A remote-sensing-based method for estimating forest above-ground biomass and its carbon sink value

Abstract

The application discloses a kind of forest above-ground biomass and its carbon sink value estimation method based on remote sensing, it is related to forest above-ground biomass measurement technical field, its technical solution key points are: obtaining ground-based and unmanned aerial LiDAR data, extracting single tree height, crown area and diameter at breast height;Through feature level point cloud fusion and diameter at breast height optimization model, combined with allometric growth model, the biomass of single tree is calculated and the total biomass of sample plot is summarized;Point cloud is voxelized, aggregated according to satellite image resolution, and multi-scale biomass distribution map is generated;Finally, the spectral and texture features of satellite image are extracted, and a remote sensing estimation model is constructed with the biomass of sample plot as the dependent variable to perform regional inversion. This method integrates multi-source data, solves the problem of high-precision collection and scale conversion of sample data, and realizes large-scale high-precision forest biomass and carbon sink value estimation.

Description

Technical Field

[0001] This invention relates to the field of forest aboveground biomass measurement technology, and more specifically, to a remote sensing-based method for estimating forest aboveground biomass and its carbon sink value. Background Technology

[0002] Accurate estimation of forest aboveground biomass (AGB) and its carbon sink value is a key scientific issue for global carbon cycle research, climate change response, and sustainable forest management.

[0003] With the rapid development of remote sensing technology, especially the fusion of LiDAR and multispectral remote sensing, revolutionary methods have been provided for forest parameter extraction. LiDAR acquires three-dimensional point clouds of ground features by actively emitting laser pulses, directly retrieving forest vertical structure parameters and overcoming the saturation problem of traditional optical remote sensing in biomass estimation. Specifically:

[0004] Airborne lidar (ALS): Typically mounted on UAVs or aircraft platforms, it can quickly acquire large-scale forest canopy height models (CHMs), suitable for regional-scale tree height extraction and canopy structure analysis.

[0005] Ground-based lidar (TLS): Deployed on the ground, it acquires fine three-dimensional point clouds of individual trees in the sample plot through multi-station scanning. It has sub-centimeter accuracy, especially in diameter at breast height (DBH) measurement. However, it is limited by line-of-sight obstruction and is prone to underestimation in tree height measurement.

[0006] However, single technologies often have limitations: ALS has insufficient penetration under dense canopies, affecting the acquisition of understory structures; TLS is limited by scanning range and viewing angle, making it difficult to fully capture the top of the canopy. Therefore, multi-platform collaboration between space, air, and ground has become a current research frontier, achieving multi-scale, high-precision forest parameter extraction by integrating satellite remote sensing (providing spectral and texture information), ALS (providing canopy height), and TLS (providing diameter at breast height and fine structure).

[0007] At the data processing level, point cloud processing algorithms (such as filtering, segmentation, and feature extraction) and machine learning models (such as random forests and multivariate regression) are widely used to invert biomass from point clouds and images. Meanwhile, accuracy verification systems are becoming increasingly sophisticated, including methods such as cross-validation, error propagation analysis, and comparison with global products, to ensure the reliability and comparability of the estimation results.

[0008] In light of this, this invention proposes a combined "space-air-ground" technical approach integrating GF-2 satellite imagery, UAV ALS, and ground-based TLS, aiming to address the inadequacy of single data sources and improve the accuracy and efficiency of forest biomass and carbon sink estimation. This research, along with user file sharing of in-depth explorations into the application of TLS technology, jointly promotes the standardization and practical application of lidar in quantitative remote sensing of forestry. Summary of the Invention

[0009] The purpose of this invention is to provide a remote sensing-based method for estimating forest aboveground biomass and its carbon sequestration value, thereby solving the aforementioned problems.

[0010] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0011] The first aspect of this invention provides a method for estimating forest aboveground biomass and its carbon sink value based on remote sensing, comprising the following steps:

[0012] Preprocessed ground-based LiDAR data and UAV-borne LiDAR data of the sample plots were acquired and feature extracted to obtain the individual tree height and crown area of ​​the sample plots and the individual tree diameter at breast height (DBH) of the sample plots. The ground-based and UAV-borne LiDAR data were fused at the feature level to obtain individual tree parameters. The individual tree height and crown area of ​​the sample plots were input into the DBH optimization model to obtain the DBH of the sample plots. An allometric growth model was constructed, and the DBH and tree height of the sample plots were input to obtain the individual tree biomass. The individual tree biomass was then summarized to obtain the total biomass of the sample plots. The UAV-borne LiDAR data was pixelated, and the biomass of a single point cloud voxel size was calculated based on the individual tree parameters. Upscaling aggregation was performed, and the biomass of each pixel was calculated based on the total biomass of the sample plots to obtain the aboveground biomass distribution map of the sample plots. Satellite imagery was acquired, and spectral and texture features were extracted. Using spectral and texture features as independent variables and the aboveground biomass distribution map of the sample plots as the dependent variable, a biomass remote sensing estimation model was constructed and inverted to estimate the forest aboveground biomass.

[0013] A second aspect of the present invention also provides an apparatus / device / system for estimating forest aboveground biomass and its carbon sink value based on remote sensing, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0014] A third aspect of the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0015] A fourth aspect of the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0016] In summary, the present invention has the following beneficial effects:

[0017] This invention provides a high-precision forest biomass estimation method integrating "space-air-ground" technology. Using UAV remote sensing technology as a link, it enables rapid and refined collection of sample plot data, integrates ground, UAV, and satellite data, and constructs a high-precision biomass remote sensing estimation model to achieve large-area estimation of biomass and carbon sink value.

[0018] This invention proposes a feature-level point cloud fusion method for extracting individual tree parameters at the plot scale. This method fully leverages the advantages of LiDAR data from different platforms to solve the problem of high-precision and rapid acquisition of plot data, providing training samples for establishing biomass remote sensing estimation models. It can also provide validation samples that meet the requirements of different spatial scales for large-scale satellite remote sensing applications.

[0019] This invention voxelizes point cloud data at the plot scale, allowing for arbitrary aggregation based on satellite image resolution to obtain biomass at satellite image resolution. This enables forest biomass estimation at different satellite scales within a region. This method ensures both the sample size and accuracy of measured data while achieving forest biomass estimation at variable satellite scales. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for estimating forest aboveground biomass and its carbon sink value based on remote sensing, as described in Embodiment 1 of the present invention.

[0021] Figure 2 This is a diagram showing the relationship between quadrats, sample plots, and regions in Embodiment 1 of the present invention;

[0022] Figure 3 This is a point cloud data map of the airborne radar in Embodiment 1 of the present invention;

[0023] Figure 4 This is a flowchart of the biomass model construction process in Embodiment 1 of the present invention;

[0024] Figure 5 This is a diagram showing the correspondence between satellite pixels and voxels in Embodiment 1 of the present invention;

[0025] Figure 6 This is a forest scene diagram from Embodiment 3 of the present invention;

[0026] Figure 7 This is a simulated airborne point cloud diagram from Embodiment 3 of the present invention;

[0027] Figure 8 This is a ground point cloud simulation diagram in Embodiment 3 of the present invention;

[0028] Figure 9 This is a ground point cloud simulation diagram in Embodiment 3 of the present invention;

[0029] Figure 10 This is the accuracy verification of pixel AGB weight allocation in Embodiment 3 of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1:

[0032] A remote sensing-based method for estimating forest aboveground biomass and its carbon sequestration value, such as... Figure 1 As shown, it includes the following steps:

[0033] S100. Acquire the pre-processed ground-based LiDAR data of the sample plot and the UAV-borne LiDAR data of the sample plot, and perform feature extraction to obtain the individual tree height and crown area of ​​the sample plot and the individual tree diameter at breast height of the sample plot.

[0034] S200. Feature-level fusion of ground-based and UAV-borne LiDAR data is performed to obtain individual tree parameters. The individual tree height and crown area of ​​the sample plot are input into the diameter at breast height (DBH) optimization model to obtain the DBH of the sample plot.

[0035] S300. Construct an allometric growth model, input the diameter at breast height (DBH) and tree height of the sample plot, obtain the biomass of individual trees, summarize the biomass of individual trees, and obtain the total biomass of the sample plot;

[0036] S400. The UAV-borne LiDAR data is processed into voxels, and the biomass of a single point cloud voxel is calculated by combining individual tree parameters. The data is then aggregated at upscale and combined with the total biomass of the sample plot to calculate the biomass of each pixel, thereby obtaining a map of the aboveground biomass distribution of the sample plot.

[0037] S500. Acquire satellite imagery, extract spectral and texture features, use spectral and texture features as independent variables and the aboveground biomass distribution map of the sample plot as the dependent variable, construct a biomass remote sensing estimation model, perform inversion, and estimate the aboveground biomass of the forest.

[0038] S600. Based on forest aboveground biomass, determine carbon storage and calculate carbon sink value.

[0039] In step S100, the preprocessing includes: ground point detection, point cloud denoising, normalization, and single-tree segmentation. Specifically, step S100 includes the following steps:

[0040] S110. Collect basic data and remote sensing images related to the forest in the project area (target).

[0041] S120. Use high-resolution (GF-2) optical remote sensing images with sampling time and region that meet the work requirements for coarse classification. Based on the results of this classification, select a certain number of sample plots for field reconnaissance.

[0042] For the reconnaissance of forest sample plots in the project area, a field reconnaissance route map was drawn, a total of 566 points were surveyed, and points that could be used as candidate sample plots were photographed and tree species information was recorded.

[0043] S130. Remote sensing image classification: Based on the results of field surveys, the interpretation markers are adjusted to make accurate classifications of remote sensing images (broadleaf forest A, broadleaf forest B, broadleaf forest C, broadleaf forest D, and coniferous forest).

[0044] S140. Based on the classification and field survey results, select the sample plot range for each category and obtain ground-based LiDAR, UAV-borne LiDAR and visible light images of the sample plot area;

[0045] Field survey results indicate that the main tree species in the project area include poplar, ash, catalpa, locust, ailanthus, golden elm, and Chinese pine. Specific growth characteristics are shown in Table 1.

[0046] Table 1. Description of characteristics of each tree species

[0047]

[0048] Referring to the forest types in the national standard "Classification and Code of Forest Types for Forestry Resources," and based on the land cover status and application needs of the study area, combined with the characteristics of remote sensing imagery and the ability to accurately identify forest types, this invention comprehensively considers the scientific validity, systematicity, and hierarchy of the classifications. The main forest land types are divided into broadleaf forests and coniferous forests. As shown in Table 3, each category of broadleaf forest has unique spectral reflectance characteristics, and significant spectral differences can serve as a basis for identifying and classifying forest resources. Therefore, broadleaf forests can be divided into four categories: Broadleaf Forest A, Broadleaf Forest B, Broadleaf Forest C, and Broadleaf Forest D. Field surveys revealed that coniferous forests are mainly composed of Pinus tabuliformis. Non-forest land is not the focus of this invention's investigation; therefore, non-forest land types are not further subdivided. The specific classification system and tree species composition are shown in Table 2.

[0049] Table 2 Types of Broadleaf Forests

[0050]

[0051] Because coniferous forests cover relatively small areas and do not appear in contiguous patches, they are difficult to classify and extract from satellite imagery. Therefore, this invention combines field surveys with manual delineation of the coniferous forest boundaries on satellite imagery. The aboveground biomass of different forest types in the project area can then be estimated based on the classification results.

[0052] Selecting suitable data from a wide variety of remote sensing images is fundamental to accurately estimating forest aboveground biomass in the project area. The data used in this invention is GF-2 data from July 2023. GF-2 data preprocessing mainly includes radiometric calibration, atmospheric correction, orthorectification, and image fusion. Since panchromatic images cannot be atmospherically corrected, quantitative remote sensing research generally only uses multispectral data for atmospheric correction, followed by orthorectification before fusing the panchromatic and multispectral data.

[0053] Object-oriented classification

[0054] (1) Forest land and non-forest land

[0055] Threshold classification was performed using NDVI (Normalized Difference Vegetation Index) to separate forest land from non-forest land.

[0056] (2) Broadleaf forests and coniferous forests

[0057] Because coniferous forests cover a small area and do not appear in large patches, they are difficult to classify and extract from satellite imagery. Furthermore, the fused GF-2 satellite imagery possesses both high spatial resolution and spectral texture information, facilitating manual delineation. Therefore, based on field surveys, the location of coniferous forests was manually delineated on the fused GF-2 satellite imagery, with the remaining vegetation classified as broadleaf forest.

[0058] (3) Broadleaf forest AD category

[0059] By visually interpreting the samples, select an appropriate number of AD type samples based on obvious differences, ensuring that the samples are evenly distributed throughout the region.

[0060] Using spectral features (spectral mean of each band and vegetation index) and texture features (GLCM dissimilarity, GLCM entropy, GLCM contrast, etc.), various classifiers (SVM, RF and decision tree) are selected for classification. The optimal combination is obtained through various combinations of classification features and classifiers.

[0061] The area of ​​broad-leaved forest type A (poplar) is approximately 113 km²; the area of ​​broad-leaved forest type B (ash and catalpa) is approximately 53 km². 2 The area of ​​broad-leaved forest category C (locust, ailanthus, and goldenrain trees) is approximately 124 km². 2 Broadleaf forest type D (golden elm) has a relatively small distribution area (approximately 1 km²). 2In addition, the area of ​​coniferous forests (Pinus tabuliformis) is relatively small (approximately 2 km²). 2 And they are relatively concentrated.

[0062] In step S200, sample plot selection: Based on the classification and field reconnaissance results, select the sample plot range for each category, and acquire ground-based LiDAR, UAV-borne LiDAR, and visible light images of the sample plot area;

[0063] The content of the sample plot survey includes: tree species composition, diameter at breast height (DBH), tree height, growth status, main vegetation type and coverage in the understory, geographical location, etc.

[0064] A forest stand refers to a forest area with largely similar internal characteristics (such as tree species composition, canopy layer, age, canopy closure, origin, status class, or timber yield) that is clearly distinct from the surrounding forests. Forest stands are often used as the basis for determining forest management measures; different forest stands require different management measures. A sample plot has the largest area, while a quadrat has a smaller area, generally referring to a smaller space within a sample plot.

[0065] In this invention, multi-scale refers to "quadrat-plot-region", and the relationship between the three is as follows: Figure 2 As shown.

[0066] This invention employs a method combining remote sensing image interpretation and field surveys to select and establish forest plots in the project area. Based on the analysis of remote sensing images and the results of field surveys, the broadleaf forests in the study area are mainly divided into four categories: broadleaf forest A (15 plots), broadleaf forest B (6 plots), broadleaf forest C (7 plots), and broadleaf forest D (2 plots), with a total of 30 plots established in each category.

[0067] The details of the sample plots and quadrats are shown in Table 3.

[0068] Table 3 Distribution of sample plots and quadrats

[0069]

[0070] Overall, the sample plots in each study area were widely distributed, covering different types of broad-leaved forests, providing diverse basic data for further research.

[0071] Visible light data acquisition: The entire aerial photography operation process includes: (1) handling airspace application and flight approval procedures, and preparing relevant materials, equipment and personnel; (2) completing the aerial photography technical design and approval filing; (3) organizing UAV aerial photography according to design requirements, and conducting formal operations after test flights and inspections; (4) implementing a real-time quality inspection system, processing and feeding back image data in a timely manner, and promptly re-flying if problems are found; (5) organizing aerial photography results and submitting them for acceptance, ensuring that the results are complete and intact. A total of 30 UAV visible light orthophoto images were collected, covering an area of ​​1.77 km². 2 .

[0072] Airborne LiDAR Data Acquisition: Data acquisition was conducted using a DJI M300 RTK (a multi-rotor drone equipped with a DJI L1 camera). The DJI L1 sensor includes an RGB (red, green, blue) sensor and a LiDAR sensor, capable of simultaneously acquiring RGB and LiDAR data. Airborne LiDAR point cloud data was collected from 30 sample plots, covering an area of ​​1.77 km². 2 .

[0073] After data collection, a quality check is required to ensure complete data coverage, compliance with flight path design requirements, and preliminary checks such as image clarity. If problems are found, a supplementary flight is conducted. During the supplementary flight, the flight path design is optimized based on the cause to ensure data quality meets requirements. Specific airborne radar point cloud data is shown in Figure 3.

[0074] Ground-based LiDAR data acquisition: TLS (Ground-based Laser Scanning) data was acquired using a Stonex X300 laser scanner. The measurement accuracy was 4 mm at 50 m, and the measurement distance was 300 m. Data was acquired in fine mode, with a single station set up at each sample plot, and the average point cloud density was 2281.92 pts·m. -2 The geographic coordinates of the ground-based laser point cloud (TLS) base station and the target sphere were measured using a Trimble R8, and the TLS data was oriented using a single-station orientation method.

[0075] Remote sensing feature extraction: This invention extracts texture features from remote sensing images based on the gray-level co-occurrence matrix (GLCM). The eight extracted texture features are shown in Table 4.

[0076] Table 4 Texture Features

[0077]

[0078] The spectral features of this invention are extracted based on GF-2 multispectral images, and the texture features are obtained by resampling GF-2 panchromatic images to 4m resolution. Therefore, this invention is processed entirely based on 4m resolution images.

[0079] Visible light data: Visual interpretation is achieved by acquiring visible light data from UAVs. The UAV aerial image processing workflow includes raw image data collection and camera calibration, correction of image distortion and color differences to ensure image consistency, improvement of image matching stability through feature matching and grayscale matching combined with the SIFT operator, aerial triangulation using the regional network method to obtain high-precision exterior orientation elements, and generation of a digital surface model (DSM). During DSM generation, object-space matching is used to improve accuracy while reducing manual editing workload, and fully automated light and color homogenization solves image stitching problems. Finally, the original image is corrected using the DSM, and orthorectified images are generated through resampling, providing reliable support for high-quality aerial survey results.

[0080] By processing airborne lidar point cloud data, the height of individual trees can be accurately extracted. Since the biomass of individual trees is calculated using the allometric growth equation, and the parameters of the allometric growth equation are determined based on the tree height and diameter at breast height (DBH) data, accurate acquisition of tree height is of great significance for biomass estimation.

[0081] The acquired RGB (Red, Green, Blue) and LiDAR data were systematically processed. RGB data was used to generate digital orthophotos (DOMs) for subsequent visual interpretation. For airborne laser scanning (ALS) point cloud data, the processing included denoising, ground point identification, point cloud normalization, individual tree segmentation, and tree feature extraction. First, an algorithm based on smooth surface growth was used to detect isolated points. Then, an irregular triangular mesh progressive densification filtering method was used to filter the point cloud and identify its location, generating a canopy height model (CHM). The normalized height of the CHM was used to identify tall vegetation on the ground. Finally, a watershed segmentation method for linear entity extraction was used to process the point cloud to extract parameter information for each tree.

[0082] Based on the PointCloudAutomata software platform, LiDAR point cloud data is filtered and classified. This invention uses the Kriging interpolation method to process the point cloud data. The ground points are interpolated using Kriging to obtain a digital elevation model (DEM), retaining only the first echo points. Similarly, Kriging interpolation is applied to obtain a digital surface model (DSM). Subtracting the DEM from the DSM yields the tree canopy height model (normalized digital surface model).

[0083] Because LiDAR point clouds contain gaps in the branches and leaves that penetrate the canopy, the laser return points that directly hit the ground will have some null values ​​in the canopy cloud model (CHM). Extracting individual tree structure parameters solely from the CHM can lead to misclassification and omission errors. Therefore, to reduce these errors, this invention creates a closed canopy region in the image and performs a morphological "opening" operation on the CHM surface. In the CHM, each grid cell value is replaced with the maximum height of the laser point cloud in its neighborhood, resulting in the canopy maxima model (CMM).

[0084] The coarseness of the CMM (Content Mirror Model) can lead to misidentification of tree crown vertices. Gaussian filtering is typically used to process the CMM. Experiments have shown that a smaller standard deviation of the Gaussian template results in a brighter image with minimal impact on subsequent crown extraction. Therefore, this invention proposes using a 5×5 Gaussian matrix to smooth the CMM, removing noise and obtaining a GCMM (Gaussfilter CMM) image. Using GCMM to extract individual tree crown vertices can effectively reduce false detections of crown vertices.

[0085] The height of a single tree is obtained by detecting the position of its crown apex. On GCMM images, a local maximum method with a variable-size dynamic window is used to detect the crown apex position. The point with the local maximum height in the dynamic window is taken as the crown apex, and the pixel value at the crown apex is taken as the tree height. The size of the dynamic window is determined by the lower limit of the 95% (α=0.05) prediction interval of the regression analysis of the measured crown width and tree height in the sample plot. Some studies have indicated that reducing α can further reduce the missed detection error, but a small α may produce a large misclassification error. This invention aims to find an α value that can effectively balance the misclassification and missed detection errors through an iterative method.

[0086] Traditional watershed segmentation can lead to oversegmentation. Therefore, this invention aims to improve the watershed algorithm by adopting a marker-controlled watershed segmentation algorithm. This algorithm finds internal and external markers on the original image to guide the segmentation process, thereby preventing oversegmentation.

[0087] To simulate the immersion process, the GCMM is first processed to calculate the complement of the two water collection basins. Assuming each basin has a hole at its minimum value, the basin will fill as water is continuously added. This algorithm can be considered as automatically constructing a dam between two adjacent water collection basins, which acts as a water divider and is used to separate individual trees.

[0088] The local maximum algorithm is used to mark the tree crown vertices, and these vertices are used as control symbols to mark and control the watershed segmentation of GCMM. High spatial resolution remote sensing imagery is used as auxiliary data to obtain the tree crown boundary and thus determine the size of the individual tree crown.

[0089] Ground-based LiDAR data

[0090] Ground-based lidar point cloud data can provide relatively accurate diameter at breast height (DBH) parameters for individual trees. Combining DBH with tree height obtained from airborne point cloud data allows for the accurate determination of individual tree biomass through allometric growth equations, laying the groundwork for subsequent large-scale forest biomass inversion at regional scales.

[0091] The main processing flow of ground-based lidar (TLS) includes location detection, point cloud normalization, tall vegetation identification, individual tree segmentation, and feature parameter extraction. The diameter at breast height (DBH) is calculated using an algorithm based on a distance measurement function, determined at a height of 1.3m above the ground (trees with a DBH less than 5cm are not considered).

[0092] Parameter extraction at the quadrat-plot scale

[0093] The total aboveground biomass of the sample plot was obtained by constructing an allometric growth relationship using individual tree height and diameter at breast height (DBH) in the standard specifications. First, tree height data obtained from UAV laser point cloud and DBH data obtained from ground-based lidar were fused to obtain high-precision individual tree parameters. Then, a regression relationship between UAV tree height and ground-based DBH was established to realize the conversion of DBH between quadrat and sample plot, obtaining large-scale measured data. Finally, the total biomass of the sample plot was calculated based on the biomass of individual trees within the sample plot, thereby realizing parameter extraction at the sample plot scale.

[0094] In step S200, the feature-level fusion of ground-based and UAV-borne LiDAR data is performed by using the tree trunk position extracted from the ground-based and UAV-borne LiDAR data to associate other features.

[0095] Specifically, it includes the following steps:

[0096] Airborne LiDAR point cloud data can provide high-precision parameters for individual tree height, while ground-based LiDAR point cloud data can provide high-precision parameters for individual tree diameter at breast height (DBH). To obtain accurate biomass data for sample plots based on tree height and DBH data, it is necessary to fuse airborne and ground-based LiDAR point cloud data.

[0097] This invention utilizes TLS and ALS for feature-level fusion to achieve individual tree measurement. Furthermore, it can use ALS data to estimate the accurate diameter at breast height (DBH) in areas without TLS coverage, thus obtaining a wider range of accurate DBH, crown width, and other individual tree characteristics.

[0098] By processing the acquired airborne and ground-based point cloud data and using a fusion scheme, large-scale individual tree features can be obtained. These features can then be used as sample tree information to estimate parameters such as diameter at breast height (DBH) and number of trees for each tree species in the entire study area.

[0099] This method addresses this issue by employing DGNSS technology during the data acquisition phase. Specifically, high-precision DGNSS RTK / PPK measurements are used to assign accurate geographic coordinates to both ALS and TLS data, meaning that the fusion of the two point cloud data is achieved using a common geographic reference. In this invention, during ALS data acquisition, georegistration is directly performed using data provided by the UAV-borne IMU and DGNSS system, employing the China Geodetic Coordinate System 2000 (CGCS2000) coordinate system and the Gauss-Krüger 3-dimensional projection method, with the central meridian at 117°E. After processing, the horizontal accuracy of the ALS point cloud is 10cm, and the vertical accuracy is 5cm. During TLS data acquisition, a "base station location + single target sphere location" approach is used, i.e., the positions of the base station and the target sphere are measured separately, and the point cloud orientation parameters are solved through 3D coordinate transformation and least squares adjustment to achieve georegistration of the TLS point cloud data. The same projection and coordinate system as ALS are used in the processing, resulting in a horizontal accuracy of 8mm and a vertical accuracy of 15mm for the processed point cloud. ALS data processing is implemented in DJI Terra v3.5, while TLS data processing is implemented in Stonex Siscan v3.0.

[0100] Leveraging the fact that the tree trunk positions extracted from the two point cloud datasets are very close, all other features extracted from the two point clouds are correlated to achieve feature-level fusion of the point cloud data. This ultimately yields individual tree features such as tree height, diameter at breast height (DBH), and crown width. The correlation rule, as shown in the formula, identifies trees as belonging to the same tree based on trunk positions extracted from both data sources not exceeding 2 meters.

[0101] ;

[0102] Given two feature datasets A and B, and a radius r, the spatial feature connection (A,B) returns a pair of spatial objects (a,b), where a belongs to A and b belongs to B. L(a,b) is the distance between all pairs of points within r. If the distance between the tree trunk locations obtained from ground-based and UAV-borne LiDAR data exceeds r, they are considered not to be the same tree.

[0103] This invention combines tree height data obtained from UAV lidar within the sample plot with diameter at breast height (DBH) data obtained from ground-based lidar within the sample plot to obtain large-scale DBH data. To improve the efficiency and accuracy of data acquisition, this invention employs ground-based or mobile lidar technology to extract DBH parameters from sample plots within the forest.

[0104] Meanwhile, ground-based and mobile lidar technologies precisely measure the three-dimensional structure of trees by emitting lasers and receiving reflected signals. This technology is characterized by high precision and efficiency, enabling the acquisition of large amounts of diameter at breast height (DBH) data in a short time. However, despite the high-precision measurement data provided by ground-based or mobile lidar, its limited coverage makes this data insufficient for remote sensing data modeling.

[0105] To compensate for this deficiency, the application of drone technology has become particularly important. Drones can fly over large forest areas in a short time, acquiring several times or even tens of times more data than traditional quadrat measurements. High-resolution cameras and sensors equipped on drones can capture detailed canopy images and obtain tree height data through algorithmic analysis. However, drones have certain limitations in diameter at breast height (DBH) extraction; using drone data alone is insufficient to obtain accurate DBH information. Therefore, combining ground-based / mobile lidar with drone technology has become an effective solution. By combining tree height data acquired by drones with DBH data acquired by ground-based / mobile lidar, accurate DBH extraction can be achieved at the quadrat scale. This integration of multi-source data not only improves the accuracy of DBH measurement but also enhances the reliability and applicability of the data model.

[0106] This invention utilizes features extracted from the ALS point cloud of quadrats, namely tree height and crown width, to establish a tree height-diameter-at-breast-width (DBH) model according to the following formula, used to predict the DBH of individual trees in a quadrat plot. This achieves scale transformation from quadrat to plot. Specifically, the DBH of the plot is obtained using the fused feature dataset obtained in the previous section.

[0107] The breast diameter optimization model is as follows:

[0108] ;

[0109] Among them, DBH is the diameter at breast height of a single tree measured by ground-based LiDAR; H and CA are the tree height and crown area of ​​a single tree obtained by airborne LiDAR; a and b are undetermined coefficients, which are determined by using quadrat survey data and the least squares method.

[0110] In step S300, according to the technical regulations for forestry carbon sequestration measurement, the aboveground biomass of each category (Saussurea, Oleaceae, Fabaceae, Betulaceae, and coniferous forests) can be obtained from allometric growth models constructed using tree height and diameter at breast height. The allometric growth models corresponding to each category are shown in Table 5. Finally, the total biomass of the sample plot is obtained by summing the biomass of all individual trees within the plot.

[0111] Table 5 Classification System

[0112]

[0113] Where W is the aboveground biomass of a single tree (AGB) (kg), D is the diameter at breast height of a single tree (cm), and H is the tree height (m).

[0114] In step S400, the point cloud data acquired by airborne LiDAR consists of a large number of discrete points, a characteristic significantly different from optical remote sensing image data, which can represent the spectral characteristics of continuously distributed target surfaces. Since each point in the LiDAR data has independent three-dimensional coordinates and their spatial density distribution is uneven and discrete, specific processing is required to extract the three-dimensional structural features of the sample plot area. This study uses segmented single-tree point clouds within the sample plot to extract the morphological and structural features contained in the point cloud dataset at the single-tree scale. This part utilizes voxelization to process the single-tree point clouds, obtaining the number of voxels for each point cloud voxel size within the sample plot based on a voxel grid, and calculating the biomass of a single point cloud voxel size. Then, by upscaling and aggregating the point cloud voxel size with the satellite image pixel ratio, the biomass of each pixel is calculated, obtaining the aboveground biomass distribution map of the sample plot. Finally, a regional-scale biomass estimation model is constructed using the correspondence between the aboveground biomass distribution map of the sample plot and satellite remote sensing indices. The specific process is shown in Figure 4.

[0115] S410. Point cloud voxelization

[0116] The airborne LiDAR point cloud of the sample area was voxelized using the stereopixel method, and empty voxels were removed. The voxel resolution (side length) should be better than the spatial resolution of the satellite imagery. The higher the voxel resolution setting, the finer the voxelization result, but the longer the processing time. The point cloud voxel size used in this invention is 1m.

[0117] The voxel resolution should generally be smaller than the crown diameter, and is usually determined by the crown diameter of trees in the plot and the spatial resolution of satellite imagery. It should also satisfy the condition that the quotient of the spatial resolution of satellite imagery and the voxel resolution is an integer, so as to facilitate upscaling aggregation.

[0118] S420. Voxel-scale biomass calculation

[0119] The biomass of each pixel at voxel resolution is calculated using the following formula to obtain the aboveground biomass Map_v of the plot at voxel resolution:

[0120] The biomass calculation method for the size of a single point cloud voxel is as follows:

[0121] ;

[0122] in, It is the first The biomass of each pixel, B is the total biomass of the plot, and T is the total number of elements in the plot. It represents the number of all voxels in the vertical direction of that pixel, as shown in Figure 5.

[0123] Map_v is upscaled and aggregated according to the following formula to obtain Map_s, a plot aboveground biomass distribution map with satellite image resolution (aggregated to 4m*4m in this invention). Since the spectral information used in this invention comes from GF-2 multispectral imagery with a spatial resolution of 4m, the point cloud is upscaled and aggregated to 4m*4m.

[0124] The formula for calculating upscaling aggregation is:

[0125] ;

[0126] ;

[0127] in, It is the first Biomass per satellite image pixel For satellite image resolution, For voxel resolution, It is the first one to participate in the aggregation within the spatial range corresponding to the satellite pixel. One pixel.

[0128] In step S500, the training sample set is obtained by utilizing the correspondence between the aboveground biomass distribution map Map_s of the sample plot and the satellite imagery. Considering the complexity of the model, the training sample set is randomly divided into a training set and a validation set in a 7:3 ratio, and the final selected remote sensing indices do not exceed 5.

[0129] The specific process of constructing the biomass estimation model in this invention involves using the acquired remote sensing feature factors (spectral and textural features) as independent variables and the acquired plot biomass as the dependent variable. Based on the results obtained from the training dataset, a multiple linear model is used to construct the biomass remote sensing estimation model. The modeling accuracy for various broadleaf forest categories is shown in Table 6 below.

[0130] Table 6. Statistics on the modeling accuracy of broadleaf forests in the project area

[0131]

[0132] For each type of broadleaf forest in the project area, corresponding models were established, totaling four models. Among them, coniferous forest accounted for a small proportion and was too sparsely planted, and was severely affected by the background (soil), resulting in very low modeling and inversion accuracy. Therefore, the aboveground biomass of coniferous forest in the entire area was calculated manually.

[0133] Table 11 shows the AGB modeling accuracy for the four categories of broadleaf forests in the project area. Category A performed best, with high accuracy and low error in both modeling and validation data, demonstrating good generalization ability. Category B's accuracy was slightly lower than Category A, especially in the validation data. Category C had low accuracy and large error, indicating poor model applicability and generalization ability. Although Category D performed consistently in both modeling and validation data, its high error may be due to sample complexity or data quality. The biomass-independent feature factors selected for biomass independence were all spectral / texture feature factors.

[0134] Based on the classification results, satellite imagery and biomass inversion models for each category were used to calculate the total biomass of each category in the entire region.

[0135] In step S600, based on the forest aboveground biomass calculation results, the forest aboveground biomass is converted into its carbon storage using the Intergovernmental Panel on Climate Change (IPCC) carbon content factor. The calculation formula is as follows:

[0136] Ci = Wi * CFi;

[0137] In the formula, Ci represents the carbon storage of a certain type of tree; Wi represents the biomass of a certain tree species; and CFi represents the carbon content coefficient of a certain tree species. According to the "Guidelines for Carbon Sequestration Measurement and Monitoring of Afforestation Projects", the carbon content coefficient of broad-leaved forests is 0.48, and the carbon content coefficient of coniferous forests is 0.51.

[0138] Finally, based on the closing price of RMB 0.05975 / kg CO2 in the national carbon market on July 12, 2023, published by the national carbon emission trading institution, and the price of RMB 0.063 / kg CO2 on the Hebei Province carbon reduction product value realization service platform, the carbon sink value was calculated. The estimated results of biomass, carbon storage, and carbon sink in the project area are shown in Table 7.

[0139] Table 7 Biomass Statistics

[0140]

[0141] As shown in Table 7, broadleaf forest type A (poplar) has the highest biomass. Therefore, broadleaf forest type A (poplar) also has the highest biomass, carbon storage, and carbon sink value, indicating that it has the greatest potential for carbon sinking. Broadleaf forest types B (ash and catalpa) and C (locust, ailanthus, and goldenrain tree) also make significant contributions to carbon sinking, while broadleaf forest type D (golden elm) and coniferous forests make relatively smaller contributions.

[0142] Example 2:

[0143] The present invention also provides an apparatus / device / system for estimating forest aboveground biomass and its carbon sink value based on remote sensing, comprising a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the above method.

[0144] The present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0145] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0146] Example 3: Accuracy Verification

[0147] The accuracy of biomass estimation was validated from both methodological and estimation result perspectives. First, the effectiveness of the classification method was tested by evaluating the classification accuracy of remote sensing images, and various pixel AGB allocation weighting methods were compared to determine the optimal weighting strategy. Then, the reliability of the models was verified by comparing the accuracy of multiple biomass estimation models from the perspectives of different types and sample plots. Furthermore, the study comprehensively validated the accuracy of the biomass estimation methods and results by analyzing the impact of measured data accuracy on the model, analyzing the accuracy of validation points, comparing the model accuracy at different scales with the estimation results of traditional models, and comparing with global canopy height products.

[0148] The following evaluation metrics were selected: overall classification accuracy, user accuracy, production accuracy, and Kappa coefficient. These metrics play a crucial role in judging the ability of the classification algorithm. The higher the value of these metrics, the higher the classification accuracy and the better the classification effect.

[0149] The accuracy of remote sensing image classification was verified using overall classification accuracy, user accuracy, production accuracy, and Kappa coefficient. Specific classification accuracies are shown in Table 8 below.

[0150] Table 8 Classification Accuracy

[0151]

[0152] Among the various categories, the producer accuracy for broadleaf forest category B is relatively high at 82.08%, meaning that the proportion of correctly classified trees in this category is relatively high; its user accuracy is even higher at 88.34%, indicating that when the model predicts a broadleaf forest category B, the prediction is likely to be accurate. The producer and user accuracy for broadleaf forest categories A and C (locust, ailanthus, and goldenrain trees) are also quite respectable, both around 80%, indicating that these two categories perform relatively stably in the classification process. The producer accuracy for broadleaf forest category D is 80.53%, and the user accuracy is 81.57%, showing acceptable classification performance. The producer accuracy for coniferous forest is 81.25%, while the user accuracy is relatively low at 76.31%.

[0153] Overall, the classification accuracy reached 81.30%, and the Kappa coefficient was 0.76, indicating that the classification results have a certain degree of reliability.

[0154] Comparison of pixel AGB weight allocation methods

[0155] This invention generates individual models of 30 poplar trees based on a radiative transfer model and constructs a forest scene. Figure 6 This assesses the ability of voxels, cover, and vegetation indices to assign weights to pixel AGBs in this forest scene. Individual tree height, diameter at breast height (DBH), crown diameter, and individual tree biomass are shown in Table 9 below.

[0156] Table 9. Statistics of Single Tree Parameters

[0157]

[0158] Based on a virtual scenario, airborne, ground-based laser point cloud, and optical remote sensing data were simulated respectively. The simulated airborne radar equipment was a RIEGL VQ 780i, with a flight altitude of 100m and a point cloud density of 28695.9 pts·m. -2，模拟 The effect is as follows Figure 7 The ground-based radar equipment is a RIEGL VZ-1000 VQ-450, with a point cloud density of 30521.5 pts·m. -2 The simulated ground-based point cloud is as follows: Figure 8 As shown, simulated airborne and ground-based point cloud data are fused, and various indicators such as vegetation index (NDVI), field cover (FVC), and voxels are extracted from them. The fused laser point cloud is shown below. Figure 9 As shown.

[0159] Based on the above three indicators, the single-tree AGB was proportionally allocated to centimeter-level pixels, and the allocated pixel AGB was re-aggregated to the single-tree AGB to verify its accuracy. The accuracy results are shown in Figure 10 below.

[0160] As can be seen from Figure 10, the method based on voxel-based pixel AGB weight allocation performs best, exhibiting the highest coefficient of determination (R²). 2 The highest estimation accuracy was achieved with the lowest root mean square error (RMSE = 33.6 kg) and relative error (rRMSE = 25.36%). The method based on NDVI-assigned AGB weights had the second-highest fitting accuracy (R = 0.9703). 2 =0.9406, RMSE=47kg), but the relative error is relatively large (rRMSE=35.47%); the fitting accuracy based on coverage allocation is the lowest (R 2 =0.9267, RMSE=50.74kg, rRMSE=38.3%. Overall, the voxel distribution-based estimation method performed best, with the highest R². 2 With the lowest RMSE and rRMSE, it is suitable as the preferred method for AGB estimation. In contrast, estimation methods based on NDVI and coverage, while effective to some extent, have larger errors and require further optimization or integration with other data sources. This may be because NDVI and coverage are only two-dimensional parameters and are prone to saturation, while voxels can reflect vertical information and can be arbitrarily aggregated according to satellite image resolution.

[0161] The impact of research methods on model accuracy

[0162] This invention analyzes the accuracy and precision of the research method for biomass estimation from two aspects: remote sensing image classification accuracy and a comparison of pixel AGB allocation weighting methods. The results show that:

[0163] (1) The object-oriented + random forest classification method constructed in this invention achieves an overall classification accuracy of 81.30% and a Kappa coefficient of 0.76, indicating that the classification results have a certain degree of reliability. Looking at the producer and user accuracy for each category, the accuracy for broadleaf forests (A, B, C, D) is all above 80%, indicating that these four categories perform relatively stably during the classification process. The producer accuracy for coniferous forests is 81.25%, while the user accuracy is relatively low at 76.31%, indicating a slightly worse classification effect.

[0164] (2) The method based on voxel-based pixel AGB weighting constructed in this invention has the best performance compared to the other two methods (NDVI and coverage), and has the highest coefficient of determination (R). 2 It has the highest estimation accuracy with the lowest root mean square error (RMSE=33.6kg) and relative error (rRMSE=25.36%).

[0165] Accuracy Validation of Various Types of Biomass Estimation Models

[0166] Biomass inversion models were constructed using multiple linear regression (MLR), random forest (RF), and support vector machine (SVM), respectively, and R was used to... 2 Accuracy was evaluated using coefficient of determination (RMSE), root mean square error (RMSE), and rRMSE (rRMSE percentage). The estimation results for each model are shown in Table 10.

[0167] Table 10 Accuracy of Various AGB Estimation Models

[0168]

[0169] Table 10 shows that different tree species and estimation models exhibit significant differences in performance during the training and validation phases. (Using R...) 2 The main evaluation criteria are the error indices (RMSE and rRMSE).

[0170] Broadleaf forest type A (poplar): The MLR model performed best in both the training and validation phases (R 2 The R values ​​were 0.64 and 0.66 respectively, with rRMSE of 20.33% and 19.41% respectively, indicating high estimation accuracy. In contrast, the R values ​​of the RF and SVM models were lower. 2 The values ​​are very low and the errors are large, especially the rRMSE of the SVM model, which is close to 40%, making it unsuitable for estimating the AGB of poplar trees in the project area.

[0171] Broadleaf forest type B (ash and catalpa): R of MLR and RF models 2 The R value is close to (approximately 0.65) and the error is also relatively small (rRMSE is between 36% and 38%), while the SVM model has a lower R value during the training phase. 2 The performance was good (0.64), but the validation phase R... 2 The significant decrease (0.56) indicates that its generalization ability is weak.

[0172] Broadleaf forest type C (locust, ailanthus, and goldenrain trees): R of the MLR model 2 (Approximately 0.52) Slightly lower than Class B, with rRMSE close to 38%, showing relatively stable performance. The errors of the RF and SVM models are significantly higher, especially the SVM model, whose rRMSE exceeds 120%, making it unsuitable for practical estimation.

[0173] Broadleaf forest type D (golden elm): The MLR model also performed better (R 2 The R² value is 0.59, and the rRMSE is approximately 25%, but the overall accuracy is lower than that of the A-class model. The R² value of the RF model is... 2 The result is almost zero, indicating poor performance. The SVM model has extremely high errors, with rRMSE exceeding 350%, making it unsuitable for estimating this tree species.

[0174] In summary, the MLR model performs better overall for different tree species, especially for broadleaf forest type A and some other tree species (such as broadleaf forest type D). The RF and SVM models, on the other hand, generally have too large an error and are not suitable for biomass estimation.

[0175] Compared with traditional models

[0176] Parametric regression is a traditional method for quantifying biophysical variables in optical remote sensing. This traditional and simple method explicitly determines parametric expressions for a limited number of remote sensing features associated with the biophysical variable of interest. It is an empirical model that relies on the selection of remote sensing features that are highly sensitive to the variable of interest, and vegetation indices are one of the commonly used remote sensing features.

[0177] Based on the research of Dong et al., a statistical regression model of NDVI (Normalized Difference Vegetation Index) and forest biomass was constructed using a large dataset. The model was found to be useful for representing the relationship between forest aboveground biomass and NDVI at spatial and ecological scales over relatively long time periods. The forest aboveground biomass estimation method was used in this invention to compare the accuracy differences between the traditional parametric regression model and the multi-feature regression model used in this invention. The specific results are shown in Table 11 below.

[0178] Table 11 Accuracy of AGB estimation model based on NDVI

[0179]

[0180] The AGB estimation model based on NDVI exhibits significant performance differences across different broadleaf forest categories. For broadleaf forest category A, the R... 2 The R value is relatively high, and the relative root mean square error is the lowest, indicating relatively good model performance; while the R value for broadleaf forest type C is relatively low. 2 The lowest rRMSE and highest rRMSE indicate poor prediction accuracy. The performance of broadleaf forest classes B and D falls between the two, with class D showing the highest RRMSE for both training and validation. 2 The discrepancies are significant, potentially indicating some degree of overfitting. Overall, the model's predictive ability varies across different types, suggesting the need to optimize variable selection and model structure to improve the consistency and reliability of prediction accuracy. Based on the results in the table above, the following conclusions can be drawn:

[0181] (1) When the traditional regression model is applied to the region of this invention, the estimation accuracy is lower than that of the multi-feature regression model selected in this invention.

[0182] (2) Traditional models are limited to a single preset parameter and a simple model, which makes it difficult to reflect the complex interaction between environmental variables and remote sensing features. They also rely on empirical relationships in specific areas and usually have poor generalizability, and may need to be recalibrated.

[0183] (3) Traditional models have low requirements for image spatial resolution. In low-resolution images, since pixels may contain multiple land cover types, traditional models can smooth out the mixed pixel effect to a certain extent and quickly provide large-scale estimation results through global empirical relationships. However, in the higher-resolution images selected in this invention, traditional models cannot make full use of spatial detail information.

[0184] Finally, it should be noted that in this article, relational terms such as first and second are only used to refer to a single...

[0185] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A remote sensing-based method for estimating forest aboveground biomass and its carbon sequestration value, characterized by: Includes the following steps: Acquire preprocessed ground-based LiDAR data of the sample plots and UAV-borne LiDAR data of the sample plots, and perform feature extraction to obtain the individual tree height, crown area and individual tree diameter at breast height of the sample plots. Feature-level fusion of ground-based and UAV-borne LiDAR data was performed to obtain individual tree parameters. The individual tree height and crown area of ​​the sample plot were then input into the diameter at breast height (DBH) optimization model to obtain the DBH of the sample plot. Construct an allometric growth model, input the diameter at breast height (DBH) and tree height of the sample plot, obtain the biomass of individual trees, summarize the biomass of individual trees, and obtain the total biomass of the sample plot; The UAV-borne LiDAR data was processed into voxels, and the biomass of a single point cloud voxel was calculated by combining individual tree parameters. The data was then aggregated at upscale and combined with the total biomass of the sample plot to calculate the biomass of each pixel, thus obtaining a map of the aboveground biomass distribution of the sample plot. Satellite imagery was acquired, spectral and texture features were extracted, and a biomass remote sensing estimation model was constructed using spectral and texture features as independent variables and the aboveground biomass distribution map of the sample plot as the dependent variable. The model was then inverted to estimate the aboveground biomass of the forest.

2. The estimation method according to claim 1, characterized in that: It also includes, Based on forest aboveground biomass, carbon storage is determined and carbon sink value is calculated.

3. The estimation method according to claim 1, characterized in that: The preprocessing includes: ground point detection, point cloud denoising, normalization, and single-tree segmentation.

4. The estimation method according to claim 1, characterized in that: The feature-level fusion of ground-based and UAV-borne LiDAR data involves using the trunk position extracted from the ground-based and UAV-borne LiDAR data to correlate other features.

5. The estimation method according to claim 4, characterized in that: The association rule is as follows: ； Given two feature datasets A and B, and a radius r, the spatial feature connection (A,B) returns a pair of spatial objects (a,b), where a belongs to A and b belongs to B. L(a,b) is the distance between all pairs of points within r. If the distance between the tree trunk locations obtained from ground-based and UAV-borne LiDAR data exceeds r, they are considered not to be the same tree.

6. The estimation method according to claim 1, characterized in that: The breast diameter optimization model is as follows: ； Among them, DBH is the diameter at breast height of a single tree measured by ground-based LiDAR; H and CA are the tree height and crown area of ​​a single tree obtained by airborne LiDAR; a and b are undetermined coefficients, which are determined by using quadrat survey data and the least squares method.

7. The estimation method according to claim 1, characterized in that: The UAV-borne LiDAR data was voxelized by using the stereo pixel method to convert the airborne LiDAR point cloud of the sample plot into voxels and remove empty voxels.

8. The estimation method according to claim 7, characterized in that: The biomass calculation method for the size of a single point cloud voxel is as follows: ； in, It is the first The biomass of each pixel, B is the total biomass of the plot, and T is the total number of elements in the plot. It is the number of all voxels in the vertical direction of that pixel.

9. The estimation method according to claim 8, characterized in that: The formula for calculating upscaling aggregation is: ; ； in, It is the first Biomass per satellite image pixel For satellite image resolution, For voxel resolution, It is the first one to participate in the aggregation within the spatial range corresponding to the satellite pixel. One pixel.

10. The estimation method according to any one of claims 1-9, characterized in that: The biomass remote sensing estimation model was constructed using a machine learning model.

11. A device / equipment / system for estimating forest aboveground biomass and its carbon sink value based on remote sensing, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-10.

12. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-10.

13. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any of claims 1-10.

Images