Forest carbon sequestration amount prediction method and system based on carbon sink model

Abstract

The application discloses a forest carbon fixation amount prediction method and system based on a carbon sink model, relates to the technical field of forestry carbon sinks, and comprises the following steps: S1, collecting forest sample plots, remote sensing images and terrain grid data, constructing a forest carbon sink plane coordinate system, generating a slope surface projection correction sample plot boundary and a sample plot pixel mismatch evidence base element; S2, constructing a mismatch traction atlas, calculating a mismatch traction normalized value and screening a sample plot mismatch dominant pixel set; S3, projecting the sample plot ground carbon storage label value back to the dominant pixel carbon label value to generate a carbon sink model training sample set; and S4, taking the mismatch traction normalized value as a training weight to train a forest carbon pool inversion model and outputting a target pixel carbon fixation amount prediction result. The application solves the problem that in the existing forest carbon fixation amount prediction, the sample plot ground carbon storage label and the remote sensing pixel features are difficult to accurately correspond, leading to the problems of mismatch of training labels, offset of prediction boundaries and insufficient stability of carbon fixation amount inversion in complex terrain and forest edge regions.

Description

Technical Field

[0001] This invention relates to the field of forestry carbon sequestration technology, and in particular to a method and system for predicting forest carbon sequestration based on carbon sequestration models. Background Technology

[0002] As the "dual carbon" goals are advanced, carbon emission accounting, carbon sink assessment, and carbon asset management are gradually shifting from manual statistical methods to data-driven intelligent forecasting. Current technologies for predicting carbon emissions, carbon emission trends, and carbon-related indicators typically collect regional economic data, energy consumption data, transportation data, industry operation data, or multi-source sample data. These are then combined with data processing methods such as machine learning models, graph neural network models, random forest models, and inference models to establish a mapping relationship between carbon emission-related factors and prediction results, thereby improving the automation level and data processing efficiency of carbon emission forecasting.

[0003] For example, application CN116611576B provides a carbon emission prediction method and apparatus, relating to the field of data processing technology, to address the problems of low accuracy, weak robustness, and poor generalization ability in existing carbon emission prediction methods. The method includes: acquiring carbon emission correlation data to be predicted, wherein the carbon emission correlation data to be predicted includes N sample data, where N is an integer greater than or equal to 2; calculating the relationship strength value between any two data samples in the N sample data to obtain at least one relationship strength value, and the at least one relationship strength value constitutes an adjacency matrix; inputting the adjacency matrix into a graph convolutional neural network to extract features at different scales in the carbon emission correlation data to be predicted, obtaining multi-scale features; and inputting the multi-scale features into a random forest model based on dynamic weights to predict carbon emissions.

[0004] For example, application CN114066097B discloses a carbon emission prediction method, which includes: obtaining the annual average growth rate of turnover of different modes of transportation in the target area, the annual average growth rate of mileage of transportation route construction, and the annual average change rate of carbon emissions; obtaining the annual average growth rate of GDP and the annual average growth rate of resident population in the target area; obtaining the predicted annual average growth rate of turnover of different modes of transportation in the target area; obtaining the annual average energy use optimization rate of different existing energy sources in the target area; obtaining the annual average energy structure optimization rate of different new energy sources in the target area; obtaining the predicted annual average optimization rate of carbon emission intensity of different modes of transportation in the target area; obtaining the predicted carbon emissions of different modes of transportation in the target area in the target year; and obtaining the predicted development trend of carbon emissions in the target area.

[0005] However, existing carbon prediction schemes mostly focus on the overall prediction of regional carbon emissions, carbon emission trends in transportation, or multi-source carbon emission correlation data. They typically use macroeconomic statistical data, sample relationship strength, or energy structure change parameters as model inputs, making it difficult to directly address the problem of inaccurate spatial correspondence between sample plot ground carbon storage labels and remote sensing pixel features in forest carbon sequestration prediction. Especially in cases of complex terrain, mixed forest edges, and sample plot positioning errors, sample plot boundaries may cover multiple remote sensing pixels. Single-pixel matching at the center point can easily lead to training label mismatch and prediction boundary shift, resulting in insufficient stability of forest carbon sequestration inversion.

[0006] Therefore, in order to address the above problems, there is an urgent need for a method and system for predicting forest carbon sequestration based on carbon sink models. Summary of the Invention

[0007] To address the problem that existing methods for predicting forest carbon sequestration often fail to accurately match ground carbon storage labels in sample plots with remote sensing pixel features, leading to mismatches in training labels for complex terrain and forest edges, shifted prediction boundaries, and insufficient stability in carbon sequestration retrieval, this invention provides a method and system for predicting forest carbon sequestration based on a carbon sink model. The technical solution is as follows: On the one hand, a method for predicting forest carbon sequestration based on a carbon sink model is provided, which includes: S1. Collect forest plot survey data, remote sensing image data and topographic grid data, construct a forest carbon sink plane coordinate system, generate slope projection correction plot boundary polygons, and generate plot pixel mismatch evidence primitives by combining center offset distance value, boundary coverage ratio value and remote sensing response difference value. S2, construct mismatch traction map based on sample plot pixel mismatch evidence primitives, calculate center offset normalization value, boundary coverage normalization value, slope drag normalization value, location diffusion normalization value and remote sensing difference normalization value, generate mismatch traction normalization value, and filter sample plot mismatch dominant pixel set; S3, based on the mismatched dominant pixel set of the sample plot, back-projects the values ​​of aboveground biomass, belowground biomass, litter and soil organic carbon storage into the corresponding dominant pixel carbon label values, and generates carbon label back-projection records and carbon sink model training sample set. S4 uses the mismatch traction normalization value as the training weight, trains the forest carbon pool inversion model based on the carbon sink model training sample set, generates the total carbon storage prediction value and carbon sequestration prediction value for the target pixel, and outputs the forest carbon sequestration prediction result.

[0008] Further, the specific steps for collecting forest plot survey data, remote sensing image data, and topographic grid data to construct a forest carbon sink plane coordinate system and generate slope projection-corrected plot boundary polygons are as follows: Collect forest plot survey data, remote sensing image data, and topographic grid data. The forest plot survey data includes plot number, plot center point coordinates, plot boundary vertex coordinates, plot positioning accuracy level marker, tree species code, age group code, aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, soil organic carbon storage, and plot ground carbon storage label. The remote sensing image data includes remote sensing image number, pixel row number, pixel column number, pixel center coordinates, pixel spatial resolution, normalized difference vegetation index (NDVI), enhanced vegetation index (VEI), canopy cover, and texture response. The topographic grid data includes grid center coordinates, slope, aspect, and slope projection scale. The southwest corner of the target predicted forest area's circumscribed rectangle is used as the coordinate origin. A forest carbon sink plane coordinate system is constructed with due east as the positive X-axis and due north as the positive Y-axis. The coordinates of the sample plot center point, sample plot boundary vertex coordinates, pixel center coordinates, and grid center coordinates are transformed into the forest carbon sink plane coordinate system to generate corresponding plane coordinate values. The initial sample plot boundary polygon is generated by closing and connecting the plane coordinates of the sample plot boundary vertex in the collection order. The positioning error radius value is read according to the sample plot positioning accuracy level mark. The slope value, aspect value, and slope projection ratio value within the coverage area of ​​the initial sample plot boundary polygon are read, and the average slope value, average aspect value, and average slope projection ratio value are calculated. The direction of the slope projection offset vector value is determined according to the average aspect value. The length of the slope projection offset vector value is determined according to the product of the square root of the area value of the initial sample plot boundary polygon, the tangent of the average slope value, and the average slope projection ratio value. The initial sample plot boundary polygon is translated along the slope projection offset vector value to generate the slope projection corrected sample plot boundary polygon.

[0009] Further, the specific steps for generating plot pixel mismatch evidence primitives by combining center offset distance, boundary coverage ratio, and remote sensing response difference are as follows: Generate pixel boundary polygons based on pixel center plane coordinates and pixel spatial resolution; determine the center-falling pixel number based on the plot center point plane coordinates; and generate center offset distance based on the distance between the plot center point plane coordinates and the center plane coordinates of the center-falling pixel; filter pixel boundary polygons that intersect with the slope projection-corrected plot boundary polygons to generate a boundary coverage pixel set; generate a boundary coverage ratio value based on the ratio of the intersecting area value to the slope projection-corrected plot boundary polygon area value; generate a remote sensing response difference value based on the mean absolute value of the differences between the normalized vegetation index, enhanced vegetation index, canopy cover, and texture response values ​​between the boundary coverage pixel and the center-falling pixel; and combine the center-falling pixel number, boundary coverage pixel set, positioning error radius value, slope projection offset vector value, center offset distance value, boundary coverage ratio, and remote sensing response difference value to generate plot pixel mismatch evidence primitives.

[0010] Furthermore, based on the sample plot pixel mismatch evidence primitives, the specific steps for constructing a mismatch traction map and calculating the center offset normalization value, boundary cover normalization value, slope drag normalization value, location diffusion normalization value, and remote sensing difference normalization value are as follows: Read the sample plot pixel mismatch evidence primitives, and construct the sample plot center point, the slope projection-corrected sample plot boundary polygon, the center-falling pixel, and the boundary-covering pixel as mismatch traction nodes. Establish a center offset traction edge with the center offset distance value, a boundary cover traction edge with the boundary cover ratio value, and a slope projection offset vector value. To the dragging traction edge and the positioning diffusion traction edge with the written positioning error radius value, the mismatched traction node and each traction edge are associated with the same location number to generate a mismatched traction map; the center offset distance value, the slope projection offset vector value length, and the positioning error radius value are divided by the pixel spatial resolution value to generate the center offset normalized value, the slope drag normalized value, and the positioning diffusion normalized value; the remote sensing response difference value is divided by the interquartile range of the remote sensing response difference values ​​within the same remote sensing image to generate the remote sensing difference normalized value, and the boundary coverage ratio value is used as the boundary coverage normalized value.

[0011] Further, the specific steps for generating mismatch traction normalization values ​​and screening the sample plot mismatch-dominant pixel set are as follows: Calculate the mismatch traction value for each boundary cover pixel. The numerator of the mismatch traction value is the product of the boundary cover normalization value and the remote sensing difference normalization value, plus the product of the slope drag normalization value and the location diffusion normalization value. The denominator of the mismatch traction value is the product of the center offset normalization value and the center fall suppression coefficient. Normalize the mismatch traction values ​​under the same plot number to generate mismatch traction normalization values. Accumulate the boundary cover pixels according to the mismatch traction normalization values ​​from largest to smallest. Write the boundary cover pixels corresponding to the accumulated mismatch traction normalization values ​​that reach the mismatch dominance retention threshold into the sample plot mismatch-dominant pixel set.

[0012] Furthermore, based on the mismatched dominant pixel set of the sample plot, the specific steps for back-projecting the aboveground biomass, belowground biomass, litter, and soil organic carbon storage values ​​into corresponding dominant pixel carbon label values ​​are as follows: Read the mismatched dominant pixel set, aboveground biomass carbon storage value, belowground biomass carbon storage value, litter carbon storage value, and soil organic carbon storage value of the sample plot; write the aboveground biomass carbon storage value and belowground biomass carbon storage value into the biomass carbon pool label vector, and write the litter carbon storage value and soil organic carbon storage value into the environmental constraint carbon pool label vector; read the mismatched traction normalization value, boundary coverage ratio value, and remote sensing difference normalization value of each dominant pixel under the same plot number; calculate the product of the mismatched traction normalization value, boundary coverage ratio value, and remote sensing difference normalization value to generate the initial value of biomass carbon inheritance for the dominant pixel; and normalize the sum of the initial values ​​under the same plot number to generate the biomass carbon label inheritance coefficient; based on the pixel center plane coordinates of the dominant pixel... The slope and aspect values ​​of the dominant pixels are read from the topographic grid data to generate the slope and aspect classification values ​​of the dominant pixels. Using the tree species code, age group code, slope and aspect classification values ​​of the dominant pixels corresponding to the plot number as lookup indexes, the environmental constraint baseline coefficients are read from the environmental constraint coefficient table. The environmental constraint baseline coefficients are multiplied by the boundary coverage ratio to generate the environmental constraint coefficients of the dominant pixels. The environmental constraint coefficients of all dominant pixels under the same plot number are normalized to generate the environmental constraint carbon label carrying capacity coefficients. The biomass carbon pool label vector is multiplied by the biomass carbon label carrying capacity coefficient of each dominant pixel to generate the aboveground biomass carbon label value and the belowground biomass carbon label value of the dominant pixels. The environmental constraint carbon pool label vector is multiplied by the environmental constraint carbon label carrying capacity coefficient of each dominant pixel to generate the litter carbon label value and the soil organic carbon label value of the dominant pixels.

[0013] Further, the specific steps for generating carbon-labeled back-projection records and carbon sink model training sample sets are as follows: Carbon-labeled back-projection records are generated according to the plot number, remote sensing image number, pixel row number, and pixel column number, combined with the corresponding dominant pixel carbon label value, carbon label inheritance coefficient, and mismatch traction normalization value. When multiple carbon-labeled back-projection records correspond to the same remote sensing image number, the same pixel row number, and the same pixel column number, the total back-projection carbon label value of each carbon-labeled back-projection record is calculated separately, and the range of the total back-projection carbon label value is divided by the mean of the corresponding plot's ground carbon storage label value to generate a label conflict value. Carbon-labeled back-projection records with label conflict values ​​greater than the label conflict threshold are written into the sample isolation queue, and the remaining records, along with their corresponding mismatch traction normalization values, are written into the carbon sink model training sample set. When only one carbon-labeled back-projection record corresponds to the same pixel, the corresponding record, along with its corresponding mismatch traction normalization value, is written into the carbon sink model training sample set.

[0014] Furthermore, using mismatch-driven normalized values ​​as training weights, the specific steps for training the forest carbon pool inversion model based on the carbon sink model training sample set are as follows: Read the carbon sink model training sample set, and use the normalized vegetation index, enhanced vegetation index, canopy cover, texture response, slope, aspect, slope projection ratio, tree species coding, and age group coding as model inputs. Use the dominant pixel aboveground biomass carbon label, dominant pixel belowground biomass carbon label, dominant pixel litter carbon label, and dominant pixel soil organic carbon label as model output labels, and use the corresponding mismatch-driven normalized values ​​as training weights to train the forest carbon pool inversion model.

[0015] Further, the specific steps for generating total carbon storage prediction values ​​and carbon sequestration prediction values ​​for target pixels and outputting forest carbon sequestration prediction results are as follows: Read the remote sensing image data, topographic grid data, and forest stand attribute layer data of the target predicted forest area. The forest stand attribute layer data includes stand sub-compartment boundary polygons, tree species coding values, and age group coding values; according to the pixel row number, pixel column number, and the overlap relationship between the target pixel and the stand sub-compartment boundary polygons, extract the normalized vegetation index value, enhanced vegetation index value, canopy cover value, texture response value, slope value, aspect value, slope projection ratio value, tree species coding value, and age group coding value of the target pixel to generate the target pixel carbon sink model input record; The target pixel carbon sink model input record is input into the forest carbon pool inversion model to generate the predicted values ​​of aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, and soil organic carbon storage for the target pixel. The four carbon storage prediction values ​​are then added together to generate the total carbon storage prediction value for the target pixel. The total carbon storage baseline value for the target pixel in the baseline measurement period is read, and the total carbon storage prediction value is subtracted from the total carbon storage baseline value to generate the carbon sequestration prediction value for the target pixel. The forest carbon sequestration prediction result is then generated according to the remote sensing image number, pixel row number, pixel column number, total carbon storage prediction value, and carbon sequestration prediction value for the target pixel.

[0016] On the other hand, a forest carbon sequestration prediction system based on a carbon sink model is provided. This system is applied to the forest carbon sequestration prediction method based on a carbon sink model. The system includes: The sample plot mismatch primitive generation module is used to collect forest sample plot survey data, remote sensing image data and topographic grid data, construct a forest carbon sink plane coordinate system, generate slope projection corrected sample plot boundary polygons, and generate sample plot pixel mismatch evidence primitives by combining center offset distance value, boundary coverage ratio value and remote sensing response difference value. The mismatch traction map construction module is used to construct a mismatch traction map based on the sample plot pixel mismatch evidence primitives, calculate the center offset normalization value, boundary coverage normalization value, slope drag normalization value, location diffusion normalization value and remote sensing difference normalization value, generate mismatch traction normalization value, and filter the sample plot mismatch dominant pixel set. The carbon label back projection reconstruction module is used to back project the values ​​of aboveground biomass, belowground biomass, litter and soil organic carbon storage into the corresponding carbon label values ​​of the dominant pixels based on the mismatch of the dominant pixel set of the sample plot, and generate carbon label back projection records and carbon sink model training sample set. The carbon sequestration prediction result write-back module is used to train the forest carbon pool inversion model based on the carbon sink model training sample set, using mismatch traction normalization values ​​as training weights, to generate total carbon storage prediction values ​​and carbon sequestration prediction values ​​for target pixels, and output forest carbon sequestration prediction results.

[0017] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: (1) By constructing a forest carbon sink plane coordinate system and combining the slope projection offset vector value to generate a slope projection correction plot boundary polygon, the plot boundary is no longer matched only according to the center point falling into the pixel. This can more accurately reflect the real coverage relationship between the plot and the remote sensing pixel under complex terrain, mixed forest edge and positioning error conditions, and reduce the probability of mismatch between the plot ground carbon storage label and the remote sensing feature.

[0018] (2) Based on the sample plot pixel mismatch evidence primitives, a mismatch traction map is constructed, and mismatch traction normalization value is generated by center offset normalization value, boundary coverage normalization value, slope drag normalization value, positioning diffusion normalization value and remote sensing difference normalization value. This makes the sample plot-pixel mismatch no longer a single area weight problem, but can be quantitatively characterized from multiple perspectives such as center offset, boundary coverage, slope drag, positioning error and remote sensing response difference, thereby improving the interpretability of mismatch judgment.

[0019] (3) By using the mismatch of the dominant pixel set of the sample plot, the ground carbon storage label value of the sample plot is back-projected into the aboveground biomass carbon label value, underground biomass carbon label value, litter carbon label value and soil organic carbon label value of the dominant pixel, and carbon label back-projection records and carbon sink model training sample set are generated. This avoids directly binding the overall carbon storage label of the sample plot to the central pixel and improves the consistency between the pixel-level training label and the real forest carbon storage distribution.

[0020] (4) When training the forest carbon pool inversion model, the mismatch traction normalization value is used as the training weight, so that the high confidence dominant pixels contribute more to the model training, and the contribution of samples with greater mismatch risk is suppressed, thereby improving the stability and reliability of the predicted total carbon storage and carbon sequestration of target pixels in complex terrain areas, forest edge areas and fragmented patch areas. Attached Figure Description

[0021] 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 these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for predicting forest carbon sequestration based on a carbon sink model; Figure 2 This is a schematic diagram of the structure of a forest carbon sequestration prediction system based on a carbon sink model; Figure 3 This is a schematic diagram of the sample plot mismatch-dominated pixel selection; Figure 4This is a schematic diagram of the plot pixel mismatch traction and carbon label back projection. Detailed Implementation

[0023] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0024] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0025] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0026] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0027] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0028] like Figure 1 As shown in the figure, this invention provides a method for predicting forest carbon sequestration based on a carbon sink model. The processing flow of this method may include the following steps: S1. Collect forest plot survey data, remote sensing image data and topographic grid data, construct a forest carbon sink plane coordinate system, generate slope projection correction plot boundary polygons, and generate plot pixel mismatch evidence primitives by combining center offset distance value, boundary coverage ratio value and remote sensing response difference value. S2, construct mismatch traction map based on sample plot pixel mismatch evidence primitives, calculate center offset normalization value, boundary coverage normalization value, slope drag normalization value, location diffusion normalization value and remote sensing difference normalization value, generate mismatch traction normalization value, and filter sample plot mismatch dominant pixel set; S3, based on the mismatched dominant pixel set of the sample plot, back-projects the values ​​of aboveground biomass, belowground biomass, litter and soil organic carbon storage into the corresponding dominant pixel carbon label values, and generates carbon label back-projection records and carbon sink model training sample set. S4 uses the mismatch traction normalization value as the training weight, trains the forest carbon pool inversion model based on the carbon sink model training sample set, generates the total carbon storage prediction value and carbon sequestration prediction value for the target pixel, and outputs the forest carbon sequestration prediction result.

[0029] Optionally, the specific steps for collecting forest plot survey data, remote sensing image data, and topographic grid data to construct a forest carbon sink plane coordinate system and generate slope projection-corrected plot boundary polygons are as follows: Collect forest plot survey data, remote sensing image data, and topographic grid data. The forest plot survey data includes plot number, plot center point coordinates, plot boundary vertex coordinates, plot positioning accuracy level marker, tree species code, age group code, aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, soil organic carbon storage, and plot surface carbon storage label. The plot number is generated from the field plot survey record sheet, and the plot center point coordinates... The values ​​of the sample plot boundary vertex coordinates were collected by a GNSS receiver. The sample plot positioning accuracy level was determined by the differential state, fixed solution state, horizontal accuracy factor value, and field re-measurement records of the GNSS receiver. The tree species code value and age group code value were entered by the sample plot survey personnel according to the forest stand sub-compartment survey form. The aboveground biomass carbon storage value and belowground biomass carbon storage value were calculated from the tree diameter at breast height (DBH), tree height, tree species biomass equation, and carbon content. The litter carbon storage value was calculated from the dry mass value of litter in the sample plot and the litter carbon content. The soil organic carbon storage value was calculated from the soil bulk density value, soil layer thickness value, and soil organic carbon content value. The sample plot ground carbon storage label value was obtained by... The carbon storage values ​​are obtained by summing aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, and soil organic carbon storage. Remote sensing image data includes the remote sensing image number, pixel row number, pixel column number, pixel center coordinates, pixel spatial resolution, normalized vegetation index (NVI), enhanced vegetation index (EVI), canopy cover, and texture response. The remote sensing image number is generated from the image reception batch and imaging timestamp; the pixel row and column numbers are generated from the image raster index; the pixel center coordinates are read from the remote sensing image georeferenced file; the pixel spatial resolution is read from the remote sensing image metadata; and the normalized vegetation index is calculated by combining the near-infrared reflectance value with the red band reflectance value. The reflectance values ​​of the light bands were calculated. The enhanced vegetation index was calculated from the reflectance values ​​of the near-infrared band, red band, and blue band. The canopy coverage value was calculated from the ratio of the vegetation index threshold segmentation result to the pixel area. The texture response value was calculated from the contrast and homogeneity values ​​of the near-infrared band gray-level co-occurrence matrix. The topographic grid data included grid center coordinates, slope, aspect, and slope projection scale. Among them, the grid center coordinates were obtained by raster index conversion from the digital elevation model, the slope and aspect values ​​were calculated from the first difference of adjacent elevation grids, and the slope projection scale was a dimensionless parameter with a value range of 0.05 to 0.35. The slope distance is obtained by correcting the ratio of the difference between the slope distance value and the plane distance value within the same topographic grid unit to the plane distance value using a plot resurvey correction coefficient. The slope distance value is calculated based on the three-dimensional distance between the centers of adjacent digital elevation model grids, and the plane distance value is calculated based on the horizontal distance between the plane coordinates of the centers of adjacent grids. The plot resurvey correction coefficient is taken from 0.60 to 1.20 and is determined by the ratio of the historical plot boundary resurvey offset to the digital elevation model slope unfolding offset within the same forest area. The slope projection ratio is used to characterize the degree to which the slope topography participates in the correction of the plot boundary plane projection. The larger the slope projection ratio, the greater the slope projection... The greater the translation correction magnitude of the offset vector value to the initial sample plot boundary polygon, the more accurate the forest carbon sink plane coordinate system is. Using the southwest corner of the predicted forest area's circumscribed rectangle as the origin, and the east direction as the positive X-axis and the north direction as the positive Y-axis, a forest carbon sink plane coordinate system is constructed. The coordinates of the sample plot center point, sample plot boundary vertex, pixel center, and grid center are transformed to the forest carbon sink plane coordinate system to generate corresponding plane coordinate values. Specifically, using the latitude and longitude of the origin as the translation reference, each coordinate value is first converted to a unified projected coordinate system, and then the corresponding projected coordinates of the origin are subtracted to obtain the plane coordinates of the sample plot center point and the sample plot boundary vertex. Point plane coordinates, pixel center plane coordinates, and grid center plane coordinates are used to ensure that subsequent distance, area, and boundary intersection calculations are all performed in meter-level plane units. The initial sample plot boundary polygon is generated by closing and connecting the plane coordinates of the sample plot boundary vertices in the acquisition order. The positioning error radius value is read according to the sample plot positioning accuracy level mark. The vertex connection order of the initial sample plot boundary polygon follows the clockwise acquisition order in the field records; during the closing connection, the plane coordinates of the last sample plot boundary vertex are connected to the plane coordinates of the first sample plot boundary vertex. The positioning error radius value is read from the positioning accuracy mapping table. The positioning table is established jointly by the differential state, fixed solution state, horizontal accuracy factor value, and field re-measurement closure difference value output by the GNSS receiver. When the positioning accuracy level of the sample plot is marked as GNSS fixed solution, the positioning error radius value is 0.03 meters to 0.10 meters; when the positioning accuracy level of the sample plot is marked as GNSS floating-point solution, the positioning error radius value is 0.30 meters to 1.00 meters; when the positioning accuracy level of the sample plot is marked as GNSS differential single-point solution, the positioning error radius value is 1.00 meters to 3.00 meters; when the positioning accuracy level of the sample plot is marked as GNSS single-point positioning, the positioning error radius value is 3.00 meters to 10.00 meters; when both have the same field re-measurement closure difference, take the larger of the upper limit of the corresponding level error radius and the field re-measurement closure difference as the positioning error radius value; read the slope value, aspect value and slope projection ratio value within the coverage area of ​​the initial sample plot boundary polygon, and calculate the average slope value, average aspect value and average slope projection ratio value; specifically, filter the terrain grid records whose grid center plane coordinate values ​​fall inside the initial sample plot boundary polygon, perform an arithmetic mean of the slope value and slope projection ratio value, and calculate the average aspect value using the unit direction vector synthesis method for the aspect value, specifically by converting each aspect value into a sine component and a cosine component. The mean values ​​of the sine and cosine components are calculated separately. Then, the average aspect value is generated based on the arctangent of the mean values ​​of the sine and cosine components, avoiding a jump in the mean value when the aspect angle is close to 0° and 360°. The direction of the slope projection offset vector value is determined based on the average aspect value, with the average aspect value starting at 0° in the north direction and increasing clockwise. The downhill direction corresponding to the average aspect value is taken as the positive direction of the slope projection offset vector value. The length of the slope projection offset vector value is determined by multiplying the square root of the area of ​​the initial sample plot boundary polygon, the tangent of the average slope value, and the average slope projection scale value. In the forest carbon sink plane coordinate system, the slope projection offset direction... The X-axis component of the magnitude is the product of the offset length and the sine of the average slope aspect value, while the Y-axis component of the slope projection offset vector is the product of the offset length and the cosine of the average slope aspect value. The average slope value is converted to radians before calculating the tangent. The square root of the area of ​​the initial plot boundary polygon is used to characterize the plot's planar scale, the tangent of the average slope value is used to characterize the displacement amplification relationship of the slope relative to the horizontal plane, and the average slope projection scale value is used to characterize the degree of participation of topographic projection correction. Multiplying the square root of the area of ​​the initial plot boundary polygon, the tangent of the average slope value, and the average slope projection scale value yields the length of the slope projection offset vector. This calculation... This system is suitable for plot boundary correction scenarios where the plot boundary scale is less than three times the spacing between adjacent terrain grids and the average slope value is no greater than 35°. When the average slope value is greater than 35°, the length of the slope projection offset vector value is limited to half of the pixel spatial resolution value. The initial plot boundary polygon is translated along the slope projection offset vector value to generate the slope projection corrected plot boundary polygon. Specifically, the slope projection offset vector value is added to the plane coordinate value of each vertex of the initial plot boundary polygon to obtain the corresponding corrected vertex plane coordinate value, and the corrected vertex plane coordinate values ​​are closed and connected according to the original acquisition order to generate the slope projection corrected plot boundary polygon.

[0030] In this implementation plan, forest plot survey data, remote sensing image data, and topographic grid data are unified into the forest carbon sink plane coordinate system. Based on slope, aspect, and slope projection ratio, a slope projection-corrected plot boundary polygon is generated. This ensures that the plot center point coordinates, plot boundary vertex coordinates, pixel center coordinates, and grid center coordinates have the same spatial measurement benchmark, reducing the interference of plot boundary plane position offset on subsequent pixel matching under complex terrain. At the same time, the slope projection offset vector value transforms the terrain aspect traction relationship into a calculable boundary correction amount, enabling subsequent inference models to read boundary coverage ratio, center offset distance, and remote sensing response difference values ​​based on a more accurate plot-pixel spatial correspondence, thereby improving the spatial credibility and verifiability of plot pixel mismatch evidence primitives.

[0031] Optionally, the specific steps for generating plot pixel mismatch evidence primitives by combining center offset distance values, boundary coverage ratio values, and remote sensing response difference values ​​are as follows: Generate pixel boundary polygons based on pixel center plane coordinates and pixel spatial resolution values. Specifically, using the pixel center plane coordinates as the geometric center, and half the pixel spatial resolution value as the horizontal and vertical half-side lengths, calculate the plane coordinates of the lower left, lower right, upper right, and upper left corners of the pixel, respectively. Close these coordinates in a clockwise order to generate the pixel boundary polygons, enabling the pixel row and column numbers in the remote sensing image data to be converted into planar geometric units capable of area intersection calculations. The center-falling cell number is determined based on the planar coordinates of the sample plot center point. Specifically, the planar coordinates of the sample plot center point are compared with the polygons of each cell boundary for point-to-face inclusion. The cell row number and cell column number corresponding to the polygons of the cell boundary containing the planar coordinates of the sample plot center point are written into the center-falling cell number. When the planar coordinates of the sample plot center point are located on the common boundary of two cell boundaries, the cell with the smaller cell row number is selected as the center-falling cell. When the planar coordinates of the sample plot center point are located on the common vertex of multiple cells, the cell row numbers are compared first, and the cell with the smaller cell row number is selected. Then, among the cells with the same cell row number, the cell with the smaller cell column number is selected as the center-falling cell. This unique attribution rule is used to determine the center-falling cell. This process ensures that the same plot center point's planar coordinates correspond to only one center-falling pixel number. A center offset distance value is generated based on the distance between the plot center point's planar coordinates and the center coordinates of the center-falling pixel. This center offset distance value is calculated using planar Euclidean distance and characterizes the degree of deviation of the plot center point from the geometric center of the center-falling pixel. A larger center offset distance value indicates a higher probability of spatial mismatch when using a single center point pixel to represent the plot's ground carbon storage label value. Pixel boundary polygons that intersect with the slope projection-corrected plot boundary polygons are selected to generate a set of boundary-covering pixels. Specifically, polygon intersection operations are performed between the slope projection-corrected plot boundary polygons and each pixel boundary polygon. When the area value of the intersecting region is greater than zero, the remote sensing image number, pixel row number, pixel column number, and pixel center plane coordinate value corresponding to the intersecting pixel are written into the boundary cover pixel set. This ensures that subsequent processing no longer relies solely on the pixel number where the center falls, but retains all pixels that are actually covered by the plot after slope correction. The boundary cover ratio is generated by the ratio of the intersecting area value to the area value of the slope projection correction plot boundary polygon. The intersecting area value is calculated from the overlapping area of ​​the slope projection correction plot boundary polygon and the boundary polygon of the corresponding pixel of the boundary cover pixel. The area value of the slope projection correction plot boundary polygon is used as the denominator, so that the boundary cover ratio represents the relative share of the plot plane cover contribution received by each boundary cover pixel.Remote sensing response difference values ​​are generated based on the mean absolute values ​​of the differences between boundary-covered pixels and center-fallen pixels in the normalized vegetation index (NWRI), enhanced vegetation index (VEI), canopy cover, and texture response values. Specifically, the NWRI, VEI, canopy cover, and texture response values ​​corresponding to all boundary-covered pixels and center-fallen pixels under the same remote sensing image number are first used as statistical objects. The NWRI, VEI, canopy cover, and texture response values ​​are then subjected to min-max normalization to generate standard values ​​for the NWRI, VEI, and texture response values. Response standard values; when the maximum and minimum values ​​of any indicator are equal, the corresponding standard value of that indicator is recorded as 0 to avoid the denominator being zero; calculate the absolute value of the difference between the normalized vegetation index standard value of the boundary cover pixel and the normalized vegetation index standard value of the center-falling pixel, the absolute value of the difference between the enhanced vegetation index standard value of the boundary cover pixel and the enhanced vegetation index standard value of the center-falling pixel, the absolute value of the difference between the canopy cover standard value of the boundary cover pixel and the canopy cover standard value of the center-falling pixel, and the absolute value of the difference between the texture response standard value of the boundary cover pixel and the texture response standard value of the center-falling pixel. The four absolute differences were averaged with equal weights to generate a remote sensing response difference value. Each of the four absolute differences had a weight of 0.25. The normalized vegetation index (NVI) and enhanced vegetation index (EFI) standard values ​​were used to characterize vegetation spectral response differences, canopy cover standard values ​​were used to characterize canopy cover structure differences, and texture response standard values ​​were used to characterize stand spatial texture differences. The remote sensing response difference value reflected the degree of difference between boundary cover pixels and center-falling pixels in forest carbon storage response. The values ​​included center-falling pixel number, boundary cover pixel set, positioning error radius, slope projection offset vector, center offset distance, and boundary cover... The plot cell mismatch evidence primitive is generated by combining the plot ratio value and the remote sensing response difference value. Specifically, the plot number is used as the main index of the plot cell mismatch evidence primitive, and the remote sensing image number, cell row number, and cell column number are used as the boundary cover cell sub-record index. The center-falling cell number, positioning error radius value, and slope projection offset vector value are written into the main index record. The cell center plane coordinate value, intersection area value, boundary cover ratio value, center offset distance value, and remote sensing response difference value corresponding to each boundary cover cell are written into the boundary cover cell sub-record. The association relationship between the plot number and each boundary cover cell sub-record is established.When the plot pixel mismatch evidence unit is used for subsequent mismatch-guided mapping, it reads the center-falling pixel number, positioning error radius value, and slope projection offset vector value according to the plot number. It also reads the boundary coverage ratio value, center offset distance value, and remote sensing response difference value corresponding to each boundary-covering pixel according to the remote sensing image number, pixel row number, and pixel column number. This allows the plot pixel mismatch evidence unit to simultaneously record center point mismatch, boundary cover mismatch, positioning error mismatch, slope projection mismatch, and remote sensing response mismatch.

[0032] In this implementation plan, by uniformly incorporating the pixel boundary polygon, the slope projection-corrected plot boundary polygon, the center offset distance value, the boundary coverage ratio value, and the remote sensing response difference value into the plot pixel mismatch evidence unit, the spatial deviation, coverage deviation, and response deviation between the plot ground carbon storage label value and the remote sensing pixel characteristics can form the same traceable evidence record, avoiding label binding offset caused by relying solely on the center falling into the pixel number. At the same time, the plot pixel mismatch evidence unit provides a unified input for subsequent mismatch traction normalization value calculation and plot mismatch dominant pixel set selection, transforming the plot-pixel correspondence under complex terrain from single-point matching to multi-evidence constraints, improving the spatial credibility of carbon label back projection records and the stability of forest carbon pool inversion model training data.

[0033] Optionally, the specific steps for constructing a mismatch traction map based on the sample plot pixel mismatch evidence primitives and calculating the center offset normalization value, boundary coverage normalization value, slope drag normalization value, positioning diffusion normalization value, and remote sensing difference normalization value are as follows: Read the sample plot pixel mismatch evidence primitives, extract the center-falling pixel number, boundary coverage pixel set, positioning error radius value, slope projection offset vector value, center offset distance value, boundary coverage ratio value, and remote sensing response difference value according to the sample plot number, and use the same sample plot number as the map construction index to ensure that the center point matching information, boundary coverage information, slope correction information, and positioning error information corresponding to the same sample plot are included in the same mismatch traction map. The spectrum; mismatch traction nodes are constructed by identifying the sample plot center point, the slope projection-corrected sample plot boundary polygon, the center-falling pixel, and the boundary-covering pixel, and corresponding node attributes are written for each type of mismatch traction node; Specifically, the sample plot center point node is written with the sample plot number, the plane coordinates of the sample plot center point, and the sample plot positioning accuracy level marker; the slope projection-corrected sample plot boundary polygon node is written with the sample plot number, the plane coordinates of the corrected vertex, the slope projection offset vector value, and the polygon area value; the center-falling pixel node is written with the remote sensing image number, the pixel row number, the pixel column number, the pixel center plane coordinates value, and the pixel spatial resolution value; and the boundary-covering pixel node is written with the remote sensing image number. The data includes: cell row number, cell column number, cell center plane coordinates, intersection area, and boundary coverage ratio; establishing a center offset traction edge between the sample plot center point node and the center-falling cell node, and writing the center offset distance value into the center offset traction edge; establishing a boundary coverage traction edge between the slope projection corrected sample plot boundary polygon node and each boundary coverage cell node, and writing the boundary coverage ratio value into the boundary coverage traction edge; establishing a slope aspect dragging traction edge between the slope projection corrected sample plot boundary polygon node and each boundary coverage cell node, and writing the slope projection offset vector value into the slope aspect dragging traction edge; establishing a center offset traction edge between the sample plot center point node and each boundary coverage cell node; and establishing a center offset traction edge between the slope projection corrected sample plot boundary polygon node and each boundary coverage cell node. A positioning diffusion traction edge is established between boundary-covered cell nodes, and the positioning error radius value is written into the positioning diffusion traction edge; mismatched traction nodes and each traction edge are associated according to the same land area number to generate a mismatched traction map; among them, the center offset traction edge is used to characterize the degree of deviation of the sample plot center point from the center falling into the geometric center of the cell, the boundary cover traction edge is used to characterize the share of the boundary cover cell in the sample plot plane range, the slope aspect drag traction edge is used to characterize the terrain traction strength of the slope value, slope aspect value and slope projection ratio value on the translation of the sample plot boundary, and the positioning diffusion traction edge is used to characterize the planar position uncertainty introduced by the sample plot positioning accuracy level mark.In subsequent calculations of center offset normalization, boundary coverage normalization, slope drag normalization, location diffusion normalization, and remote sensing difference normalization, the center offset distance, boundary coverage ratio, slope projection offset vector, and location error radius carried by the corresponding traction edge are read from the mismatched traction map according to the plot number. This ensures that the mismatched traction map serves not only as a graphical representation of relationships but also as a data organization carrier for normalization calculations. The center offset distance, slope projection offset vector length, and location error radius are divided by the pixel spatial resolution value to generate the center offset normalization, slope drag normalization, and location diffusion normalization values. The pixel spatial resolution value serves as a unified scale benchmark, with meters as the unit. The center offset distance, slope projection offset vector length, and positioning error radius are all converted into dimensionless proportions relative to the width of a single remote sensing pixel to avoid incomparability of mismatch degrees caused by different spatial resolution values ​​of remote sensing images. The remote sensing response difference value is divided by the interquartile range of remote sensing response difference values ​​within the same remote sensing image to generate a normalized remote sensing difference value, and the boundary coverage ratio is used as the normalized boundary coverage value. The statistical population of remote sensing response difference values ​​within the same remote sensing image is the remote sensing response difference value corresponding to the boundary coverage pixel in all sample plot pixel mismatch evidence primitives under the same remote sensing image number. The remote sensing response difference values ​​in the statistical population are arranged in ascending order, and the 25th percentile is read from each value. The interquartile range (IQR) is calculated by subtracting the 25th percentile from the 75th percentile. When the number of remote sensing response differences under the same remote sensing image number is less than 8, the IQR corresponding to the same tree species code value and age group code value within the same target predicted forest area is included in the statistical population. If the number of IQR differences after inclusion is still less than 8, 0.10 is used as a substitute IQR. When the calculated IQR is less than 0.05, 0.05 is used as a substitute denominator to avoid abnormal amplification of the normalized remote sensing difference value due to the IQR being zero or too small. Through IQR normalization, the amplification effect of extreme pixel responses on the normalization result is weakened. The system aims to ensure that the remote sensing difference normalization value can stably characterize the differences in vegetation spectrum, canopy structure, and stand texture between boundary cover pixels and center-falling pixels. The boundary cover ratio value itself is the ratio of the intersection area to the area of ​​the slope projection-corrected plot boundary polygon, which is already within the dimensionless ratio range and therefore directly used as the boundary cover normalization value. Through center offset normalization, boundary cover normalization, slope drag normalization, location diffusion normalization, and remote sensing difference normalization, center point offset, boundary cover, terrain drag, location diffusion, and remote sensing response differences are converted into mismatch drag inputs at the same scale, providing a verifiable data foundation for subsequent mismatch drag value calculation and selection of the plot mismatch-dominated pixel set.

[0034] In this implementation plan, by converting the center offset distance value, boundary coverage ratio value, slope projection offset vector value, positioning error radius value, and remote sensing response difference value into center offset normalized value, boundary coverage normalized value, slope drag normalized value, positioning diffusion normalized value, and remote sensing difference normalized value, the spatial offset, boundary connection, terrain drag, positioning diffusion, and remote sensing response difference in the sample plot pixel mismatch evidence primitives can be entered into the mismatch traction map at the same scale, reducing the traction bias of the single center-falling pixel number on the sample plot ground carbon storage label value. At the same time, the mismatch traction map can convert the geometric contribution, terrain influence, and vegetation response difference of boundary cover pixels into comparable mismatch traction inputs, improving the stability of subsequent mismatch traction value calculation and sample plot mismatch-dominated pixel set selection, so that the carbon label back projection record can better reflect the true correspondence between sample plots and remote sensing pixels under complex terrain.

[0035] Optionally, the specific steps for generating mismatch traction normalization values ​​and filtering the set of mismatch-dominant pixels in the sample plot are as follows: Figure 3As shown, the mismatch traction value for each boundary cover cell is calculated. Before calculation, the boundary cover normalization value, remote sensing difference normalization value, slope drag normalization value, location diffusion normalization value, and center offset normalization value corresponding to each boundary cover cell under the same location number are read. The normalization values ​​corresponding to the same boundary cover cell are then written into the same cell traction calculation record, ensuring that the mismatch traction value corresponds to a specific remote sensing image number, cell row number, and cell column number. The numerator of the mismatch traction value is the boundary cover normalization value plus a remote sensing difference value. The product of the heterogeneous normalization values, plus the product of the slope drag normalization value and the location diffusion normalization value, is used to characterize the geometrical inheritance contribution and the remote sensing difference dominance contribution of boundary cover pixels within the sample plot coverage area. In this step, the remote sensing difference normalization value does not represent the matching reliability between boundary cover pixels and center-falling pixels, but rather indicates that when center-falling pixels are difficult to represent the vegetation spectrum, canopy structure, and stand texture characteristics of boundary cover pixels, the boundary cover pixels need to be included separately. The intensity of mismatch correction within the carbon label coverage area; therefore, the larger the remote sensing difference normalization value, the less representative the single-pixel matching result of the center point is of the boundary coverage pixel, and the higher the traction contribution of the boundary coverage pixel to correcting the plot-pixel mismatch; the product of the slope drag normalization value and the positioning diffusion normalization value is used to characterize the spatial uncertainty contribution caused by the combined effects of slope projection offset and field positioning error; the denominator of the mismatch traction value is the product of the center offset normalization value and the center fall suppression coefficient, where the center fall suppression coefficient is derived from the plot-pixel mismatch. The verification parameter table is read, with values ​​ranging from 0.10 to 0.40, determined by the proportion of single-pixel matching error at the center point of the historical sample plot. When the proportion of single-pixel matching error at the center point of the historical sample plot is less than 10%, the center-fall-in suppression coefficient is taken as 0.10 to 0.18; when the proportion of single-pixel matching error at the center point of the historical sample plot is 10% to 25%, the center-fall-in suppression coefficient is taken as 0.18 to 0.30; when the proportion of single-pixel matching error at the center point of the historical sample plot is greater than 25%, the center-fall-in suppression coefficient is taken as 0.30 to 0.40; The denominator is used to suppress the center pixel traction deviation caused when the center point of the sample plot deviates from the center and falls into the geometric center of the pixel, so that the traction strength of the boundary cover pixel is not controlled by the single pixel matching result of the center point; In this implementation plan, the product of the boundary cover normalization value and the remote sensing difference normalization value, and the product of the slope drag normalization value and the positioning diffusion normalization value are all added into the numerator of the mismatch traction value by equal weighting, and no longer set manually adjustable weights, so as to avoid changing the screening result of the sample plot mismatch dominant pixel set due to human weight setting; The mismatch traction values ​​under the same plot number are normalized to generate the mismatch traction normalization value. Specifically, the sum normalization method is adopted, and all mismatch traction values ​​under the same plot number are first normalized. The mismatched drag values ​​of boundary-covered pixels are summed to generate a total mismatched drag value. Then, the mismatched drag value of each boundary-covered pixel is divided by the total mismatched drag value to obtain the corresponding mismatched drag normalization value for that pixel, ensuring that the sum of all mismatched drag normalization values ​​under the same land area number is 1. When the total mismatched drag value under the same land area number is 0, the boundary coverage ratio value of each boundary-covered pixel under the same land area number is used as a substitute normalization basis, and normalization is performed according to the sum of the boundary coverage ratio values. When the sum of the boundary coverage ratio values ​​is still 0, the mismatched drag normalization value of each boundary-covered pixel under the same land area number is set to the same value, which is 1 divided by the number of boundary-covered pixels. The normalization is then performed from largest to smallest. For boundary cover pixels with small cumulative values, the boundary cover pixels corresponding to the point where the cumulative mismatch traction normalization value reaches the mismatch dominance retention threshold are written into the sample plot mismatch dominance pixel set. Specifically, boundary cover pixels are first sorted in descending order of mismatch traction normalization value, and then the mismatch traction normalization value is accumulated one by one starting from the first position. Accumulation stops when the cumulative value first reaches the mismatch dominance retention threshold, and the boundary cover pixel number, remote sensing image number, pixel row number, pixel column number, mismatch traction value, and mismatch traction normalization value participating in the accumulation are written into the sample plot mismatch dominance pixel set. When multiple boundary cover pixels have the same mismatch traction normalization value, the boundary cover pixels with larger boundary cover ratio values ​​are prioritized. When the mismatch traction normalization value is... When the coverage ratio and boundary ratio are the same, the pixels are arranged in ascending order of row number and column number to ensure the uniqueness of the set of mismatch-dominated pixels in the sample plot. The mismatch-dominated retention threshold is read from the sample plot verification parameter table and ranges from 0.75 to 0.90, determined by the carbon storage error distribution of the verification sample plots under the same forest stand type. When the coefficient of variation of the carbon storage error of the verification sample plots under the same forest stand type is less than 0.15, the mismatch-dominated retention threshold is 0.75 to 0.80; when the coefficient of variation is 0.15 to 0.30, the mismatch-dominated retention threshold is 0.80 to 0.85; when the coefficient of variation is greater than 0.30, the mismatch-dominated retention threshold is 0.85 to 0.90; The mismatch dominance retention threshold is used to retain boundary cover pixels that primarily bear carbon storage values ​​from the four types of sample plots, while excluding edge pixels with low traction contributions, ensuring that subsequent carbon label back-projection records are concentrated in dominant pixels with major mismatch traction contributions.

[0036] In this implementation plan, by incorporating boundary cover normalization value, remote sensing difference normalization value, slope drag normalization value, location diffusion normalization value, and center offset normalization value into the mismatch traction value, and converting the mismatch traction value into a mismatch traction normalization value, the set of mismatch-dominant pixels in the sample plots is screened. This transforms the relationship between boundary cover pixels and the back projection of carbon storage values ​​of the four types of sample plots from area overlap judgment to multi-factor traction judgment, reducing the one-way interference of edge pixels and center-falling pixels on carbon label allocation. At the same time, the mismatch dominance retention threshold can constrain the range of pixels participating in carbon label back projection records, so that the set of mismatch-dominant pixels in the sample plots mainly retains the boundary cover pixels that make the main contribution, improving the relevance of subsequent generation of carbon label values ​​for dominant pixels and the reliability of training data for forest carbon pool inversion models.

[0037] Optionally, based on the mismatched dominant pixel set of the sample plot, the specific steps for back-projecting the aboveground biomass, belowground biomass, litter, and soil organic carbon storage values ​​into corresponding dominant pixel carbon label values ​​are as follows: Read the mismatched dominant pixel set of the sample plot, and extract the aboveground biomass carbon storage value, belowground biomass carbon storage value, litter carbon storage value, and soil organic carbon storage value under the same sample plot number. Write the aboveground biomass carbon storage value and the belowground biomass carbon storage value into the biomass carbon pool label vector, and write the litter carbon storage value and the soil organic carbon storage value into the environmental constraint carbon pool label vector; wherein, the biomass carbon pool label vector is used to connect with canopy spectrum, canopy cover, and stand texture. For carbon pool labels with closer responsiveness, environmentally constrained carbon pool label vectors are used to inherit carbon pool labels that are more closely related to long-term environmental conditions such as tree species, age group, slope, and aspect. This ensures that aboveground biomass carbon storage values, belowground biomass carbon storage values, litter carbon storage values, and soil organic carbon storage values ​​are processed separately according to the carbon pool response mechanism during back projection. The mismatch traction normalization value, boundary cover ratio value, and remote sensing difference normalization value of each dominant pixel under the same site number are read. The product of the mismatch traction normalization value, boundary cover ratio value, and remote sensing difference normalization value is calculated to generate the initial value for dominant pixel biomass carbon inheritance. This initial value is then normalized according to the sum of the initial values ​​under the same site number to generate the biomass carbon label inheritance system. Specifically, the initial biomass carbon carrying capacity of each dominant pixel is divided by the sum of the initial biomass carbon carrying capacity of all dominant pixels under the same land area number to obtain the biomass carbon tag carrying capacity coefficient of the corresponding dominant pixel. This ensures that the sum of the biomass carbon tag carrying capacity coefficients of all dominant pixels is 1, thereby guaranteeing the conservation of aboveground and belowground biomass carbon storage values ​​among the dominant pixels. When the sum of the initial biomass carbon carrying capacity of all dominant pixels under the same land area number is 0, the boundary cover ratio of each dominant pixel is used as a substitute carrying capacity basis, and normalized according to the sum of the boundary cover ratio values. When the sum of the boundary cover ratio values ​​is still 0, the biomass carbon carrying capacity of each dominant pixel under the same land area number is... The biomass carbon label carrying capacity coefficient is set to the same value, which is 1 divided by the number of dominant pixels. Among them, the mismatch traction normalization value is used to characterize the traction strength of the dominant pixels to the back projection of the sample plot carbon label, the boundary coverage ratio value is used to characterize the area carrying share of the dominant pixels to the slope projection correction sample plot boundary polygon, the remote sensing difference normalization value is used to characterize the vegetation response difference of the dominant pixels relative to the center falling pixels, and the addition of the remote sensing difference normalization value is used to retain the basic carrying capacity of pixels with consistent remote sensing response. Based on the pixel center plane coordinate value of the dominant pixels, the slope value and aspect value of the dominant pixels are read from the topographic grid data to generate the slope classification value and aspect classification value of the dominant pixels.Specifically, the topographic grid records where the plane coordinates of the dominant pixel center fall are used as the source of topographic values. The slope values ​​of the dominant pixels are divided into four intervals: greater than or equal to 0° and less than or equal to 5°, greater than 5° and less than or equal to 15°, greater than 15° and less than or equal to 25°, and greater than 25°, with the unit of slope value being degrees. The aspect values ​​of the dominant pixels are divided into northward, eastward, southward, and westward, with the unit of aspect value being degrees. When the plane coordinates of the dominant pixel center are exactly... When the data falls on the boundary of two terrain grids, the terrain grid record with the smaller grid row number is read first. When the plane coordinates of the dominant pixel center fall exactly on the common vertex of multiple terrain grids, the terrain grid record with the smaller grid row number is selected first, and then the terrain grid record with the smaller grid column number is selected from the terrain grid records with the same grid row number, ensuring that the dominant pixel slope classification value and the dominant pixel aspect classification value have unique values. Using the tree species code value, age group code value, dominant pixel slope classification value, and dominant pixel aspect classification value corresponding to the plot number as the lookup index, the environmental constraint benchmark coefficient is read from the environmental constraint coefficient table. The environmental constraint benchmark coefficient is multiplied by the boundary coverage ratio value to generate the environmental constraint coefficient of the dominant pixel, and the same... The environmental constraint coefficients of all dominant pixels under the site number are normalized to generate environmental constraint carbon labeling carrying coefficients. The environmental constraint coefficient table is indexed according to tree species code values, age group code values, slope classification values, and aspect classification values. Each index record corresponds to an environmental constraint baseline coefficient, which is a dimensionless coefficient ranging from 0.20 to 1.00. The environmental constraint coefficient table is jointly calibrated using historical sample plot data from the target predicted forest area and sample plot data from similar forest stands in adjacent forest areas. Specifically, under the same tree species code values, age group code values, slope classification values, and aspect classification values, the ratio of the sum of litter carbon storage and soil organic carbon storage values ​​in historical sample plots to the sample plot's surface carbon storage label value is calculated. For example, the average proportion is linearly normalized to the minimum and maximum values ​​of all average proportions within the target predicted forest area and mapped to 0.20 to 1.00 to obtain the corresponding environmental constraint benchmark coefficient. When a combination of tree species code value, age group code value, slope classification value, and aspect classification value does not have a corresponding record in the environmental constraint coefficient table, the average environmental constraint benchmark coefficient corresponding to adjacent slope classification values ​​and adjacent aspect classification values ​​under the same tree species code value and the same age group code value is first read as a substitute value. When adjacent slope classification values ​​and adjacent aspect classification values ​​still do not have a corresponding record, the average environmental constraint benchmark coefficient corresponding to all slope classification values ​​and all aspect classification values ​​under the same tree species code value and the same age group code value is read as a substitute value.When no corresponding records exist for the same tree species code and the same age group code, the average of all environmental constraint baseline coefficients within the target predicted forest area is used as a substitute value. The environmental constraint baseline coefficient is multiplied by the boundary cover ratio to obtain the dominant pixel environmental constraint coefficient. Then, each dominant pixel's environmental constraint coefficient is divided by the sum of all dominant pixel environmental constraint coefficients under the same location number to obtain the corresponding dominant pixel's environmental constraint carbon label carrying capacity coefficient, ensuring that the sum of all dominant pixel environmental constraint carbon label carrying capacity coefficients under the same location number is 1. When the sum of all dominant pixel environmental constraint coefficients under the same location number is 0, the boundary cover ratio of each dominant pixel is used. The value is used as a substitute for acceptance, and normalized according to the sum of the boundary coverage ratio values; when the sum of the boundary coverage ratio values ​​is still 0, the environmental constraint carbon label acceptance coefficient of each dominant cell under the same land number is set to the same value, which is 1 divided by the number of dominant cells; the environmental constraint carbon label acceptance coefficient is used to determine the spatial acceptance share of litter carbon storage value and soil organic carbon storage value among dominant cells, and is not used to change the numerical ratio between litter carbon storage value and soil organic carbon storage value; since both litter carbon storage value and soil organic carbon storage value are affected by the long-term forest land environmental constraints formed by tree species, age group, slope and aspect, and both are in this step Both belong to relatively stable environmentally constrained carbon pool labels. Therefore, the same environmentally constrained carbon tag acceptance coefficient is used to determine the spatial distribution position of the two between the dominant pixels. At the same time, the litter carbon storage value and the soil organic carbon storage value are retained as two independent input values ​​to avoid merging the two types of carbon pool labels. In the environmental constraint coefficient table calibration stage, the back projection error of the litter carbon label and the back projection error of the soil organic carbon label after sharing the environmental constraint carbon tag acceptance coefficient are cross-validated. When both types of errors are lower than the allowable error threshold for the re-measurement of the target predicted forest area sample plot, the corresponding tree species code value, age group code value, slope classification value and aspect classification value are combined and written into the environmental constraint coefficient table. The binding coefficient table is used to generate the aboveground biomass carbon label value and the belowground biomass carbon label value of the dominating pixel by multiplying the biomass carbon pool label vector by the biomass carbon label carrying coefficient of each dominating pixel. Specifically, the aboveground biomass carbon storage value is multiplied by the biomass carbon label carrying coefficient of the dominating pixel to obtain the aboveground biomass carbon label value of the dominating pixel, and the belowground biomass carbon storage value is multiplied by the same biomass carbon label carrying coefficient of the dominating pixel to obtain the belowground biomass carbon label value of the dominating pixel. The environmental constraint carbon pool label vector is multiplied by the environmental constraint carbon label carrying coefficient of each dominating pixel to generate the litter carbon label value and the soil organic carbon label value of the dominating pixel.Specifically, the litter carbon storage value is multiplied by the environmental constraint carbon label carrying coefficient of the dominating pixel to obtain the dominating pixel litter carbon label value. The soil organic carbon storage value is multiplied by the same environmental constraint carbon label carrying coefficient to obtain the dominating pixel soil organic carbon label value. The aboveground biomass carbon label value, belowground biomass carbon label value, litter carbon label value, and soil organic carbon label value of the dominating pixel are then bound to the remote sensing image number, pixel row number, and pixel column number to form a pixel-level carbon label record that can participate in the training of the forest carbon pool inversion model.

[0038] In this implementation plan, aboveground biomass carbon storage values ​​and belowground biomass carbon storage values ​​are written into the biomass carbon pool label vector, while litter carbon storage values ​​and soil organic carbon storage values ​​are written into the environmental constraint carbon pool label vector. Then, based on the biomass carbon label carrying capacity coefficient and the environmental constraint carbon label carrying capacity coefficient, dominant pixel aboveground biomass carbon label values, dominant pixel belowground biomass carbon label values, dominant pixel litter carbon label values, and dominant pixel soil organic carbon label values ​​are generated. This allows the carbon storage values ​​of the four types of sample plots to be converted into pixels corresponding to the remote sensing image number, pixel row number, and pixel column number. Level carbon labeling avoids binding the entire ground carbon storage label value of the sample plot to the central pixel, which would cause training label mismatch. At the same time, the biomass carbon pool label vector and the environmental constraint carbon pool label vector are constrained by mismatch-driven normalization value, boundary coverage ratio value, remote sensing difference normalization value, tree species coding value, age group coding value, dominant pixel slope classification value, and dominant pixel aspect classification value, respectively. This ensures that the training samples obtained by the inference model have spatial continuity, remote sensing response, and topographic environment evidence, thereby improving the stability and verifiability of the forest carbon pool inversion model in predicting carbon sequestration in forest areas with complex terrain.

[0039] Optionally, the specific steps for generating carbon-labeled back-projection records and carbon sink model training sample sets are as follows: Based on the plot number, remote sensing image number, pixel row number, and pixel column number, and combined with the corresponding dominant pixel carbon label value, carbon label carrying capacity coefficient, and mismatch traction normalization value, a carbon-labeled back-projection record is generated. Specifically, the plot number, remote sensing image number, pixel row number, pixel column number, aboveground biomass carbon label value, belowground biomass carbon label value, litter carbon label value, soil organic carbon label value, mismatch traction normalization value, biomass carbon label carrying capacity coefficient, and environmental constraints corresponding to the same dominant pixel are used to generate the carbon-labeled back-projection record. Carbon labeling continuity factors are written into the same carbon labeling back-projection record, ensuring that plot-scale carbon storage labels, after back-projection, can be bound to specific remote sensing image numbers and pixel locations. Specifically, the dominant pixel's aboveground biomass carbon label value, dominant pixel's belowground biomass carbon label value, dominant pixel's litter carbon label value, and dominant pixel's soil organic carbon label value are all uniformly converted to carbon storage values ​​per unit pixel area, with the unit of measurement being tons of carbon per pixel. This allows carbon labeling back-projection records from different plot areas to undergo conflict detection and model training at the same remote sensing pixel scale. When the same remote sensing image number, the same pixel row number, and the same pixel column number correspond to multiple... When recording carbon label back projections, the total back projection carbon label value for each carbon label back projection record is calculated separately. The range of the total back projection carbon label value is then divided by the mean of the corresponding plot's ground carbon storage label value to generate a label conflict value. The total back projection carbon label value is the sum of the dominant pixel's aboveground biomass carbon label value, dominant pixel's belowground biomass carbon label value, dominant pixel's litter carbon label value, and dominant pixel's soil organic carbon label value within the same carbon label back projection record. There is a partial summation relationship between the total back projection carbon label value and the plot's ground carbon storage label value; the plot's ground carbon storage label value is the sum of the aboveground biomass carbon storage value and the belowground biomass carbon storage value. The total back-projected carbon label value is the sum of litter carbon storage and soil organic carbon storage values. This total back-projected carbon label value is the pixel-level total carbon storage value after being allocated to the dominant pixels by the corresponding carbon label carrying capacity coefficients for the four types of sample plots. The range of the total back-projected carbon label value is the difference between the maximum and minimum total back-projected carbon label values ​​under the same remote sensing image number, pixel row number, and pixel column number. The mean of the ground carbon storage label values ​​for the corresponding sample plots is the arithmetic mean of the ground carbon storage label values ​​of all sample plots participating in the conflict judgment for the same pixel, under a unified pixel area caliber. When the mean of the ground carbon storage label values ​​for the corresponding sample plots is less than 0.05 tons of carbon per pixel, it will be 0.0.05 tons of carbon per pixel is used as a substitute denominator in the calculation of label conflict values ​​to avoid abnormally amplified label conflict values ​​due to a denominator of 0 or too small. Label conflict values ​​characterize the degree of label divergence when the same remote sensing pixel receives carbon labels from different back-projected plots. Back-projected carbon label records with label conflict values ​​greater than the label conflict threshold are written into a sample isolation queue, while those with label conflict values ​​less than or equal to the label conflict threshold are written into a training candidate record group. The label conflict threshold is read from the plot verification parameter table, ranging from 0.12 to 0.25, and is determined by the carbon storage label difference distribution of repeated plots under the same forest stand type. When the coefficient of variation of carbon storage label differences of repeated plots under the same forest stand type is less than 0.15, the label conflict threshold is 0.12 to 0.16; when the coefficient of variation is 0.15 to 0.30, the label conflict threshold is 0.16 to 0.21; and when the coefficient of variation is greater than 0.30, the label conflict threshold is 0.21 to 0.25; The sample isolation queue is used to temporarily store carbon-labeled backprojection records where the label divergence exceeds the allowable range, preventing conflicting labels from entering the forest carbon pool inversion model training process; When multiple carbon-labeled backprojection records exist in the training candidate record group corresponding to the same remote sensing image number, the same pixel row number, and the same pixel column number, the multiple carbon-labeled backprojection records are weighted and fused according to the mismatch traction normalization value to generate pixel fusion training samples; Specifically, the mismatch traction normalization values ​​of each carbon-labeled backprojection record in the training candidate record group are first normalized to generate fusion weight values, and then the dominant pixels are calculated separately. The weighted average of the aboveground biomass carbon label, the dominant pixel's belowground biomass carbon label, the dominant pixel's litter carbon label, and the dominant pixel's soil organic carbon label is used as the output label for the four types of models in the pixel fusion training samples. When the sum of all mismatched traction normalization values ​​within the training candidate record group is 0, the sum is normalized using the boundary coverage ratio of each carbon label backprojection record to generate the fusion weight value. When the sum of the boundary coverage ratio values ​​is still 0, the fusion weight value of each carbon label backprojection record within the training candidate record group is set to the same value, where the same value is 1 divided by the training candidate. Select the number of records within the record group; simultaneously retain the normalized vegetation index, enhanced vegetation index, canopy cover, texture response, slope, aspect, slope projection ratio, tree species code, and age group code for the records written into the carbon sink model training sample set, ensuring that each training sample simultaneously possesses remote sensing input features, topographic input features, stand attribute input features, dominant pixel carbon label value, and mismatch-driven normalization value; for the pixel fusion training samples generated by weighted fusion, write the plot number, mismatch-driven normalization value, and fusion weight value of each carbon label back-projection record within the training candidate record group into the source record. The recording field ensures that when the same pixel corresponds to multiple low-conflict carbon label back-projection records, it does not create training bias caused by the same input corresponding to multiple different labels. When the same pixel corresponds to only one carbon label back-projection record, the corresponding record, along with the corresponding mismatch traction normalization value, is directly written into the carbon sink model training sample set. Since there is no cross-plot label divergence when the same pixel corresponds to only one carbon label back-projection record, directly retaining the corresponding record maintains the effective number of training samples in the plot mismatch-dominated pixel set, and the mismatch traction normalization value participates in sample contribution control as a training weight when subsequently training the forest carbon pool inversion model.

[0040] In this implementation scheme, a label conflict value is introduced to constrain the consistency of carbon label back-projection records under the same remote sensing image number, the same pixel row number, and the same pixel column number. This ensures that the dominant pixel carbon label value undergoes conflict screening before entering the carbon sink model training sample set, reducing label discrepancies caused by different landform ground carbon storage label values ​​for the same pixel. At the same time, the carbon sink model training sample set retains mismatch traction normalization values, remote sensing input features, topographic input features, and forest stand attribute input features. This allows the subsequent forest carbon pool inversion model to control sample contributions using mismatch traction normalization values ​​after excluding high-conflict samples, thereby improving the spatial consistency, label reliability, and carbon sequestration prediction stability of the training samples.

[0041] Table 1. Back-projection data of carbon label-controlled pixels in sample plots.

[0042] This embodiment uses plot YD-01 as the object, with a plot area of ​​400m² and a plot center point plane coordinate of (123.6m, 84.2m). The plot is located in a terrain area with a slope of 18° and a slope aspect of 135°. By constructing a forest carbon sink plane coordinate system, the initial plot boundary is corrected according to the slope projection offset vector value to obtain the slope projection corrected plot boundary polygon. Then, based on the intersection relationship between the slope projection corrected plot boundary polygon and the remote sensing pixel boundary, R08C12, R08C13, R09C12, and R09C13 are determined as boundary coverage pixels. The mismatch traction normalization value is calculated by combining the boundary coverage ratio value, remote sensing difference normalization value, slope drag normalization value, and location diffusion normalization value, and a set of mismatch dominating pixels for the plot is selected. Subsequently, the aboveground biomass carbon storage value, underground biomass carbon storage value, litter carbon storage value, and soil organic carbon storage value are back-projected to each dominating pixel, and the carbon label values ​​of the four types of dominating pixels are summed to generate the total carbon label value of the corresponding dominating pixels.

[0043] Table 1 records the mismatch drag and carbon label back projection results for each boundary cover pixel in the mismatch dominant pixel set of the sample plots, where tC represents the unit of measurement as tons of carbon per pixel. As shown in Table 1, R08C12 has the highest boundary cover ratio of 0.38 among the four boundary cover pixels, but its mismatch drag normalization value is 0.240, and its total carbon label value for dominant pixels is 1.080 tC, which is not the highest. R09C12 has a boundary cover ratio of 0.22, lower than R08C12 and R08C13, but its mismatch drag normalization value is 0.303, its biomass carbon label carrying capacity is 0.358, and its total carbon label value for dominant pixels is 1.374 tC, all of which are relatively high. The value is the highest among the four pixels, indicating that the mismatch traction normalization value is not determined solely by the boundary cover ratio value, but is jointly determined by the boundary cover normalization value, remote sensing difference normalization value, slope drag normalization value, location diffusion normalization value, and center offset normalization value. The boundary cover ratio value of R09C13 is 0.15, the mismatch traction normalization value is 0.182, and the total carbon label value of the dominant pixel is 0.579tC, all of which are at a low level, indicating that this pixel has a small contribution to the back projection of carbon storage values ​​of the four types of sample plots.

[0044] like Figure 4 The figure illustrates the mismatch between sample plots and remote sensing pixels, as well as the carbon label back-projection process in this invention. Using the forest carbon sink plane coordinate system as a reference, the target forest area is divided into multiple remote sensing pixel grids. The dashed box represents the initial sample plot boundary formed based on the plane coordinate values ​​of the sample plot boundary vertices, while the solid box represents the corrected sample plot boundary based on the slope value, aspect value, and slope projection ratio. The sample plot center point is located near R08C12, but the corrected sample plot boundary simultaneously covers four boundary-covering pixels: R08C12, R08C13, R09C12, and R09C13. In the figure, arrows pointing from the center point of the sample plot to each covered cell represent mismatch traction relationships. The thicker the arrow and the cell boundary line, the higher the mismatch traction normalization value of the corresponding cell. In the cell label, T represents the mismatch traction normalization value, and C represents the total carbon label value of the dominant cell, which is obtained by summing the aboveground biomass carbon label value, the belowground biomass carbon label value, the litter carbon label value, and the soil organic carbon label value of the dominant cell. As can be seen from the figure, this embodiment reconstructs the sample plot-cell correspondence under complex terrain by using sample plot cell mismatch evidence primitives, mismatch traction maps, and carbon label back projection records, rather than simply allocating carbon storage labels according to the center point or area ratio of the sample plot.

[0045] Optionally, using the mismatch-driven normalized value as training weight, the specific steps for training the forest carbon pool inversion model based on the carbon sink model training sample set are as follows: Read the carbon sink model training sample set, extract each training sample record according to the plot number, remote sensing image number, pixel row number, and pixel column number, and simultaneously read the corresponding mismatch-driven normalized value, aboveground biomass carbon label value of the dominant pixel, belowground biomass carbon label value of the dominant pixel, litter carbon label value of the dominant pixel, and soil organic carbon label value of the dominant pixel; where the mismatch-driven normalized value is not used to indicate that the sample has no mismatch risk, but rather to indicate the carbon label inheritance contribution intensity of the dominant pixel after completing the selection of the mismatch dominant pixel set in the plot, the generation of carbon label back projection records, and the isolation of label conflicts. A larger normalized value indicates a higher contribution of the corresponding dominant pixel to the carbon storage label of the sample plot, and a higher contribution to the training samples of the forest carbon pool inversion model. The normalized vegetation index (NVI), enhanced vegetation index (VEI), canopy cover, texture response, slope, aspect, slope projection ratio, tree species coding, and age group coding are used as model inputs. Specifically, the NVI and VEI are used to characterize the pixel's spectral response to vegetation, canopy cover to characterize the degree of canopy coverage, texture response to characterize the spatial structure differences of the forest stand, slope, aspect, and slope projection ratio to characterize the topographic constraints of the pixel, and tree species coding and age group coding to characterize the forest stand attributes of the pixel. These values ​​are used before inputting into the forest carbon pool inversion model. The normalized vegetation index, enhanced vegetation index, canopy cover, texture response, slope, aspect, and slope projection ratio are subjected to max-min normalization. Tree species and age group codes are converted to one-hot codes, ensuring that continuous and categorical input features are recorded in the same training sample. When the maximum and minimum values ​​of continuous input features are equal, the normalization result for the corresponding continuous input feature is recorded as 0 to avoid a normalization denominator of 0. The dominant pixel aboveground biomass carbon label, dominant pixel belowground biomass carbon label, dominant pixel litter carbon label, and dominant pixel soil organic carbon label are used as model output labels, and the corresponding mismatch traction normalization values ​​are used as training weights to train the forest carbon pool inversion model. Specifically, The forest carbon pool inversion model employs a multi-output gradient boosting regression tree that supports sample weight input. The model includes a shared input feature set and four carbon pool output branches. These four branches output predicted values ​​for aboveground biomass carbon, belowground biomass carbon, litter carbon, and soil organic carbon, respectively. The four carbon pool output branches use the same training sample input features and the same sample weights, and calculate the prediction error for the corresponding carbon pool label. Before training, the mismatch traction normalization values ​​of all training samples in the same training batch are summed and normalized so that the sum of all training sample weights is 1. When the sum of all mismatch traction normalization values ​​in the training batch is 0, the weights of all training samples are set to the same value, which is 1 divided by the number of training samples.The weighted loss value for a single training sample record is the training sample weight multiplied by the sum of squared errors of the four carbon label predictions. These four carbon label prediction errors are the differences between the predicted aboveground biomass carbon value and the aboveground biomass carbon label value of the dominant pixel, the predicted belowground biomass carbon value and the belowground biomass carbon label value of the dominant pixel, the predicted litter carbon value and the litter carbon label value of the dominant pixel, and the predicted soil organic carbon value and the soil organic carbon label value of the dominant pixel. The training objective of the forest carbon pool inversion model is to minimize the sum of the weighted loss values ​​of all training sample records, ensuring that training samples with higher mismatch traction normalization values ​​contribute more to tree split node selection and leaf node error calculation, while reducing the error contribution of training samples with lower mismatch traction normalization values ​​during training. After each iteration of the tree model, the predicted values ​​generated by the four carbon pool output branches are restricted to be no less than 0, ensuring that the predicted aboveground biomass carbon, belowground biomass carbon, litter carbon, and soil organic carbon values ​​conform to carbon storage. Non-negativity constraints are implemented. During training, the carbon sink model training sample set is divided into training subsets and validation subsets according to the plot number, ensuring that training samples corresponding to the same plot number do not simultaneously enter the training subset and validation subset. Training stops when the weighted loss value of the validation subset decreases by less than 0.001 over 10 consecutive iterations, and stops after 300 training rounds. Training ends when either of the two stopping conditions is met first. After training, the following parameters are saved: the feature input order, continuous input feature normalization parameters, tree species coding values ​​and age group coding values' unique thermal coding mapping table, four carbon label output order, mismatch traction normalization value weighted training parameters, non-negativity output constraint parameters, and training termination conditions. This ensures that subsequent target pixel carbon sink model input records can generate predicted values ​​for aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, and soil organic carbon storage according to the same input order and feature processing rules.

[0046] In this implementation plan, by introducing mismatched traction normalization values ​​into the training of the forest carbon pool inversion model, the sample contributions of different dominant pixels in the carbon sink model training sample set are consistent with the intensity of the dominant pixels' contribution to the carbon label of the sample plot. This reduces the interference of low-traction samples on the aboveground biomass carbon label value, belowground biomass carbon label value, litter carbon label value, and soil organic carbon label value of the dominant pixels. At the same time, the normalized vegetation index value, enhanced vegetation index value, canopy cover value, texture response value, slope value, aspect value, slope projection ratio value, tree species coding value, and age group coding value are trained in a fixed input order. This allows the subsequent inference model to output the predicted carbon pool values ​​of each target pixel using the same feature caliber, improving the stability and verifiability of the forest carbon sequestration prediction results in complex terrain and mixed sample plot boundary areas.

[0047] Optionally, the specific steps for generating total carbon storage prediction values ​​and carbon sequestration prediction values ​​for target pixels and outputting forest carbon sequestration prediction results are as follows: Read the remote sensing image data, topographic grid data, and stand attribute layer data of the target predicted forest area. The stand attribute layer data includes stand sub-compartment boundary polygons, tree species coding values, and age group coding values. Based on the pixel row number, pixel column number, and the overlap relationship between the target pixel and the stand sub-compartment boundary polygons, extract the normalized vegetation index, enhanced vegetation index, canopy cover value, texture response value, slope value, aspect value, slope projection ratio value, tree species coding value, and age group coding value of the target pixel to generate the target pixel carbon sink model input record. Specifically, remote sensing... Normalized Difference Vegetation Index (NDV), Enhanced Vegetation Index (VEA), Canopy Cover, and Texture Response values ​​in image data are used to characterize the vegetation spectral response, canopy cover status, and stand texture structure of target pixels. Slope, aspect, and slope projection scale values ​​in topographic grid data are used to characterize the topographic constraints of target pixels. Tree species and age group codes in stand attribute layer data are used to characterize the stand type corresponding to the target pixel. When a target pixel overlaps with the boundary polygon of a stand sub-compartment, the overlap area between the target pixel's boundary polygon and the boundary polygons of each stand sub-compartment is calculated. The tree species and age group codes corresponding to the stand sub-compartment with the largest overlap area are selected as the tree species and age group codes for the target pixel. Code values; when multiple forest stand sub-compartments have the same overlap area value with the target pixel, the tree species code value and age group code value corresponding to the forest stand sub-compartment with the smaller forest stand sub-compartment number are selected first to ensure that the forest stand attribute value of the target pixel is unique; when generating the target pixel carbon sink model input record, the normalized vegetation index value, enhanced vegetation index value, canopy cover value, texture response value, slope value, aspect value, slope projection ratio value, tree species code value, and age group code value are arranged according to the feature input order saved in the training stage of the forest carbon pool inversion model, and the normalization parameters and coding rules consistent with the training stage are used to process continuous input features and category input features, so that the target pixel carbon sink model input record and the carbon sink model training sample are consistent. The data sets maintain the same caliber; the normalized continuous input features and the uniquely thermally encoded category input features are used as inputs for model calculations, without changing the unit of measurement for subsequent carbon storage predictions; the target pixel carbon sink model input records are input into the forest carbon pool inversion model to generate predicted values ​​for aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, and soil organic carbon storage of the target pixel, and these four carbon storage prediction values ​​are summed to generate the total carbon storage prediction value for the target pixel; all four carbon storage prediction values ​​correspond to the same target pixel boundary range, the same target prediction time point, and the same unit of measurement, with the unit of measurement being uniformly ton carbon per pixel;The forest carbon pool inversion model outputs predicted values ​​for aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, and soil organic carbon storage of target pixels in the order of the four carbon labels generated during the training phase. This ensures that the predicted total carbon storage of target pixels retains the sources of each carbon pool and supports subsequent item-by-item verification. The model reads the baseline total carbon storage value of the target pixel during the baseline measurement period, subtracts the baseline total carbon storage value from the predicted total carbon storage value to generate the predicted carbon sequestration value for the target pixel, and then... The forest carbon sequestration prediction results are generated using remote sensing image numbers, pixel row numbers, pixel column numbers, predicted total carbon storage of target pixels, and predicted carbon sequestration of target pixels. Specifically, the baseline measurement period is the historical comparison measurement period prior to the target prediction date, with a length of one year. The start date of the baseline measurement period is the date corresponding to twelve months prior to the target prediction date, and the end date is the date corresponding to one day prior to the target prediction date. When forestry carbon sink accounting requires the use of a fixed annual caliber, the baseline measurement period adopts the previous year of the target prediction. The total carbon storage baseline value is calculated from January 1st to December 31st of a calendar year. It is derived from historical carbon storage results corresponding to the same target forest area, the same remote sensing image number, the same row number, and the same column number within the baseline measurement period. The pixel boundary range, pixel spatial resolution, and unit of measurement in the historical carbon storage results are consistent with the predicted total carbon storage value for the target pixel. When there is a shift between the remote sensing image grid of the baseline measurement period and the remote sensing image grid of the target prediction time, the baseline value of total carbon storage for the baseline measurement period is adjusted according to the image grid of the target prediction time. The boundary range of the target pixel is resampled and area-weighted to ensure that the baseline total carbon storage value and the predicted total carbon storage value of the target pixel correspond to the same pixel range. The predicted carbon sequestration value of the target pixel retains its positive or negative sign to represent the increase or decrease in carbon storage in the current measurement period relative to the baseline measurement period, with the unit of measurement being tons of carbon per pixel. The forest carbon sequestration prediction results are indexed according to the remote sensing image number, pixel row number, and pixel column number of the target pixel, so that each target pixel can correspond to the predicted total carbon storage value and the predicted carbon sequestration value of the target pixel.

[0048] In this implementation plan, by ensuring that the data caliber of the input records of the target pixel carbon sequestration model is consistent with that of the carbon sequestration model training sample set, and by uniformly summarizing the predicted values ​​of aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, and soil organic carbon storage of the target pixel into a total predicted value of carbon storage for the target pixel, the forest carbon sequestration prediction results can reflect the pixel-level carbon storage status in the current measurement period, and can also express the carbon storage change relative to the baseline measurement period through the predicted values ​​of carbon sequestration of the target pixels. This reduces the prediction bias caused by the inconsistency between the training input feature caliber and the target prediction input feature caliber, and improves the traceability, itemized verification, and measurement stability of the forest carbon sequestration prediction results under different remote sensing image numbers, pixel row numbers, and pixel column numbers.

[0049] like Figure 2 As shown, another aspect of the present invention provides a forest carbon sequestration prediction system based on a carbon sink model. This system is applied to a forest carbon sequestration prediction method based on a carbon sink model. The system includes: The sample plot mismatch primitive generation module is used to collect forest sample plot survey data, remote sensing image data and topographic grid data, construct a forest carbon sink plane coordinate system, generate slope projection corrected sample plot boundary polygons, and generate sample plot pixel mismatch evidence primitives by combining center offset distance value, boundary coverage ratio value and remote sensing response difference value. The mismatch traction map construction module is used to construct a mismatch traction map based on the sample plot pixel mismatch evidence primitives, calculate the center offset normalization value, boundary coverage normalization value, slope drag normalization value, location diffusion normalization value and remote sensing difference normalization value, generate mismatch traction normalization value, and filter the sample plot mismatch dominant pixel set. The carbon label back projection reconstruction module is used to back project the values ​​of aboveground biomass, belowground biomass, litter and soil organic carbon storage into the corresponding carbon label values ​​of the dominant pixels based on the mismatch of the dominant pixel set of the sample plot, and generate carbon label back projection records and carbon sink model training sample set. The carbon sequestration prediction result write-back module is used to train the forest carbon pool inversion model based on the carbon sink model training sample set, using mismatch traction normalization values ​​as training weights, to generate total carbon storage prediction values ​​and carbon sequestration prediction values ​​for target pixels, and output forest carbon sequestration prediction results.

[0050] In this implementation plan, by forming a continuous data chain from sample plot pixel mismatch evidence primitives, mismatch-driven normalized values, sample plot mismatch-dominant pixel sets, carbon label back-projection records, and carbon sink model training sample sets, the aboveground biomass carbon storage values, belowground biomass carbon storage values, litter carbon storage values, and soil organic carbon storage values ​​can be converted from sample plot scale to pixel-level training data with spatial matching basis and carbon pool component meaning, reducing the mismatch impact caused by individually binding labels to the center-falling pixel numbers. At the same time, the mismatch-driven normalized values ​​constrain the training of the forest carbon pool inversion model, enabling the inference model to incorporate sample plot-pixel mismatch correction results when generating target pixel total carbon storage prediction values ​​and target pixel carbon sequestration prediction values, thereby improving the stability, traceability, and verification consistency of forest carbon sequestration prediction results in complex terrain and mixed forest edge areas.

[0051] The following points need to be explained: (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0052] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0053] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0054] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting forest carbon sequestration based on a carbon sink model, characterized in that, The method includes: S1. Collect forest plot survey data, remote sensing image data and topographic grid data, construct a forest carbon sink plane coordinate system, generate slope projection correction plot boundary polygons, and generate plot pixel mismatch evidence primitives by combining center offset distance value, boundary coverage ratio value and remote sensing response difference value. S2, construct mismatch traction map based on sample plot pixel mismatch evidence primitives, calculate center offset normalization value, boundary coverage normalization value, slope drag normalization value, location diffusion normalization value and remote sensing difference normalization value, generate mismatch traction normalization value, and filter sample plot mismatch dominant pixel set; S3, based on the mismatched dominant pixel set of the sample plot, back-projects the values ​​of aboveground biomass, belowground biomass, litter and soil organic carbon storage into the corresponding dominant pixel carbon label values, and generates carbon label back-projection records and carbon sink model training sample set. S4 uses the mismatch traction normalization value as the training weight, trains the forest carbon pool inversion model based on the carbon sink model training sample set, generates the total carbon storage prediction value and carbon sequestration prediction value for the target pixel, and outputs the forest carbon sequestration prediction result.

2. The forest carbon sequestration prediction method based on a carbon sink model according to claim 1, characterized in that, The specific steps for collecting forest plot survey data, remote sensing image data, and topographic grid data, constructing a forest carbon sink plane coordinate system, and generating slope projection-corrected plot boundary polygons are as follows: Forest plot survey data, remote sensing image data, and topographic grid data were collected. Forest plot survey data included plot number, plot center point coordinates, plot boundary vertex coordinates, plot positioning accuracy level markers, tree species codes, age group codes, aboveground biomass carbon storage, belowground biomass carbon storage, litter carbon storage, soil organic carbon storage, and plot ground carbon storage label values. Remote sensing image data included remote sensing image number, pixel row number, pixel column number, pixel center coordinates, pixel spatial resolution, normalized difference vegetation index (NDVI), enhanced vegetation index (VEI), canopy cover, and texture response. Topographic grid data included grid center coordinates, slope, aspect, and slope projection scale. Using the southwest corner of the target predicted forest area's outer rectangle as the origin, and the east direction as the positive X-axis direction, and the north direction as the positive Y-axis direction, a forest carbon sink plane coordinate system is constructed. The coordinate values ​​of the sample plot center point, sample plot boundary vertex, pixel center, and grid center are transformed into the forest carbon sink plane coordinate system to generate the corresponding plane coordinate values. The initial sample plot boundary polygon is generated by closing and connecting the plane coordinates of the vertex of the sample plot boundary in the order of collection, and the positioning error radius value is read according to the sample plot positioning accuracy level mark. Read the slope, aspect, and slope projection ratio values ​​within the coverage area of ​​the initial sample plot boundary polygon, and calculate the average slope, average aspect, and average slope projection ratio. The direction of the slope projection offset vector value is determined based on the average slope aspect value. The length of the slope projection offset vector value is determined based on the product of the square root of the area value of the initial plot boundary polygon, the tangent of the average slope value, and the average slope projection ratio value. The initial plot boundary polygon is then translated along the slope projection offset vector value to generate the slope projection corrected plot boundary polygon.

3. The forest carbon sequestration prediction method based on a carbon sink model according to claim 2, characterized in that, The specific steps for generating plot pixel mismatch evidence primitives by combining center offset distance values, boundary coverage ratio values, and remote sensing response difference values ​​are as follows: The pixel boundary polygon is generated based on the pixel center plane coordinates and pixel spatial resolution. The number of the center-falling pixel is determined based on the sample plot center point plane coordinates. The center offset distance value is generated based on the distance between the sample plot center point plane coordinates and the center plane coordinates of the center-falling pixel. Filter out the boundary polygons of the plots that intersect with the boundary polygons of the slope projection correction plots, and generate a set of boundary-covered pixels; generate a boundary coverage ratio value according to the ratio of the intersection area value to the area value of the boundary polygons of the slope projection correction plots; generate a remote sensing response difference value according to the mean of the absolute values ​​of the differences between the boundary-covered pixels and the center-falling pixels in the normalized vegetation index, enhanced vegetation index, canopy cover and texture response values. The sample plot pixel mismatch evidence primitive is generated by combining the center-falling pixel number, boundary-covered pixel set, positioning error radius value, slope projection offset vector value, center offset distance value, boundary coverage ratio value, and remote sensing response difference value.

4. The forest carbon sequestration prediction method based on a carbon sink model according to claim 3, characterized in that, The specific steps for constructing a mismatch traction map based on sample plot pixel mismatch evidence primitives and calculating the center offset normalization value, boundary cover normalization value, slope drag normalization value, location diffusion normalization value, and remote sensing difference normalization value are as follows: Read the sample plot pixel mismatch evidence primitives, construct the sample plot center point, slope projection corrected sample plot boundary polygon, center falling pixel and boundary covering pixel as mismatch traction nodes, and establish center offset traction edge with center offset distance value, boundary covering traction edge with boundary covering ratio value, slope aspect drag traction edge with slope projection offset vector value and positioning diffusion traction edge with positioning error radius value. Associate the mismatch traction nodes and each traction edge according to the same sample plot number to generate mismatch traction map; The center offset distance, slope projection offset vector length, and positioning error radius are divided by the pixel spatial resolution to generate the center offset normalization value, slope drag normalization value, and positioning diffusion normalization value. The remote sensing response difference value is divided by the interquartile range of remote sensing response difference values ​​within the same remote sensing image to generate the remote sensing difference normalization value, and the boundary coverage ratio value is used as the boundary coverage normalization value.

5. The method for predicting forest carbon sequestration based on a carbon sink model according to claim 4, characterized in that, The specific steps for generating mismatch traction normalization values ​​and filtering the set of mismatch-dominant pixels in the sample plots are as follows: Calculate the mismatch drag value for each boundary cover cell. The numerator of the mismatch drag value is the product of the boundary cover normalization value and the remote sensing difference normalization value plus the product of the slope drag normalization value and the positioning diffusion normalization value. The denominator of the mismatch drag value is the product of the center offset normalization value and the center fall suppression coefficient. Normalize the mismatch traction values ​​under the same land area number to generate mismatch traction normalized values. Accumulate boundary cover cells in descending order of mismatch traction normalized values. Write the boundary cover cells corresponding to the accumulated mismatch traction normalized values ​​that reach the mismatch dominance retention threshold into the sample land mismatch dominance cell set.

6. The method for predicting forest carbon sequestration based on a carbon sink model according to claim 5, characterized in that, The specific steps for back-projecting aboveground biomass, belowground biomass, litter, and soil organic carbon storage values ​​into corresponding dominant pixel carbon label values ​​based on the mismatched dominant pixel set of the sample plot are as follows: Read the set of mismatched dominant pixels in the sample plot, the aboveground biomass carbon storage value, the underground biomass carbon storage value, the litter carbon storage value, and the soil organic carbon storage value. Write the aboveground biomass carbon storage value and the underground biomass carbon storage value into the biomass carbon pool label vector, and write the litter carbon storage value and the soil organic carbon storage value into the environmental constraint carbon pool label vector. Read the mismatch traction normalization value, boundary coverage ratio value and remote sensing difference normalization value of each dominant cell under the same land number, calculate the product of the mismatch traction normalization value, boundary coverage ratio value and remote sensing difference normalization value, generate the initial value of biomass carbon inheritance of dominant cell, and normalize it according to the sum of the initial values ​​under the same land number to generate the biomass carbon label inheritance coefficient. Based on the pixel center plane coordinates of the dominant pixels, the slope and aspect values ​​of the dominant pixels are read from the topographic grid data to generate the slope and aspect classification values ​​of the dominant pixels. Using the tree species code, age group code, slope and aspect classification values ​​of the dominant pixels corresponding to the plot number as lookup indexes, the environmental constraint benchmark coefficients are read from the environmental constraint coefficient table. The environmental constraint benchmark coefficients are multiplied by the boundary coverage ratio to generate the environmental constraint coefficients of the dominant pixels. The environmental constraint coefficients of all dominant pixels under the same plot number are normalized to generate the environmental constraint carbon labeling acceptance coefficient. The biomass carbon pool label vector is multiplied by the biomass carbon label carrying capacity coefficient of each dominant cell to generate the aboveground biomass carbon label value and the belowground biomass carbon label value of the dominant cell; the environmental constraint carbon pool label vector is multiplied by the environmental constraint carbon label carrying capacity coefficient of each dominant cell to generate the litter carbon label value and the soil organic carbon label value of the dominant cell.

7. The method for predicting forest carbon sequestration based on a carbon sink model according to claim 6, characterized in that, The specific steps for generating carbon-labeled back-projection records and carbon sink model training sample sets are as follows: Based on the plot number, remote sensing image number, pixel row number, and pixel column number, combined with the corresponding dominant pixel carbon label value, carbon label acceptance coefficient, and mismatch traction normalization value, a carbon label back projection record is generated. When multiple carbon-labeled backprojection records correspond to the same remote sensing image number, the same pixel row number, and the same pixel column number, the total backprojection carbon label value of each carbon-labeled backprojection record is calculated separately. The range of the total backprojection carbon label value is divided by the mean of the corresponding sample plot ground carbon storage label value to generate a label conflict value. Carbon-labeled backprojection records with label conflict values ​​greater than the label conflict threshold are written into the sample isolation queue, and the remaining records, along with the corresponding mismatch traction normalization value, are written into the carbon sink model training sample set. When the same pixel corresponds to only one carbon-labeled backprojection record, the corresponding record, along with the corresponding mismatch traction normalization value, is written into the carbon sink model training sample set.

8. The method for predicting forest carbon sequestration based on a carbon sink model according to claim 7, characterized in that, The specific steps for training the forest carbon pool inversion model based on the carbon sink model training sample set, using mismatch traction normalization values ​​as training weights, are as follows: The carbon sink model training sample set was read, and the normalized vegetation index, enhanced vegetation index, canopy coverage, texture response, slope, aspect, slope projection ratio, tree species coding, and age group coding were used as model inputs. The dominant pixel aboveground biomass carbon label, dominant pixel belowground biomass carbon label, dominant pixel litter carbon label, and dominant pixel soil organic carbon label were used as model output labels. The corresponding mismatch traction normalization value was used as training weights to train the forest carbon pool inversion model.

9. The method for predicting forest carbon sequestration based on a carbon sink model according to claim 8, characterized in that, The specific steps for generating predicted total carbon storage and carbon sequestration values ​​for target pixels and outputting forest carbon sequestration prediction results are as follows: The system reads remote sensing image data, topographic grid data, and stand attribute layer data of the target predicted forest area. The stand attribute layer data includes stand sub-compartment boundary polygons, tree species coding values, and age group coding values. Based on the pixel row number, pixel column number, and the overlap relationship between the target pixel and the stand sub-compartment boundary polygons, the system extracts the normalized vegetation index, enhanced vegetation index, canopy cover value, texture response value, slope value, aspect value, slope projection ratio value, tree species coding value, and age group coding value of the target pixel to generate the target pixel carbon sink model input record. Input the target pixel carbon sink model input record into the forest carbon pool inversion model to generate the target pixel aboveground biomass carbon storage prediction value, target pixel underground biomass carbon storage prediction value, target pixel litter carbon storage prediction value and target pixel soil organic carbon storage prediction value, and add the four carbon storage prediction values ​​to generate the target pixel total carbon storage prediction value. The baseline value of total carbon storage for the target pixel in the baseline measurement period is read. The predicted value of total carbon storage for the target pixel is subtracted from the baseline value of total carbon storage to generate the predicted value of carbon sequestration for the target pixel. The forest carbon sequestration prediction result is generated according to the remote sensing image number, pixel row number, pixel column number, predicted value of total carbon storage for the target pixel, and predicted value of carbon sequestration for the target pixel.

10. A forest carbon sequestration prediction system based on a carbon sink model, employing the forest carbon sequestration prediction method based on a carbon sink model as described in any one of claims 1-9, characterized in that, include: The sample plot mismatch primitive generation module is used to collect forest sample plot survey data, remote sensing image data and topographic grid data, construct a forest carbon sink plane coordinate system, generate slope projection corrected sample plot boundary polygons, and generate sample plot pixel mismatch evidence primitives by combining center offset distance value, boundary coverage ratio value and remote sensing response difference value. The mismatch traction map construction module is used to construct a mismatch traction map based on the sample plot pixel mismatch evidence primitives, calculate the center offset normalization value, boundary coverage normalization value, slope drag normalization value, location diffusion normalization value and remote sensing difference normalization value, generate mismatch traction normalization value, and filter the sample plot mismatch dominant pixel set. The carbon label back projection reconstruction module is used to back project the values ​​of aboveground biomass, belowground biomass, litter and soil organic carbon storage into the corresponding carbon label values ​​of the dominant pixels based on the mismatch of the dominant pixel set of the sample plot, and generate carbon label back projection records and carbon sink model training sample set. The carbon sequestration prediction result write-back module is used to train the forest carbon pool inversion model based on the carbon sink model training sample set, using mismatch traction normalization values ​​as training weights, to generate total carbon storage prediction values ​​and carbon sequestration prediction values ​​for target pixels, and output forest carbon sequestration prediction results.

Images