A high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing

Abstract

The present application relates to the technical field of image extraction, in particular to a high-precision forest biomass estimation method combining airborne laser radar and satellite remote sensing, comprising the following steps: obtaining satellite remote sensing images, airborne laser radar point clouds and sample plot biomass records and unifying coordinates, extracting forest objects according to band values, vegetation indexes, texture edges and spatial connection relationships, cutting point clouds according to forest object boundaries, correcting point cloud attribution according to overlap rates and centroid offset amounts, generating a fusion feature table under object numbering, calculating initial estimation values in combination with multi-source data quality weight, and generating partition correction parameters according to sample plot residuals. The present application associates point cloud structure features, spectral texture features and sample plot records in the same spatial object through objectified image extraction, boundary attribution judgment, quality weight constraint and residual feedback correction, and reduces the influence of boundary mismatch and feature range deviation on estimation results.

Description

Technical Field

[0001] This invention relates to the field of image extraction technology, and in particular to a high-precision method for estimating forest biomass that combines airborne lidar and satellite remote sensing. Background Technology

[0002] Image extraction technology focuses on pixel reading, geometric correction, registration, feature extraction, classification and recognition, spatial indexing, and result storage based on remote sensing images, point cloud data, and ground cover images. Acquisition terminals can include imaging satellites, airborne lidar, and ground recording equipment. Processing terminals establish correspondences based on image grayscale, spectral bands, point cloud coordinates, elevation values, and geographical location, and write the extraction results to local storage or a server database. A traditional high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing involves estimating tree height, canopy structure, vegetation cover, and surface reflectance information in forest areas. This method uses airborne lidar to acquire three-dimensional point cloud and canopy height data, and satellite remote sensing to acquire multispectral or optical images. After coordinate unification, noise removal, image cropping, plot record matching, and feature variable calculation, point cloud height indices, spectral indices, texture features, and measured plot biomass records are input into a regression model or machine learning model. The estimated forest biomass value is then output according to the plot location and pixel range.

[0003] Traditional forest biomass estimation mainly relies on point cloud height indices, spectral indices, texture features, and plot records for direct modeling. Even after coordinate unification, point clouds and images are still affected by boundary mismatches, scale differences, acquisition time differences, and cloud shadow states. Fixed pixels or plot buffers are difficult to correspond to actual canopy objects. When the range of plot records and remote sensing features is inconsistent, the height dispersion and spectral dispersion within the same forest stand enter the same model input. The source range of structural features and image features deviates, forming local biases in regional estimation near plot boundaries. Summary of the Invention

[0004] The purpose of this invention is to provide a high-precision forest biomass estimation method that combines airborne lidar and satellite remote sensing. By using forest objects as multi-source data association units, spatial correspondence, feature fusion, quality weight constraints, and residual correction are performed on airborne lidar point clouds, satellite remote sensing images, and sample plot biomass records. This allows point cloud structural features, image spectral texture features, and sample plot records to participate in forest biomass estimation at a unified object scale.

[0005] To achieve the purpose of this invention, a high-precision method for estimating forest biomass by combining airborne lidar and satellite remote sensing is provided, comprising the following steps: acquiring satellite remote sensing images of forest areas, airborne lidar point clouds, and sample plot biomass records, and performing coordinate unification and overlapping region extraction on the satellite remote sensing images and airborne lidar point clouds; extracting forest objects based on the band values, vegetation indices, texture edges, and spatial connectivity of the satellite remote sensing images, generating forest object boundaries and object numbers; cropping the airborne lidar point clouds according to the forest object boundaries, and determining the point cloud affixation based on the overlap rate and centroid offset between the point cloud projection boundary and the forest object boundary, wherein the overlap rate meets a preset overlap ratio threshold and the centroid offset is within the spatial registration allowable range. Within the tolerance range, the attribution relationship is preserved. When any condition is not met, the corresponding point cloud is marked as a boundary undetermined point cloud and reallocation is performed. The vertical structure features of the point cloud, the spectral texture features of the image, and the biomass record of the sample plot are associated under the object number to generate a fusion feature table. Multi-source data quality weights are generated based on the point cloud sampling density, image cloud shadow state, acquisition time difference, and object feature dispersion. The fusion feature table is input into a preset biomass estimation model, and the features are weighted and calculated in combination with the multi-source data quality weights to obtain the initial estimate of forest biomass. The residual between the sample plot biomass record and the initial estimate of forest biomass is used to generate zoning correction parameters, and the forest biomass estimation result corrected by the zoning correction parameters is output.

[0006] Furthermore, satellite remote sensing images, airborne lidar point clouds, and plot biomass records for the forest area are acquired, and coordinate unification and overlapping area extraction are performed. This includes: reading the imaging time, geographic coordinate range, and band records of the satellite remote sensing images; reading the flight strip number, three-dimensional coordinates, and elevation records of the airborne lidar point clouds; and reading the plot center location and plot boundary of the plot biomass records. When the geographic coordinate range of the satellite remote sensing images overlaps with the projection range of the airborne lidar point clouds, the overlapping area is extracted as a common processing range. The satellite remote sensing images are resampled according to the same geographic coordinate reference, and the coordinates of the airborne lidar point clouds are transformed. When the geographic coordinate range does not overlap with the projection range, invalid data is removed, and feature association of the corresponding area is stopped. After extracting the common processing range, the temporal phase and plot location matching are further performed, retaining the satellite remote sensing images and the airborne lidar point cloud that are within the same vegetation growth window when the imaging time and the flight acquisition time are located; using the center location of the plot as the verification point, the plot boundaries falling within the common processing range are converted to the same geographic coordinate reference; when the plot boundary exceeds the common processing range, the corresponding plot biomass record is marked as not participating in subsequent object association.

[0007] Further, the forest object is extracted based on the band values, vegetation index, texture edges, and spatial connectivity of the satellite remote sensing image, including: marking cloud image elements and removing invalid pixels from the satellite remote sensing image, and forming a normalized image raster with the retained band values ​​according to the same spatial resolution and preset radiometric calibration parameters; calculating the vegetation index from the red band values ​​and near-infrared band values ​​of the normalized image raster, and including image pixels whose vegetation index is within a preset vegetation response range and are not marked as cloud image elements into a candidate vegetation pixel set; truncating the candidate vegetation pixel set according to texture edges, and merging adjacent candidate vegetation pixels according to spatial connectivity; when the merged connected region reaches the minimum effective forest area, meets the edge closure condition and the internal vegetation index consistency condition, the connected region is determined as the forest object and the forest object boundary and the object number are generated. When generating the forest object boundary and the object number, the outer edge pixels of the connected region are converted into vector boundaries, and the image coordinate range, boundary closure state, number of internal pixels, average vegetation index, and texture edge intensity corresponding to the vector boundaries are recorded to form object attribute records. When the interval between two adjacent vector boundaries is less than a preset spatial adjacency critical distance, and the difference in average vegetation index and texture edge intensity corresponding to the two adjacent vector boundaries both meet the same type determination condition, the two adjacent vector boundaries are merged into the same forest object boundary, and the object attribute records before merging are written into the same object number. When any of the same type determination conditions are not met, different object numbers are generated respectively.

[0008] Furthermore, when clipping the airborne lidar point cloud according to the forest object boundary, the forest object boundary is converted to the point cloud clipping range under the same geographic coordinate reference, and the three-dimensional coordinates, elevation records, and echo information falling within the point cloud clipping range are read; when the point cloud sampling density reaches the lower limit of the effective structural feature extraction density, the corresponding point cloud is retained as a point cloud set within the object; when the point cloud sampling density does not reach the lower limit of the density, the corresponding point cloud is marked as a low-density point cloud set, and the multi-source data quality weight corresponding to the point cloud structural features is reduced during subsequent quality weight generation. When determining the point cloud affiliation, the point cloud set clipped based on the forest object boundary is projected onto the image plane to form a point cloud projection boundary. The overlap area between the point cloud projection boundary and the forest object boundary is compared, and the centroid offset between the centroid of the point cloud projection boundary and the centroid of the forest object boundary is calculated. When the ratio of the overlap area to the forest object boundary area meets a preset overlap ratio threshold, and the centroid offset is within the spatial registration tolerance range, the point cloud affiliation is retained. When any condition is not met, the point cloud located in the boundary overlap area is assigned to the boundary set to be assigned, and the boundary set to be assigned is reassigned according to the boundary distance of adjacent forest objects.

[0009] Further, when associating point cloud vertical structure features, image spectral texture features, and the plot biomass records under the object number to generate a fusion feature table, the elevation records and echo information of the point cloud set within the object are read according to the object number, and the height quantiles, average canopy height, canopy height dispersion, and canopy void ratio are statistically analyzed to form point cloud vertical structure features; the band values, vegetation index, texture edge intensity, and number of internal pixels within the corresponding forest object are read according to the object number to form image spectral texture features; the plot boundary and the boundary of adjacent forest objects are sorted by coverage area, and the forest object with the highest coverage area and whose canopy height dispersion and vegetation index dispersion match the plot internal records are identified as the plot-associated object; when the forest object with the highest coverage area does not match the plot internal records, the next forest object is verified sequentially according to the coverage area until the plot-associated object is determined; the point cloud vertical structure features, the image spectral texture features, and the corresponding plot biomass records are written into the same object feature record to form the fusion feature table.

[0010] Further, when generating the multi-source data quality weights, the point cloud sampling density, image cloud shadow status, acquisition time difference, canopy height dispersion, and vegetation index dispersion in the object feature record are read according to the object number; the sorting position of the point cloud sampling density in the same area is converted into a point cloud quality level, the effective pixel ratio corresponding to the image cloud shadow status is converted into an image quality level, and the degree of overlap of vegetation growth periods corresponding to the acquisition time difference is converted into a temporal consistency level; the lowest level among the point cloud quality level, the image quality level, and the temporal consistency level is used as a benchmark to determine the multi-source data quality weights corresponding to the object feature record; when the canopy height dispersion and vegetation index dispersion point to different stand states, the multi-source data quality weights are downgraded.

[0011] Furthermore, the pre-set biomass estimation model is a random forest biomass estimation model containing multiple decision trees. When generating zoning correction parameters, the corresponding point cloud vertical structure features, image spectral texture features, and multi-source data quality weights are called according to the object number to form a model input record. Based on the model input record, the multi-source data quality weights are used as feature scaling coefficients and multiplied by the point cloud vertical structure features and the image spectral texture features respectively. The weighted features are then input into the random forest biomass estimation model to generate an initial estimate of forest biomass for the corresponding forest object. The residual between the initial estimate of forest biomass and the measured biomass record of the sample plot is calculated, and the residual is grouped according to the topographic zoning, canopy height sorting interval, and image object category of the forest object. When the residual direction is consistent within the same group, the corresponding zoning correction parameters are generated and the estimated value of the forest object within the same group is corrected. When the residual direction is inconsistent within the same group, the initial estimate of forest biomass is retained and the corresponding group is marked as a verification group.

[0012] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, forest objects extracted from satellite remote sensing imagery are used as spatial indices. Airborne lidar point cloud cropping and image object boundaries form a unified object range. Overlap rate and centroid offset jointly limit the retention of boundary point clouds, reducing weighting or reassignment, thus minimizing residual local mismatches after coordinate unification. Point cloud vertical structure features, image spectral texture features, and plot biomass records form a fusion feature table under object numbers. Multi-source data quality weights constrain inputs affected by cloud shadows or insufficient density. Zonal correction parameters correct initial estimates based on plot residuals, ensuring consistency between forest biomass estimation results and object-scale data sources, and reducing result fluctuations caused by feature range deviations. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0014] Figure 1 This is the main flowchart of the high-precision forest biomass estimation method of the present invention; Figure 2 This invention provides a data flow graph for coordinate unification and overlapping region extraction. Figure 3 This is a partial flowchart of the forest object extraction process of the present invention; Figure 4 This is a logic diagram for point cloud cropping and point cloud attribution determination in this invention; Figure 5 This invention integrates a feature table, multi-source data quality weights, and a feedback closed-loop diagram of partition correction parameters. Detailed Implementation

[0015] The following combination Figures 1 to 5 The technical solutions described in this invention are explained below. Subsequent embodiments are based on the technical solutions described in the application documents, and use consistent terminology for satellite remote sensing images of forest areas, airborne lidar point clouds, plot biomass records, forest objects, forest object boundaries, object numbers, point cloud projection boundaries, point cloud vertical structure features, image spectral texture features, fusion feature tables, multi-source data quality weights, initial estimates of forest biomass, zoning correction parameters, and forest biomass estimation results. Satellite remote sensing images refer to forest area image data including imaging time, geographic coordinate range, and band records; airborne lidar point clouds refer to point cloud data including flight strip numbers, three-dimensional coordinates, elevation records, and echo information; plot biomass records refer to data including plot center location, plot boundaries, and measured biomass values. New accompanying figures are added later. Figure 1 Start numbering, Figure 1 Used to demonstrate the main process. Figures 2 to 5 The accompanying diagrams are used to illustrate local data flow, local processing logic, and feedback loops. They are only for explaining the method flow, data flow direction, judgment relationship, and output status, and do not include entity component labels.

[0016] Please see Figures 1 to 5 This embodiment provides a high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing. The method establishes multi-source data correspondences around forest objects within a forest area, using the boundaries and object numbers of forest objects generated from satellite remote sensing imagery as spatial indices. This allows airborne lidar point clouds, image features, and plot biomass records to enter the same object-oriented processing chain. The method includes six main steps, S1 to S6. S1 is used for data acquisition, coordinate unification, and overlapping area extraction; S2 is used for forest object extraction; S3 is used for point cloud cropping and point cloud attribution determination; S4 is used for generating a fusion feature table; S5 is used for generating multi-source data quality weights; and S6 is used for initial forest biomass estimation, residual grouping correction, and output of forest biomass estimation results. The data output from each main step serves as input or verification for the next main step. The final output forest biomass estimation results can be traced through object numbers, zoning correction parameters, model input records, and verification grouping markers.

[0017] S1 acquires satellite remote sensing imagery, airborne lidar point clouds, and plot biomass records for the forest area, and performs coordinate unification and overlap region extraction on the satellite remote sensing imagery and airborne lidar point clouds. The inputs to this step are satellite remote sensing imagery, airborne lidar point clouds, and plot biomass records for the forest area. The outputs are the common processing area, the resampled satellite remote sensing imagery, the coordinate-transformed airborne lidar point clouds, and the plot boundaries after geographic coordinate benchmark transformation. After this step, subsequent forest object extraction, point cloud cropping, and plot association are all performed within the common processing area, avoiding unfounded associations between different data sources outside the spatial scope.

[0018] S111 reads the imaging time, geographic coordinate range, and band records of satellite remote sensing imagery. The imaging time is used to match the flight acquisition time of the airborne lidar point cloud; the geographic coordinate range is used to extract overlapping areas with the projection range of the airborne lidar point cloud; and the band records are used for subsequent vegetation index calculations and image spectral texture feature formation. During the reading process, the completeness of the band record fields, the spatial closure of the geographic coordinate range, and the existence of the imaging time are verified. When a band record is missing, the geographic coordinate range cannot form a closed region, or the imaging time is missing, an invalid data marker is set for the corresponding satellite remote sensing image. Satellite remote sensing images with invalid data markers are not included in subsequent feature association.

[0019] S112: Read the flight strip number, 3D coordinates, and elevation records of the airborne lidar point cloud. The flight strip number is used to distinguish the point cloud acquisition path and subsequent point cloud sampling density statistics. The 3D coordinates are used for coordinate transformation and point cloud clipping, and the elevation records are used to form the vertical structural features of the point cloud. During the reading process, the 3D coordinates and elevation records under the same flight strip number are verified. If the same point cloud record lacks 3D coordinates or elevation records, the corresponding point cloud record is discarded as invalid data and will not proceed to the next processing step. After the reading is completed, the airborne lidar point cloud is projected onto the geographic coordinate reference used by the satellite remote sensing image to form the projection range.

[0020] S113, read the plot center location and plot boundary from the plot biomass record. The plot center location is used to determine whether the plot falls within the common treatment range, the plot boundary is used to sort the coverage area with the forest object boundary, and the measured plot biomass record is used to calculate the residual with the initial forest biomass estimate. During the reading process, the plot center location, plot boundary, and measured plot biomass record are checked as necessary fields for the same plot biomass record; when the plot boundary cannot correspond to the plot center location, the plot biomass record is marked as not participating in subsequent object association.

[0021] S114. When the geographic coordinate range of the satellite remote sensing image overlaps with the projected range of the airborne lidar point cloud, the overlapping area is extracted as a common processing range. The common processing range is formed by the spatial intersection of the geographic coordinate range and the projected range, and serves as the spatial boundary for subsequent data clipping and object association. After extracting the common processing range, the satellite remote sensing image is resampled according to the same geographic coordinate reference, and the coordinates of the airborne lidar point cloud are transformed so that the pixel positions of the satellite remote sensing image, the projected positions of the airborne lidar point cloud, and the sample plot boundaries are all on the same geographic coordinate reference. When the geographic coordinate range of the satellite remote sensing image does not overlap with the projected range of the airborne lidar point cloud, the data in the corresponding area is marked as invalid data, and feature association in the corresponding area is stopped.

[0022] S115: After extracting the common processing range, perform temporal and plot location matching. Retain satellite remote sensing images and airborne lidar point clouds whose imaging time and flight acquisition time fall within the same vegetation growth window. The vegetation growth window is determined based on the vegetation growth stage records corresponding to the forest area, used to correlate the image spectral state and point cloud canopy structure state within the same growth stage. Using the plot center as the verification point, the plot boundaries falling within the common processing range are transformed to the same geographic coordinate reference; when the plot boundary exceeds the common processing range, the corresponding plot biomass record is marked as not participating in subsequent object association. This marking serves as the input status for whether to write the plot biomass record when generating the fusion feature table in S4.

[0023] S2 extracts forest objects based on the band values, vegetation indices, texture edges, and spatial connectivity of satellite remote sensing imagery, generating forest object boundaries and object numbers. The input to this step is the satellite remote sensing imagery within the common processing range formed in S1, and the output includes forest objects, forest object boundaries, object numbers, and object attribute records. Forest objects serve as the spatial units for subsequent point cloud cropping, feature association, and estimation results. Forest object boundaries are used to crop the airborne lidar point cloud, and object numbers are used to associate point cloud vertical structure features, image spectral texture features, and plot biomass records.

[0024] S211 performs cloud imagery markers and removes invalid pixels from satellite remote sensing imagery, then uses the retained band values ​​to form a normalized image raster with the same spatial resolution and preset radiometric calibration parameters. Cloud imagery markers are generated based on the cloud cover status of the image. Invalid pixels include pixels lacking band records, whose coordinates fall outside the common processing range, or pixels already marked as invalid data in S1. The normalized image raster retains pixel positions, red band values, near-infrared band values, and other band values ​​that can be used for texture edge calculation, and serves as input for vegetation index calculation, candidate vegetation pixel set generation, and boundary truncation.

[0025] S212: The vegetation index is calculated from the red and near-infrared band values ​​of the normalized image raster. Image pixels whose vegetation index falls within a preset vegetation response range and are not marked as cloud image pixels are included in the candidate vegetation pixel set. The preset vegetation response range represents the range of pixels in the normalized image raster that conform to the spectral response of forest vegetation. Pixels not marked as cloud image pixels are used to exclude pixels affected by cloud shadows. The candidate vegetation pixel set records at least the pixel location, vegetation index, band value source, and cloud image pixel marking status, which are used to subsequently form forest objects based on texture edges and spatial connectivity.

[0026] S213: The candidate vegetation pixel set is truncated according to texture edges, and adjacent candidate vegetation pixels are merged according to spatial connectivity. Texture edges are formed by the grayscale changes and edge continuity of adjacent pixels in the normalized image raster, and are used to delineate the boundary positions of canopy or forest patches in the candidate vegetation pixel set. Spatial connectivity is used to determine whether adjacent candidate vegetation pixels belong to the same continuous forest area. When the merged connected region reaches the minimum effective forest area, meets the edge closure condition, and the internal vegetation index consistency condition, the connected region is identified as a forest object, and the forest object boundary and object number are generated. Connected regions that do not reach the minimum effective forest area, have unclosed edges, or have inconsistent internal vegetation indices are not included as forest objects in subsequent point cloud pruning.

[0027] In step S214, when generating forest object boundaries and object numbers, the outer edge pixels of connected regions are converted into vector boundaries, and the image coordinate range, boundary closure status, number of internal pixels, average vegetation index, and texture edge intensity corresponding to the vector boundaries are recorded to form object attribute records. Each object attribute record corresponds one-to-one with an object number and serves as the data source for image spectral texture features and object feature dispersion in steps S4 and S5. When the interval between two adjacent vector boundaries is less than a preset spatial adjacency threshold distance, and the difference in average vegetation index and texture edge intensity corresponding to the two adjacent vector boundaries both meet the same criteria, the two adjacent vector boundaries are merged into the same forest object boundary, and the object attribute records before merging are written to the same object number. When any of the criteria is not met, different object numbers are generated.

[0028] S3 involves cropping the airborne LiDAR point cloud according to the forest object boundary and determining the point cloud's affiliation based on the overlap rate between the point cloud's projected boundary and the forest object boundary, as well as the centroid offset. The inputs to this step are the forest object boundary and object number generated in S2, and the airborne LiDAR point cloud after coordinate transformation in S1. The outputs are a set of point clouds within the object, a set of low-density point clouds, point clouds with undetermined boundaries, a set of boundary objects to be assigned, and the point cloud affiliation relationships. This step establishes an object-level correspondence between the 3D point cloud structure data and the image object boundary.

[0029] S311, when clipping airborne lidar point clouds according to forest object boundaries, the forest object boundaries are transformed to the point cloud clipping range under the same geographic coordinate reference, and the 3D coordinates, elevation records, and echo information falling within the point cloud clipping range are read. The point cloud clipping range is used to limit the spatial range entering object-level point cloud statistics, the 3D coordinates are used to determine the position of the point cloud relative to the forest object boundary, and the elevation records and echo information are used to form the vertical structural features of the point cloud. When the point cloud sampling density reaches the lower limit of the effective structural feature extraction density, the corresponding point cloud is retained as the point cloud set within the object; when the point cloud sampling density does not reach the lower limit of the density, the corresponding point cloud is marked as a low-density point cloud set, and the multi-source data quality weight corresponding to the point cloud structural features is reduced in subsequent quality weight generation.

[0030] S312, when determining the point cloud affixation, the point cloud set clipped based on the forest object boundary is projected onto the image plane to form the point cloud projection boundary. The overlap area between the point cloud projection boundary and the forest object boundary is compared, and the centroid offset between the centroid of the point cloud projection boundary and the centroid of the forest object boundary is calculated. The point cloud projection boundary represents the spatial coverage of the point cloud set within the object on the image plane, the overlap area reflects the coverage correspondence between the point cloud projection boundary and the forest object boundary, and the centroid offset reflects the consistency of their center positions.

[0031] S313: When the ratio of the overlapping area to the boundary area of ​​the forest object meets the preset overlap ratio threshold, and the centroid offset is within the allowable tolerance range of spatial registration, the point cloud attribution relationship is preserved. The preserved point cloud attribution relationship is written to the point cloud attribution record corresponding to the object number and is available for reading by S4.

[0032] S314: When any condition is not met, the corresponding point cloud is marked as a boundary pending point cloud, and the point clouds located in the boundary overlapping area are included in the boundary pending set. When reassigning the boundary pending set, candidate forest objects are determined according to the boundary distance between adjacent forest objects, and the boundary pending set is assigned to forest objects whose boundary distance correspondence meets the assignment conditions. The reassignment result and the object number together form the updated point cloud assignment record. Point clouds that cannot be reassigned are not included in the object's point cloud set statistics.

[0033] S4, under the object number, associates the vertical structure features of the point cloud, the spectral texture features of the image, and the biomass records of the sample plot to generate a fused feature table. The inputs to this step are the object attribute records generated in S2, the point cloud set and point cloud attribution records within the object generated in S3, and the sample plot biomass records matched by the sample plot location in S1; the outputs are the object feature records and the fused feature table. The fused feature table is used as the input source for the subsequent biomass estimation model.

[0034] S411, read the elevation records and echo information of the point cloud set within the object according to the object number, and statistically analyze the height quantiles, average canopy height, canopy height dispersion, and canopy void ratio to form the vertical structure features of the point cloud. Height quantiles characterize the distribution of point cloud elevations within the same object; average canopy height characterizes the concentration of canopy heights within the object; canopy height dispersion characterizes the differences in canopy structure within the object; and canopy void ratio characterizes the canopy void state in the point cloud echo. All the above statistical results are written into the object feature record using the object number as an index.

[0035] S412 reads the band values, vegetation index, texture edge intensity, and number of internal pixels within the corresponding forest object according to the object number, forming the image spectral texture features. The band values ​​and vegetation index are derived from the normalized image raster, the texture edge intensity is derived from the boundary truncation process in S2, and the number of internal pixels is derived from the object attribute record. The image spectral texture features and the point cloud vertical structure features are associated with the same object number, so that the planar image extraction results and the 3D point cloud structure results are entered into the same object feature record.

[0036] S413, the cover area of ​​the sample plot boundary and the boundary of adjacent forest objects are sorted separately. The forest object with the highest cover area and whose average canopy height and stand characteristics match the records within the sample plot are identified as the sample plot associated object. Cover area sorting is used to determine the main forest objects covered by the sample plot boundary, while average canopy height and stand characteristics are used to verify the consistency of the measured tree height and forest species with the structure and attributes of the candidate forest objects within the sample plot. When the forest object with the highest cover area does not match the records within the sample plot, the next forest object is verified in order of cover area until the sample plot associated object is determined. After the sample plot associated object is determined, the point cloud vertical structure characteristics, image spectral texture characteristics, and the corresponding sample plot biomass record are written into the same object feature record to form a fused feature table. The fused feature table includes at least the object number, point cloud vertical structure characteristics, image spectral texture characteristics, sample plot biomass record, and the status of the sample plot associated object.

[0037] S5 generates multi-source data quality weights based on point cloud sampling density, image cloud shadow status, acquisition time difference, and object feature dispersion. The inputs to this step are the object feature records in the fusion feature table, as well as the point cloud sampling density, image cloud shadow status, acquisition time difference, canopy height dispersion, and vegetation index dispersion generated in S1 to S4. The outputs are the multi-source data quality weights and degradation labels corresponding to the object feature records. The multi-source data quality weights participate in the weighted calculation of features within the fusion feature table in S6.

[0038] S511: Read the point cloud sampling density, image cloud shadow state, acquisition time difference, canopy height dispersion, and vegetation index dispersion from the object feature record according to the object number. The point cloud sampling density comes from the point cloud cropping result in S3, the image cloud shadow state comes from the cloud image meta-labels in S2, the acquisition time difference is formed by the imaging time and the flight acquisition time, the canopy height dispersion comes from the vertical structure features of the point cloud, and the vegetation index dispersion comes from the image spectral texture features. All of the above data items maintain the same object number and are used as inputs to generate multi-source data quality weights.

[0039] S512 converts the sorting position of point cloud sampling density within the same region into point cloud quality level, the effective pixel ratio corresponding to cloud shadow status in the image into image quality level, and the degree of overlap of vegetation growth periods corresponding to the acquisition time difference into temporal consistency level. Point cloud quality level represents the data support status of point cloud structural features within an object; image quality level represents the state of spectral texture features of the image within an object affected by cloud shadows; and temporal consistency level represents the correspondence between satellite remote sensing imagery and airborne lidar point clouds during the vegetation growth period. Using the lowest level among point cloud quality level, image quality level, and temporal consistency level as a benchmark, the multi-source data quality weights corresponding to object feature records are determined, ensuring that the lower-quality data source among multiple data sources constrains subsequent estimation inputs.

[0040] In S513, when the canopy height dispersion and vegetation index dispersion point to different stand states, a downgrade label is applied to the multi-source data quality weights. Stand state is determined by combining the structural differences reflected by the canopy height dispersion and the spectral differences reflected by the vegetation index dispersion within an object. When the two indicate inconsistent stand states, it indicates that the 3D structural features and image features under the same object number have inconsistent attributes. In this case, the downgrade label is written into the object feature record along with the multi-source data quality weights. The downgrade label is used as part of the model input record in S6 to constrain the weighted calculation of the corresponding object.

[0041] S6: Input the fused feature table into the pre-set stochastic forest biomass estimation model, perform weighted calculations on the features based on multi-source data quality weights to obtain the initial forest biomass estimate, and generate zoning correction parameters based on the residuals between the plot biomass records and the initial forest biomass estimate. Output the forest biomass estimation result corrected by the zoning correction parameters. The inputs to this step are the fused feature table, multi-source data quality weights, downgrade markers, and plot biomass records; the outputs are the model input records, the initial forest biomass estimate, residual groupings, zoning correction parameters, verification groupings, and the forest biomass estimation result.

[0042] The pre-built random forest biomass estimation model's algorithm architecture includes multiple decision trees based on classification and regression tree algorithms, achieving biomass regression prediction through an ensemble learning mechanism. The model's training process is as follows: Data entries containing measured plot biomass records (i.e., plot-related objects) from the fusion feature table are extracted to form a training set. Input features are point cloud vertical structure features and image spectral texture features, with the output target variable being the measured biomass value. The training set is then sampled with replacement using the Bootstrap method to form multiple training subsets for training each decision tree. During the node splitting process of each decision tree, a feature subset is randomly selected, and the multi-source data quality weights from the fusion feature table are used as sample weights for the corresponding training samples. The optimal splitting features and splitting points are found based on minimizing the mean squared error after sample weight adjustment, ensuring that high-quality data samples provide a greater gradient contribution when constructing the tree structure, thus completing the model training and construction.

[0043] S611: Based on the object number, the corresponding point cloud vertical structure features, image spectral texture features, and multi-source data quality weights are retrieved to form a model input record. The model input record retains the correspondence between the object number, point cloud vertical structure features, image spectral texture features, multi-source data quality weights, and degradation markers. After reading the model input record, the pre-built random forest biomass estimation model performs a weighted calculation on the point cloud vertical structure features and image spectral texture features to generate an initial estimate of the forest biomass for the corresponding forest object. Specifically, the weighted calculation is implemented as follows: First, based on the model input record, the multi-source data quality weights are parsed into point cloud weight factors and image weight factors according to their source evaluation indicators. Then, the point cloud weight factors are used as feature scaling coefficients and multiplied with the point cloud vertical structure feature vector, and the image weight factors are used as feature scaling coefficients and multiplied with the image spectral texture feature vector (such as band values, vegetation indices, etc.). This dynamically suppresses the feature representation of low-quality data sources and enhances the feature representation of high-quality data sources before entering the model. Finally, the fused features after feature weighting are input into a pre-set random forest biomass estimation model. Each decision tree in the forest outputs an independent prediction value, and the prediction results of all decision trees are arithmetically averaged to generate an initial forest biomass estimate for the corresponding forest object. The initial forest biomass estimate and the object number are written together into the intermediate estimation record as input for residual calculation.

[0044] S612 calculates the residuals between the initial estimate of forest biomass and the measured biomass records of the sample plots, and groups the residuals according to the topographic zoning, canopy height sorting interval, and image object category of the forest object. Topographic zoning is used to distinguish the topographic background of the forest object, canopy height sorting interval is used to distinguish the vertical structural layers of the forest object, and image object category is used to distinguish the image category of the forest object in satellite remote sensing imagery. After the residuals are grouped, the direction of the residuals within the same group is used as the basis for determining the generation of zoning correction parameters.

[0045] S613: When the residual direction remains consistent within the same group, generate the corresponding zoning correction parameters and correct the estimated values ​​of forest objects within the same group. The zoning correction parameters maintain a correspondence with topographic zoning, canopy height sorting intervals, and image object categories. The corrected estimated values, together with the object numbers, form the forest biomass estimation results.

[0046] S614, when the residual directions are inconsistent within the same group, retain the initial forest biomass estimate and mark the corresponding group as a verification group. The verification group label is output along with the initial forest biomass estimate to indicate that the group was not corrected using zoning correction parameters. Through the above processing, the forest biomass estimation results can be traced back to the fusion feature table, multi-source data quality weights, residual grouping, and zoning correction parameters.

[0047] The processes, data items, object numbers, judgment conditions, feature formation methods, quality weight generation methods, residual grouping methods, and output states described in this embodiment are used to explain the possible implementations of the present invention and should not be limited to the processing sequence details, field record formats, or data storage formats listed in this embodiment. Without departing from the scope of the technical solutions described in this invention, any equivalent substitutions, equivalent modifications, equivalent combinations, order adjustments, equivalent transformations of field names, equivalent inheritance of execution entities, or equivalent changes in carrier formats that can be conceived by those skilled in the art should fall within the scope of protection of this patent; however, name changes must not alter the substantial correspondence between satellite remote sensing images, airborne lidar point clouds, plot biomass records, forest objects, forest object boundaries, object numbers, fusion feature tables, multi-source data quality weights, zoning correction parameters, and forest biomass estimation results.

Claims

1. A high-precision method for estimating forest biomass by combining airborne lidar and satellite remote sensing, characterized in that, include: We acquired satellite remote sensing images, airborne lidar point clouds, and plot biomass records of forest areas, and performed coordinate unification and overlapping area extraction on the satellite remote sensing images and airborne lidar point clouds. Forest objects are extracted based on the band values, vegetation index, texture edges, and spatial connectivity of the satellite remote sensing images, and forest object boundaries and object numbers are generated. The airborne lidar point cloud is cropped according to the boundary of the forest object, and the point cloud ownership is determined based on the overlap rate between the point cloud projection boundary and the boundary of the forest object and the centroid offset. When the overlap rate meets the preset overlap ratio threshold and the centroid offset is within the spatial registration tolerance range, the ownership relationship is retained. When either condition is not met, the corresponding point cloud is marked as a boundary undetermined point cloud and reallocation is performed. Under the object number, the vertical structure features of the point cloud, the spectral texture features of the image, and the biomass records of the sample plot are associated to generate a fused feature table; Multi-source data quality weights are generated based on point cloud sampling density, image cloud shadow status, acquisition time difference, and object feature dispersion. The fused feature table is input into a preset biomass estimation model. The features are weighted and calculated in combination with the multi-source data quality weights to obtain an initial estimate of forest biomass. Based on the residual between the sample plot biomass records and the initial estimate of forest biomass, zoning correction parameters are generated, and the forest biomass estimation result corrected by the zoning correction parameters is output.

2. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 1, characterized in that, Acquire satellite remote sensing imagery, airborne lidar point clouds, and plot biomass records for forest areas, and perform coordinate unification and overlapping area extraction, including: The imaging time, geographic coordinate range, and band records of the satellite remote sensing image are read; the flight strip number, three-dimensional coordinates, and elevation records of the airborne lidar point cloud are read; and the center location and boundary of the plot are read from the plot biomass record. When the geographic coordinate range of the satellite remote sensing image overlaps with the projection range of the airborne lidar point cloud, the overlapping area is extracted as a common processing range, the satellite remote sensing image is resampled according to the same geographic coordinate reference, and the coordinates of the airborne lidar point cloud are transformed. When the geographic coordinate range does not overlap with the projection range, invalid data is removed and feature association for the corresponding region is stopped.

3. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 2, characterized in that, After extracting the common processing range, the time phase and sample plot location are further matched: The satellite remote sensing images and the airborne lidar point clouds whose imaging time and flight acquisition time fall within the same vegetation growth window are retained; Using the center location of the sample plot as the verification point, the boundaries of the sample plots falling within the common processing range are transformed to the same geographical coordinate reference. When the sample plot boundary exceeds the common treatment range, the corresponding sample plot biomass record will be marked as not participating in subsequent object association.

4. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 1, characterized in that, The forest objects are extracted based on the band values, vegetation indices, texture edges, and spatial connectivity of the satellite remote sensing imagery, including: The satellite remote sensing images are marked with cloud image elements and invalid pixels are removed. The remaining band values ​​are then used to form a normalized image raster with the same spatial resolution and preset radiometric calibration parameters. The vegetation index is calculated from the red band value and near-infrared band value of the normalized image raster, and image pixels whose vegetation index is in the preset vegetation response range and are not marked as cloud image pixels are included in the candidate vegetation pixel set. The candidate vegetation cell set is truncated according to the texture edge, and adjacent candidate vegetation cells are merged according to spatial connectivity. When the merged connected region reaches the minimum effective forest area, meets the edge closure condition and the internal vegetation index consistency condition, the connected region is determined as the forest object and the boundary and object number of the forest object are generated.

5. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 4, characterized in that, When generating the forest object boundary and the object number, the outer edge pixels of the connected region are converted into vector boundaries, and the image coordinate range, boundary closure state, number of internal pixels, average vegetation index and texture edge intensity corresponding to the vector boundaries are recorded to form object attribute records; When the interval between two adjacent vector boundaries is less than the preset spatial adjacency critical distance, and the difference in average vegetation index and the difference in texture edge intensity corresponding to the two adjacent vector boundaries both meet the same judgment condition, the two adjacent vector boundaries are merged into the same forest object boundary, and the object attribute records before merging are written into the same object number. When any of the aforementioned criteria is not met, different object numbers are generated respectively.

6. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 1, characterized in that, When clipping the airborne lidar point cloud according to the boundary of the forest object, the boundary of the forest object is converted to the clipping range of the point cloud under the same geographic coordinate reference, and the three-dimensional coordinates, elevation records and echo information falling within the clipping range of the point cloud are read. When the point cloud sampling density reaches the lower limit of the effective structural feature extraction density, the corresponding point cloud is retained as the point cloud set within the object. When the point cloud sampling density does not reach the lower limit of the density, the corresponding point cloud is marked as a low-density point cloud set, and the multi-source data quality weight corresponding to the point cloud structural features is reduced during subsequent quality weight generation.

7. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 6, characterized in that, When determining the point cloud affiliation, the point cloud set clipped based on the forest object boundary is projected onto the image plane to form a point cloud projection boundary. The overlapping area between the point cloud projection boundary and the forest object boundary is compared, and the centroid offset between the centroid of the point cloud projection boundary and the centroid of the forest object boundary is calculated. When the ratio of the overlapping area to the boundary area of ​​the forest object meets the preset overlap ratio threshold, and the centroid offset is within the spatial registration tolerance range, the point cloud attribution relationship is preserved. If any condition is not met, the point cloud located in the boundary overlapping area is included in the boundary to be assigned set, and the boundary to be assigned set is reassigned according to the boundary distance of adjacent forest objects.

8. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 6, characterized in that, When associating point cloud vertical structure features, image spectral texture features, and the sample plot biomass records under the object number and generating a fusion feature table, the elevation records and echo information of the point cloud set within the object are read according to the object number, and the height quantiles, average canopy height, canopy height dispersion, and canopy void ratio are statistically analyzed to form point cloud vertical structure features. According to the object number, the band value, vegetation index, texture edge intensity and number of internal pixels in the corresponding forest object are read to form the image spectral texture features; The cover area of ​​the sample plot boundary and the boundary of the adjacent forest object are sorted separately. The forest object with the highest cover area and whose canopy height dispersion and vegetation index dispersion match the records inside the sample plot are identified as the sample plot associated object. When the forest object ranked first by coverage area does not match the record inside the sample plot, the next forest object is checked in order of coverage area until the associated object of the sample plot is determined. The point cloud vertical structure features, the image spectral texture features, and the corresponding plot biomass records are written into the same object feature record to form the fused feature table.

9. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 8, characterized in that, When generating the multi-source data quality weights, the point cloud sampling density, image cloud shadow status, acquisition time difference, canopy height dispersion, and vegetation index dispersion in the object feature record are read according to the object number. The sorting position of the point cloud sampling density within the same area is converted into point cloud quality level, the effective pixel ratio corresponding to the image cloud shadow state is converted into image quality level, and the degree of overlap of vegetation growth period corresponding to the acquisition time difference is converted into time consistency level. The lowest level among the point cloud quality level, the image quality level, and the temporal consistency level is used as a benchmark to determine the multi-source data quality weight corresponding to the object feature record; When the dispersion of canopy height and vegetation index points to different stand states, the quality weights of the multi-source data are downgraded.

10. The high-precision forest biomass estimation method combining airborne lidar and satellite remote sensing according to claim 9, characterized in that, The preset biomass estimation model is a random forest biomass estimation model that includes multiple decision trees. When generating partition correction parameters, the corresponding point cloud vertical structure features, image spectral texture features and multi-source data quality weights are called according to the object number to form the model input record; Based on the model input record, the multi-source data quality weights are used as feature scaling coefficients to perform dot product weighting with the point cloud vertical structure features and the image spectral texture features, respectively. The weighted features are then input into the random forest biomass estimation model to generate initial estimates of forest biomass for the corresponding forest objects. The residuals between these estimates and the measured biomass records are calculated, and the residuals are grouped according to the topographic zoning, canopy height sorting interval, and image object category of the forest objects. When the residual direction is consistent within the same group, the corresponding partition correction parameters are generated and the estimated values ​​of forest objects within the same group are corrected. When the residual directions are inconsistent within the same group, the initial estimate of forest biomass is retained and the corresponding group is marked as a verification group.

Images