types include three main types: pure stands of Masson pine, mixed forests of Quercus acutissima and Liquidambar formosana, and mixed coniferous and broad-leaved forests. The forest age ranges from 10 to 60 years, with canopy closure ranging from 0.4 to 0.95. A total of 80 30m × 30m ground plots were established, including 25 coniferous forest plots, 30 broad-leaved forest plots, and 25 mixed forest plots. The average tree density within each plot was 620 trees / hm². 2 The average diameter at breast height (DBH) is 14.8 cm, the average tree height is 12.6 m, and the measured volume ranges from 45 to 285 m³. 3 / hm 2 The average volume is 147m³. 3 / hm 2 .

[0052] Point cloud data was acquired using a DJIM 300RTK UAV equipped with a Zenmuse L1 LiDAR module. The flight altitude was 120m, the flight speed was 4m / s, the flight path overlap rate was 35%, and the acquired point cloud density was approximately 15pts / m. 2It supports up to 3 echo recordings. Data acquisition is conducted during the peak vegetation growth period to ensure the integrity of the canopy structure. The multispectral satellite imagery uses the Sentinel-2Level-2A surface reflectance product with spatial resolutions of 10m (near-infrared band B8) and 20m (red-edge band B5). The time interval between acquisition and UAV flight does not exceed 15 days. After radiometric correction and geometric fine registration, the spatial registration error is controlled within 1 pixel.

[0053] Eighty sample plots were randomly divided into a training set (56 plots) and an independent validation set (24 plots) in a 7:3 ratio. A random forest regression model (300 decision trees, with candidate features representing one-third of the total dimensions) was constructed on the training set, and three rounds of feature importance feedback iterations were performed. After three rounds of iterations, 11 features out of the initial 28 dimensions were discarded due to importance scores below the threshold, ultimately retaining 17 valid features for modeling. Ten-fold cross-validation results show that the method of this invention... , , The accuracy on the independent validation set is , , The small difference between the two indicates that the model has good generalization performance and no significant overfitting.

[0054] Further analysis of the model's performance at different accumulation levels: in the low accumulation range (less than 100m³) 3 / hm 2 ), , In the medium-volume section (100~200m) 3 / hm 2 ), , In high-volume sections (greater than 200m) 3 / hm 2 ), , It is worth noting that the model performance in the high-volume segment significantly outperformed the traditional spectral vegetation index model in the same volume segment (the latter...). Only 0.35, (Up to 38.5%), fully verifying the key role of three-dimensional structural parameters in overcoming the problem of spectral saturation.

[0055] To conduct comparative analysis, three sets of control models were constructed: a traditional regression model using only spectral vegetation indices (NDVI, EVI) as independent variables, and... , Random forest models that only use three-dimensional structural parameters (excluding texture features) have... , A random forest model that uses 3D structural parameters and texture features but does not include feature importance feedback. , The comparative results show that the method of this invention, compared with the traditional spectral vegetation index model, An increase of 0.34, It reduced the spectral saturation problem by 15.9 percentage points; after adding texture features... An improvement of 0.04 was achieved, validating the synergistic effect of fusing 3D structure and spectral texture. Texture features supplement the spatial heterogeneity information of the canopy from the horizontal direction, complementing the vertical information provided by the 3D structure parameters, thus achieving a nonlinear gain greater than the sum of its parts. After introducing feature importance feedback iteration... It has been further improved by 0.03. The reduction of 1.8 percentage points confirms the optimization effect of the adaptive screening mechanism. By eliminating redundant features, the dimensionality burden of the model is reduced, allowing the model to focus more on learning the mapping relationship between high-contribution features and accumulation.

[0056] Feature importance analysis showed that the top five features were, in order: the 80th percentile of canopy height, etc. Canopy coverage Standard deviation of canopy height Canopy bulk density profile, 8th segment The uniformity and texture features in the near-infrared band indicate that high percentile canopy height and canopy cover contribute most to the prediction of stock volume, while upper segment information and near-infrared texture uniformity in the canopy volume density profile also make significant contributions. This result aligns with forestry theory, which predicts that stock volume is positively correlated with dominant tree height and stand density, while the upper canopy volume density is closely related to the canopy width distribution of the main contributors to stock volume (dominant and sub-dominant trees). Furthermore, multiple segment characteristics in the canopy volume density profile ( to All of them were included in the final 17-dimensional feature subset, further confirming the important value of the innovative parameter of canopy volume density profile in volume inversion.

[0057] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for estimating and inverting forest stand volume from remote sensing images, characterized in that, Includes the following steps: Step S1, 3D point cloud data acquisition and preprocessing: A drone equipped with a lidar scanner flies along a preset flight path above the target forest stand to acquire 3D point cloud data of the forest stand area; progressive morphological filtering is performed on the 3D point cloud data to separate ground points from non-ground points; a digital elevation model is generated based on the ground point set through spatial interpolation; the ground elevation corresponding to the digital elevation model is subtracted from the elevation values ​​of all filtered point cloud data to complete the point cloud normalization process; and a canopy height model is constructed based on the normalized point cloud. Step S2, Extraction of three-dimensional canopy structure parameters: Based on the canopy height model, the three-dimensional structure parameters of the forest stand area are extracted using a preset sample plot grid as the statistical unit; Step S3, Multispectral satellite image texture feature extraction: Obtain multispectral satellite images corresponding to the forest stand area, select near-infrared band and red edge band, and extract texture features based on gray-level co-occurrence matrix respectively. Texture features include uniformity, contrast and correlation. Step S4, Multi-source feature fusion and adaptive filtering: The three-dimensional structural parameters extracted in step S2 and the texture features extracted in step S3 are concatenated into a comprehensive feature vector. Based on the feature importance score, each feature in the comprehensive feature vector is sorted, and features with an importance score lower than a preset threshold are removed to obtain a filtered feature subset. Step S5, Random Forest Regression Modeling and Spatial Inversion of Stock Volume: Using the measured stock volume of ground sample plots as the dependent variable and the selected feature subset as the independent variable, a random forest regression model is constructed; the generalization performance of the random forest regression model is evaluated through cross-validation, and the validated random forest regression model is applied to the whole forest area image to predict the stock volume pixel by pixel, and the spatial distribution map of stock volume is generated by rendering in a color gradient manner.

2. The method for estimating and inverting forest stand volume from remote sensing images according to claim 1, characterized in that, In step S1, the UAV flies at an altitude of 100-150m, and the point density of the acquired 3D point cloud data is not less than 10pts / m. 2 .

3. The method for estimating and inverting forest stand volume from remote sensing images according to claim 1, characterized in that, In step S1, progressive morphological filtering uses progressively increasing structuring element sizes to perform opening operations on the point cloud. The initial radius of the structuring element is 0.5m, the maximum radius is 10m, the increment step size is 0.5m, and a height difference threshold is set to distinguish between ground points and vegetation points.

4. The method for estimating and inverting forest stand volume from remote sensing images according to claim 1, characterized in that, In step S2, the three-dimensional structural parameters include the average canopy height, the standard deviation of the canopy height, nine percentiles of the canopy height, canopy coverage, and canopy volume density profile. Among them, the canopy coverage is the proportion of echo points higher than a preset height threshold to the total number of echo points. The canopy volume density profile is the vertical distribution vector formed by dividing the canopy height range into N equal segments and counting the echo point density in each segment. The canopy height percentiles include nine height values ​​from the 10th percentile to the 90th percentile at intervals of 10. The preset height threshold for canopy coverage is 2m.

5. The method for estimating and inverting forest stand volume from remote sensing images according to claim 1, characterized in that, In step S2, the method for constructing the canopy volume density profile is as follows: the maximum canopy height represented by the canopy height model is divided into 10 segments, and the proportion of the echo points of the normalized point cloud in each segment to the total number of echo points is calculated to form a 10-dimensional vertical density distribution vector. This vertical density distribution vector is used to characterize the vertical structure features of the forest stand.

6. The method for estimating and inverting forest stand volume from remote sensing images according to claim 1, characterized in that, In step S3, the calculation window size of the gray-level co-occurrence matrix is ​​3×3 pixels, the step size is 1 pixel, and the calculation directions include four directions: 0°, 45°, 90° and 135°. The texture feature is taken as the average of the four directions as the final feature value.

7. The method for estimating and inverting forest stand volume from remote sensing images according to claim 1, characterized in that, In step S4, the feature importance score is calculated based on the precision reduction method of out-of-bag data permutation inside the random forest, and the preset threshold is 50% of the average importance score of all features.

8. The method for estimating and inverting forest stand volume from remote sensing images according to claim 1, characterized in that, In step S5, the feature importance ranking output by the random forest regression model is fed back to step S4 to update the feature importance score in order to achieve iterative optimization of feature selection. The number of decision trees in the random forest regression model is 200 to 500. When splitting a node, each decision tree randomly selects one-third of the total number of candidate features, and the minimum number of leaf node samples is 3 to 5.

9. The method for estimating and inverting forest stand volume from remote sensing images according to claim 8, characterized in that, In step S5, the cross-validation adopts ten-fold cross-validation, and the coefficient of determination, root mean square error, and relative root mean square error are used as evaluation indicators of model accuracy.

10. A forest stand volume remote sensing image estimation and inversion system, used to implement the forest stand volume remote sensing image estimation and inversion method according to any one of claims 1-9, characterized in that, include: The point cloud acquisition and preprocessing module is used to acquire three-dimensional point cloud data of the target forest stand by using a drone equipped with a lidar scanner, perform progressive morphological filtering on the three-dimensional point cloud data to separate ground points, generate a digital elevation model based on the ground points, complete the point cloud normalization process, and construct a canopy height model. The three-dimensional structural parameter extraction module is used to extract the three-dimensional structural parameters of the forest stand area based on the canopy height model. The three-dimensional structural parameters include the average canopy height, the standard deviation of the canopy height, nine percentiles of the canopy height, the canopy coverage, and the canopy volume density profile. The texture feature extraction module is used to acquire multispectral satellite images and extract uniformity, contrast and correlation texture features for the near-infrared band and red edge band based on the gray-level co-occurrence matrix, respectively. The feature fusion and filtering module is used to concatenate 3D structural parameters and texture features into a comprehensive feature vector, and to perform adaptive filtering based on feature importance scores to obtain a filtered feature subset. The modeling and inversion module is used to construct a random forest regression model with the measured volume of ground sample plots as the dependent variable and the selected feature subset as the independent variable. After cross-validation, the model predicts the volume of the entire forest area pixel by pixel and generates a spatial distribution map of the volume. The feature importance ranking output by the modeling and inversion module is fed back to the feature fusion and filtering module to achieve iterative optimization.

Images