A carbon accounting method based on carbon layer partitioning

By using multispectral remote sensing image processing and feature matching technology, the problem of carbon layer delineation relying on human experience has been solved, and intelligent connection between carbon layers and carbon accounting models has been achieved, improving the accuracy and efficiency of carbon layer delineation and adapting to rapid assessment of different forest types.

CN122841981BActive Publication Date: 2026-11-17HUAJUN TECHNOLOGY (CHONGQING) CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611232743.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17
Estimated Expiration
2046-08-14

AI Technical Summary

Technical Problem

Existing carbon layer partitioning methods rely on human experience, resulting in insufficient accuracy and weak generalization ability. Furthermore, carbon layer partitioning is disconnected from the selection of carbon accounting models, leading to an overall workflow that cannot be automated in a closed loop and is inefficient.

Method used

By acquiring multispectral remote sensing image data, preprocessing it, and outputting the reflectance tensor, multidimensional feature vectors are calculated pixel by pixel. The feature matrix of the whole image is fused, principal component analysis is performed to reduce the dimensionality, the optimal number of clusters is determined, feature matrix clustering is performed, carbon layer label maps are generated, and feature matching is performed with the feature fingerprint library of pre-stored carbon accounting models to bind the optimal model to each carbon layer.

Benefits of technology

It has realized a fully automated process from multispectral remote sensing imagery to carbon layer delineation and model matching, which has improved the accuracy and adaptability of carbon layer delineation, reduced manual intervention, shortened the cycle, and met the needs of rapid assessment of large-scale forest carbon sinks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122841981B_ABST
    Figure CN122841981B_ABST
Patent Text Reader

Abstract

This invention discloses a carbon accounting method based on carbon layer partitioning, belonging to the field of carbon layer partitioning technology. The method includes: acquiring multispectral image data of a target forest area, and outputting a reflectance tensor of dimension H×W×4 after preprocessing; obtaining multidimensional feature vectors from the reflectance tensor based on the reflectance values ​​of four bands, resulting in a full-image feature matrix of dimension H×W×N; performing dimensionality reduction on the full-image feature matrix using principal component analysis based on the cumulative variance contribution rate, outputting a dimensionality-reduced feature matrix H×W×M; determining the optimal number of clusters K based on the dimensionality-reduced feature matrix, performing feature matrix clustering, obtaining cluster center coordinates, and generating a carbon layer label map; extracting the multidimensional feature vectors before dimensionality reduction from each carbon layer in the carbon layer label map, obtaining the feature mean vector, and performing feature matching with a pre-stored feature fingerprint database of carbon accounting models to bind a carbon accounting model to each carbon layer.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of carbon layer partitioning technology, specifically relating to a carbon accounting method based on carbon layer partitioning. Background Technology

[0002] Forests are the largest carbon sink in terrestrial ecosystems, playing a crucial role in the carbon cycle. Forestry carbon sequestration, as an important means of addressing climate change, absorbs atmospheric carbon dioxide and fixes it in vegetation and soil through forestry activities such as afforestation and forest management, and has become an important asset in the carbon trading market. In the development of forestry carbon sequestration projects, carbon stratification—dividing the project area into several relatively homogeneous carbon stratification units based on factors such as afforestation time, tree species composition, forest age structure, and site conditions—is a fundamental step in accurately calculating carbon storage and directly affects the accuracy of carbon sequestration calculation results.

[0003] Carbon stratification and the selection of carbon accounting models are two core tasks in forest carbon sink measurement and monitoring, and there is a close relationship between the two: carbon stratification determines spatial partitioning, and each carbon stratum needs to be matched with an appropriate carbon accounting model in order to complete the accurate estimation of carbon storage.

[0004] Forestry personnel pre-divide areas based on topographic maps and experience, measure tree diameter at breast height (DBH) and tree height in typical sample plots, calculate carbon storage in the sample plots using empirical formulas, and then extrapolate to the entire region. This approach has inherent flaws: the delineation results heavily rely on personnel experience, are highly subjective, have poor repeatability, and show significant differences in delineation results between different teams within the same area; manual surveys are time-consuming, typically requiring several months to complete surveys of areas covering hundreds of square kilometers, resulting in poor data timeliness and difficulty in responding promptly to forest dynamics; and when extrapolating sample plot data to the region, it is difficult to characterize forest spatial heterogeneity, and estimation errors are easily amplified.

[0005] Therefore, vegetation index maps are generated using satellite or UAV remote sensing data, and carbon layers are distinguished by setting thresholds for segmentation. For example, Chinese patent CN122366879A discloses a carbon sink calculation method based on multidimensional carbon layer segmentation. This method acquires vector data of ecological protection and restoration projects in the area to be calculated and performs carbon layer segmentation for different land categories based on different segmentation factors. Chinese patent CN116433748B performs carbon layer segmentation through grid division and canopy closure stratification. In addition, Chinese patent CN120298904A uses machine learning algorithms to calculate environmental variables based on reflectance data, surface temperature data, and elevation data of each pixel in remote sensing monitoring images.

[0006] While it has improved work efficiency to some extent, there are still obvious shortcomings: it is difficult to fully depict the essential differences in carbon storage among different vegetation types; in areas with complex terrain and mixed pixels, the classification accuracy is insufficient and the adaptability is poor; the classification results only provide spatial boundaries, and it is still necessary to rely on human experience to match the corresponding calculation formulas for different regions.

[0007] Therefore, designing a carbon layer partitioning and carbon accounting method that is highly automated, has reliable partitioning accuracy, and can intelligently adapt to carbon accounting models is a technical problem that urgently needs to be solved. Summary of the Invention

[0008] This invention provides a carbon accounting method based on carbon layer partitioning, realizing a closed-loop process from remote sensing image input to carbon layer partitioning and model matching. It achieves intelligent and seamless connection between carbon layer partitioning adjustment and carbon layer and carbon accounting model, thus solving the problems of insufficient partitioning accuracy and weak generalization ability caused by carbon layer partitioning relying on human experience, and the disconnect between carbon layer partitioning and carbon accounting model selection, which still requires manual intervention and results in low efficiency and an inability to achieve closed-loop automation of the overall workflow.

[0009] The technical solution adopted in this invention is as follows: A carbon accounting method based on carbon layer partitioning includes: Acquire multispectral image data of the target forest area, including at least blue, green, red, and near-infrared bands. After preprocessing, output a reflectance tensor with dimensions H×W×4, where H is the number of pixels in the image height direction, W is the number of pixels in the image width direction, and 4 corresponds to the reflectance values ​​of the four bands: blue, green, red, and near-infrared. The reflectance tensor is calculated pixel by pixel. Based on the reflectance values ​​of the four bands, a multidimensional feature vector characterizing the vegetation characteristics is obtained. The multidimensional feature vector is fused to obtain a full-image feature matrix of H×W×N, where N is the number of dimensions of the multidimensional feature vector. For the full-image feature matrix, principal component analysis is used to reduce the feature dimension based on the cumulative variance contribution rate, and the reduced feature matrix H×W×M is output, where M is the number of dimensions after feature dimensionality reduction, 1<M<N. Based on the reduced-dimensional feature matrix, the optimal number of clusters K is determined, feature matrix clustering is performed, the coordinates of the cluster centers are obtained, and a carbon layer label map is generated, where 3≤K≤8. Extract the multidimensional feature vectors of each carbon layer in the carbon layer label graph before dimensionality reduction to obtain the feature mean vector, and perform feature matching with the feature fingerprint library of the pre-stored carbon accounting model to bind the carbon accounting model to each carbon layer.

[0010] Wherein, the dimension N of the multidimensional feature vector is 8, including: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Texture Contrast Based on Gray-Level Co-occurrence Matrix (GLCM), Red Edge Position, Blue Band Reflectance, Near-Infrared Band to Red Band Reflectance Ratio, Neighborhood Standard Deviation of NDVI, and Difference Between EVI and NDVI.

[0011] The process of obtaining the multidimensional feature vector specifically includes: The reflectivity tensor is input as 4-channel input data into a convolutional neural network, which includes three convolutional layers and a global average pooling layer. The first convolutional layer yields local spectral features, the second convolutional layer yields combined low-level features, the third convolutional layer yields high-level abstract features, and the multi-dimensional feature vector is output after passing through a global average pooling layer. The number of convolution kernels in the third convolutional layer is greater than the number of convolution kernels in the second convolutional layer, and the number of convolution kernels in the second convolutional layer is greater than the number of convolution kernels in the first convolutional layer.

[0012] The process of obtaining the multidimensional feature vector specifically includes: The process is performed pixel by pixel, and the multidimensional feature vector is obtained based on the relationship between the feature vector of each dimension and the reflectivity tensor.

[0013] For the full-image feature matrix, principal component analysis is used to perform feature dimensionality reduction based on the cumulative variance contribution rate, including: The full-image feature matrix is ​​centered by subtracting the global mean of the corresponding dimension from each feature dimension; Calculate the covariance matrix between each feature dimension; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and corresponding eigenvectors. The variance contribution rates are accumulated sequentially from largest to smallest eigenvalues, and the top M principal components whose cumulative variance contribution rates reach a preset threshold are selected. The original multi-dimensional features are mapped to M-dimensional features by constructing a projection matrix using the feature vectors corresponding to the selected M principal components, and the reduced feature matrix is ​​output. Where M takes the value 3.

[0014] Based on the dimensionality-reduced feature matrix, the optimal number of clusters K is determined, including: Traverse the range of cluster numbers [3, 8]; For each cluster value, multiple cluster centers are randomly initialized, and clustering is performed separately for each cluster. Calculate the average silhouette coefficient for each clustering result to obtain the silhouette coefficient curve; If the slope change of adjacent points is greater than or equal to the slope threshold, then the profile coefficient curve is determined to have an inflection point. The point corresponding to the maximum slope change of adjacent points is taken as the inflection point, and the clustering value corresponding to the inflection point of the profile coefficient curve is selected as the optimal number of clusters K. Otherwise, select the optimal number of clusters K, where the silhouette coefficient is greater than the coefficient threshold and is a local maximum.

[0015] Perform feature matrix clustering to obtain cluster center coordinates and generate carbon layer label maps, specifically: Iterative clustering is performed based on the determined optimal number of clusters K, and K pixels are selected as the initial cluster centers; Calculate the Euclidean distance from each pixel in the dimensionality-reduced feature matrix to the K cluster centers, and assign each pixel to the category of the nearest cluster center; Calculate the mean of all pixels in each cluster in the reduced feature space, and update the cluster center coordinates; When the distance the cluster center coordinates move in this round is less than the distance threshold, or the number of iterations reaches the upper limit, the iteration stops, and a carbon layer label map is generated based on the updated cluster center coordinates.

[0016] Extract the multidimensional feature vectors of each carbon layer in the carbon layer label image before dimensionality reduction to obtain the feature mean vector. Perform feature matching with the feature fingerprint database of the pre-stored carbon accounting model to bind the carbon accounting model to each carbon layer. Specifically, this includes: The position of the pixels contained in each carbon layer in the reflectivity tensor or the full image feature matrix is ​​inverted based on the carbon layer label map. Based on the indexed position, extract the multidimensional feature vectors of the pixels contained in each carbon layer before dimensionality reduction, calculate the average value, and obtain the feature mean vector of each carbon layer. The similarity between the feature mean vector of each carbon layer and the feature prototype vector corresponding to each carbon accounting model in the feature fingerprint database is calculated one by one. The carbon accounting model with the highest similarity is selected and bound to the corresponding carbon layer.

[0017] The similarity between the feature mean vector of each carbon layer and the feature prototype vector corresponding to each carbon accounting model in the feature fingerprint database is calculated one by one, including: Calculate the cosine similarity between the feature mean vector and the feature prototype vectors in the feature fingerprint database: The carbon nucleus model with the highest cosine similarity was selected as the binding model for this carbon layer. If the highest cosine similarity is lower than the preset threshold, the carbon layer is marked as awaiting manual review or bound to a general accounting model.

[0018] A second aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described herein.

[0019] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are as follows: This invention constructs a fully automated workflow from multispectral remote sensing image input to carbon layer and model binding relationship output: After acquiring multispectral image data, reflectance tensors are directly output after preprocessing; based on this, multidimensional feature vectors are automatically calculated pixel by pixel to construct a full-image feature matrix; then, feature dimensionality reduction is automatically completed through principal component analysis; based on the dimensionality-reduced feature matrix, the optimal number of clusters is automatically determined and clustering is performed to generate carbon layer label maps; finally, the mean vector of each carbon layer feature is automatically extracted and matched with a pre-stored feature fingerprint database to bind a carbon accounting model to each carbon layer. This eliminates the need for manual field surveys and repeated trial-and-error threshold adjustments, greatly shortening the carbon layer division and model matching cycle, and meeting the needs for rapid assessment of large-scale forest carbon sinks.

[0020] First, in the feature extraction stage, a multi-dimensional feature vector is calculated based on the reflectance values ​​of four bands (blue light, green light, red light, and near-infrared), integrating multi-dimensional information such as spectrum and texture, significantly enhancing the feature discrimination between different vegetation types and carbon storage levels. Second, in the clustering stage, this invention determines the optimal number of clusters K based on the dimensionality-reduced feature matrix, avoiding the poor adaptability problems caused by fixed thresholds or fixed number of classifications, and can adaptively divide carbon layers according to the actual characteristics of different forest types. Based on this, feature matrix clustering is performed to generate carbon layer label maps, effectively improving the scientific rigor and rationality of the division results.

[0021] Furthermore, by extracting the multidimensional feature vectors of each carbon layer in the carbon layer label map before dimensionality reduction to obtain the feature mean vector, and performing feature matching with the feature fingerprint database of pre-stored carbon accounting models, the optimal carbon accounting model is automatically bound to each carbon layer. This transforms the subjective matching process, which originally relied on human experience, into a data-driven and repeatable objective calculation process, achieving a connection between carbon layer spatial partitioning and carbon accounting model selection, and avoiding estimation bias caused by the misuse of a single model over a large area.

[0022] In summary, this invention effectively solves the technical problems in the prior art where carbon layer division relies on manual experience or fixed thresholds, resulting in insufficient accuracy and weak generalization ability, as well as the disconnect between carbon layer division and carbon accounting model selection, requiring manual intervention to connect them, leading to the inability to achieve closed-loop automation of the overall workflow and low efficiency. It is used for forest carbon sink measurement and monitoring. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the carbon accounting method based on carbon layer partitioning according to one embodiment of the present invention. Detailed Implementation

[0024] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0026] like Figure 1 As shown, a carbon accounting method based on carbon layer partitioning includes: S1: Acquire multispectral image data of the target forest area, including at least blue, green, red, and near-infrared bands. After preprocessing, output a reflectance tensor with dimensions H×W×4, where H is the number of pixels in the image height direction, W is the number of pixels in the image width direction, and 4 corresponds to the reflectance values ​​of the four bands: blue, green, red, and near-infrared.

[0027] This step acquires multispectral remote sensing image data of the target forest area and converts it into a standardized reflectance tensor through preprocessing, providing reliable, spatially continuous, and radiometrically calibrated basic data input for subsequent feature extraction.

[0028] Four wavelength bands were selected: blue (Band 2, 450-515nm), green (Band 3, 525-600nm), red (Band 4, 630-680nm), and near-infrared (Band 5, 845-885nm). The blue band is sensitive to atmospheric scattering and can be used for atmospheric correction and elimination of scattering effects. The green band, located between the two absorption bands of chlorophyll, reflects vegetation health. The red band, located in the chlorophyll absorption valley, is sensitive to changes in chlorophyll content. The near-infrared band is sensitive to changes in canopy structure and biomass. The combination of these four bands can comprehensively characterize the carbon sink-related features of forest vegetation.

[0029] It should be noted that in the preprocessing, atmospheric correction uses the FLAASH algorithm, based on the radiative transfer equation. (in The radiance received by the sensor. Radiation reflected from the Earth's surface. Atmospheric transmittance, (For atmospheric path radiation), the surface reflectance is obtained by inversion using the MODTRAN4+ radiative transfer model, and the original DN value or radiance data is converted into surface reflectance.

[0030] In addition, geometric correction and georeferencing: multispectral images are unified to the WGS84 coordinate system, and geometric fine correction is performed through RPC parameters or ground control points. Bilinear interpolation is used for resampling, and a uniform spatial resolution (e.g., 30m×30m) is output.

[0031] Spatial stitching and fusion: For UAV multi-strip or multi-scene satellite imagery, feathering fusion or weighted overlay is performed in overlapping areas to eliminate stitching seams and form a spatially continuous full-area image.

[0032] After the above processing, the output is a reflectance tensor with dimensions H×W×4, where H is the number of pixels in the image height direction, W is the number of pixels in the image width direction, and 4 corresponds to the reflectance values ​​of the four bands: blue light, green light, red light, and near-infrared light.

[0033] This step eliminates interference from atmospheric scattering and absorption through atmospheric correction, ensuring that reflectance data accurately reflects the spectral characteristics of surface vegetation and providing quantitatively accurate input for subsequent feature extraction. Geographic registration and spatial stitching ensure spatial continuity and consistency across the entire region, avoiding spatial misalignment issues caused by stitching together multi-source data. The selection of four bands strikes a balance between comprehensive information and data redundancy, covering the key spectral range required for vegetation carbon storage inversion while avoiding the computational burden and data redundancy associated with too many bands.

[0034] S2: Calculate the reflectance tensor pixel by pixel, and obtain a multi-dimensional feature vector characterizing the vegetation based on the reflectance values ​​of the four bands. Fuse the multi-dimensional feature vector to obtain a full-image feature matrix of H×W×N, where N is the number of dimensions of the multi-dimensional feature vector.

[0035] This step calculates the reflectance tensor pixel by pixel, characterizing the carbon sink-related attributes of the vegetation represented by each pixel from multiple dimensions. It expands the original four-band spectral information into a more informative and discriminative multi-dimensional feature space, providing a high-quality input foundation for subsequent carbon layer segmentation.

[0036] This invention designs a standardized multidimensional feature vector, specifically an 8-dimensional feature vector, to characterize vegetation carbon storage-related attributes from four independent dimensions. These include: the spectral feature dimension, which includes: the Normalized Difference Vegetation Index (NDVI), reflecting vegetation greenness and photosynthetic activity, and showing a significant positive correlation with aboveground biomass; the Enhanced Vegetation Index (EVI), which introduces the blue band to reduce the influence of atmospheric scattering and outperforms NDVI in high biomass areas; blue light reflectance, which is related to leaf thickness and canopy shading ratio; and the ratio of near-infrared to red light reflectance, which is strongly correlated with vegetation canopy biomass density.

[0037] Texture feature dimensions: including texture contrast based on gray-level co-occurrence matrix, quantized canopy roughness, and structural complexity.

[0038] Red edge feature dimension: including red edge position (REP), which can effectively distinguish different vegetation function types.

[0039] Spatial heterogeneity dimension: including the neighborhood standard deviation of NDVI, reflecting the spatial heterogeneity of vegetation; and the difference between EVI and NDVI, characterizing the complexity of canopy structure.

[0040] The above eight features reflect the physiological, structural and spatial information of vegetation from different perspectives.

[0041] This step overcomes the limitations of traditional methods that rely on a single vegetation index, comprehensively characterizing vegetation features from four dimensions: spectrum, texture, red edge, and spatial heterogeneity. This significantly enhances the feature differentiation between different vegetation types and carbon storage levels. Moreover, the selection of eight features balances physical interpretability with computational efficiency.

[0042] S3: For the full-image feature matrix, perform feature dimensionality reduction based on the cumulative variance contribution rate through principal component analysis, and output the dimensionality-reduced feature matrix H×W×M, where M is the number of dimensions after feature dimensionality reduction, 1<M<N.

[0043] This step performs dimensionality reduction on the high-dimensional full-image feature matrix. While preserving as much of the original feature information as possible, it reduces the data dimension, lowers the computational complexity and storage overhead of subsequent clustering calculations, and eliminates redundancy and collinearity between features.

[0044] Principal component analysis (PCA) is used for dimensionality reduction. The obtained eigenvalues ​​are sorted from largest to smallest, and the variance contribution rates are accumulated sequentially. The top M principal components whose cumulative variance contribution rates reach a preset threshold (e.g., 95%) are selected.

[0045] The projection matrix (8×M) is constructed using the eigenvectors corresponding to the selected M principal components. The original 8-dimensional feature matrix is ​​multiplied by the projection matrix to obtain the dimensionality-reduced M-dimensional feature matrix (H×W×M), where 1 < M < N.

[0046] It should be noted that a dimensionality reduction quality assessment step can be set up: compare the degree of preservation of inter-class distances in the feature space before and after dimensionality reduction. If dimensionality reduction leads to a significant reduction in inter-class distances, then the M value should be increased appropriately.

[0047] In this step, the N-dimensional features are reduced to M dimensions, significantly reducing the amount of data required for clustering calculations and greatly improving computational efficiency. Furthermore, PCA removes linear correlations between features through orthogonal transformations, eliminating collinearity issues among multiple features and making subsequent clustering more stable and reliable.

[0048] By controlling the cumulative variance contribution rate threshold (≥95%), we ensure that the main information of the original features is preserved during the dimensionality reduction process, and the reduced features can still fully characterize the carbon sink-related attributes of vegetation.

[0049] S4: Based on the reduced feature matrix, determine the optimal number of clusters K, perform feature matrix clustering, obtain the cluster center coordinates, and generate a carbon layer label map, where 3≤K≤8.

[0050] In the dimensionality-reduced feature space, pixels with similar features are grouped into the same carbon layer using unsupervised clustering, thus achieving spatial partitioning of the forest region. The core lies in adaptively determining the optimal number of clusters, enabling the granularity of carbon layer partitioning to automatically adjust according to the actual characteristics of the target forest region, avoiding the adaptation problem caused by a fixed number of classifications.

[0051] The optimal number of clusters K ranges from [3, 8]. In forestry carbon sink projects, it is necessary to distinguish at least three carbon storage levels: high, medium, and low (lower limit K ≥ 3). At the same time, too many carbon layers will result in too few pixels per carbon layer and insufficient statistical representativeness (upper limit K ≤ 8).

[0052] Furthermore, the optimal number of clusters K can be set as a configurable parameter, which users can adjust according to the actual complexity of the target forest area. This invention does not impose any restrictions on this.

[0053] For the determined optimal number of clusters K, perform clustering. After clustering is completed, output the carbon layer label map (dimension H×W, each pixel value is an integer from 1 to K, representing the carbon layer to which it belongs) and the coordinates of the cluster center (K×3 matrix).

[0054] This step abandons the traditional method of fixing the number of clusters or manually thresholding the segmentation, allowing the granularity of carbon layer division to adaptively adjust according to the feature distribution of the image itself, achieving reasonable segmentation results in different forest types (tropical rainforest, cold-temperate coniferous forest, temperate mixed forest, etc.). Furthermore, performing clustering in the dimensionality-reduced 3D feature space ensures clustering accuracy while significantly improving computational efficiency.

[0055] S5: Extract the multidimensional feature vectors of each carbon layer in the carbon layer label map before dimensionality reduction, obtain the feature mean vector, perform feature matching with the feature fingerprint library of the pre-stored carbon accounting model, and bind the carbon accounting model to each carbon layer.

[0056] After the carbon layer spatial partitioning is completed, this step automatically binds each carbon layer to an applicable carbon accounting model. The core of this process is to use a reverse indexing mechanism to trace back from the carbon layer label map to the pixel group of each carbon layer in the original image, extract the mean vector of its original 8-dimensional features before dimensionality reduction, and then match it with a pre-built feature fingerprint database to automatically select the optimal carbon accounting model for each carbon layer.

[0057] It should be noted that the clustering stage in step S4 is carried out in the low-dimensional feature space after principal component analysis compression. Its core purpose is to use principal component analysis to remove linear redundancy between features and to aggregate pixels with similar comprehensive spectral responses into several spatially continuous regional units in order to determine the spatial division boundary of the carbon layer. This stage pursues compactness and computational efficiency in the feature space.

[0058] The model matching stage, which reverts to the original multidimensional feature space before dimensionality reduction, aims to preserve the physical interpretability and independent ecological meaning of each feature dimension, thereby accurately identifying the vegetation function type to which the carbon layer belongs. Therefore, clustering is performed in the low-dimensional space to ensure efficiency, while matching is implemented in the high-dimensional space to ensure accuracy.

[0059] Among them, the feature fingerprint database is a prerequisite for automatic matching in step S5, and its construction method is as follows.

[0060] Data collection: Collect publicly available remote sensing image data and corresponding vegetation / forest type labeling data covering various forest types (tropical rainforest, cold coniferous forest, temperate mixed forest, savanna, artificial economic forest, etc.).

[0061] Feature extraction and grouping: Extract the same 8-dimensional feature vector as in step S2 for each pixel of the training data, and group them according to the vegetation type label.

[0062] Statistical Learning and Prototype Generation: For each set of data, calculate the mean vector of 8-dimensional features. Covariance Matrix Construct the feature prototype vector corresponding to this vegetation type. and its permissible range of variation (e.g.) interval, (Standard deviation).

[0063] Model binding: Based on the known applicable scenarios and vegetation types of each carbon accounting model (such as IPCC Tropical, IPCC Boreal, BIOME-BGC, CENTURY, 3-PG, etc.), the corresponding feature prototype vectors are bound to the carbon accounting model to form a feature fingerprint database.

[0064] In this step, based on the carbon layer label map, for each carbon layer label... ( Find all tag values ​​equal to The set of pixel positions.

[0065] Original feature extraction: Based on the indexed pixel position, extract all 8-dimensional feature vectors of the corresponding pixel from the full-image feature matrix (8-dimensional) before dimensionality reduction, instead of using the 3-dimensional features after dimensionality reduction.

[0066] Mean vector calculation: The original feature mean vector of the carbon layer is obtained by averaging the 8-dimensional feature vectors of all pixels in the carbon layer dimension by dimension. (8-dimensional). It represents the central position of the carbon layer in the original feature space and is a comprehensive characterization of the physiological, structural and spatial features of the vegetation in the carbon layer.

[0067] Cosine similarity matching: Compared with the feature prototype vectors of each model in the feature fingerprint database Calculate cosine similarity. Cosine similarity measures the consistency of two vectors in direction and is suitable for comparing the similarity of feature vectors.

[0068] Model binding and output: Select the carbon accounting model with the highest similarity and bind it to the carbon layer.

[0069] Understandably, the fingerprint database supports dynamic updates: when new regional data is accumulated or a new carbon accounting model is added, the fingerprint database can be quickly expanded through re-statistics or transfer learning.

[0070] In addition, a confidence report can be generated for the matching results, which will indicate the similarity score of each carbon layer match and the recommended model name for forestry experts to review and refer to.

[0071] The matching results can be further compared and verified with historical data. If the model bound to a certain carbon layer is inconsistent with the model of the same area in the same period in the past, the system will automatically prompt the user to verify whether there is a change in vegetation type.

[0072] This step achieves intelligent integration of carbon layer partitioning and model selection, transforming the subjective matching process that originally relied on human experience into a data-driven, repeatable, and objective computation process. Clustering is performed in the dimensionality reduction space to ensure computational efficiency, while model matching is backtracked to the original high-dimensional feature space to ensure information integrity, thus balancing efficiency and accuracy.

[0073] The feature fingerprint database supports dynamic updates, enabling the method to adapt to application scenarios across regions and vegetation types.

[0074] In a preferred embodiment of the present invention, the dimension N of the multidimensional feature vector is 8, including: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Texture Contrast Based on Gray-Level Co-occurrence Matrix (GLCM), Red Edge Position, Blue Band Reflectance, Near-Infrared Band to Red Band Reflectance Ratio, Neighborhood Standard Deviation of NDVI, and Difference Between EVI and NDVI.

[0075] The Normalized Difference Vegetation Index (NDVI) is one of the most widely used vegetation indices, quantifying vegetation greenness and photosynthetic activity. In carbon segregation scenarios, the primary purpose of NDVI is to characterize vegetation growth vigor and biomass levels, providing fundamental spectral characteristics for distinguishing different carbon storage levels. NDVI shows a significant positive correlation with aboveground biomass and is one of the most basic remote sensing indicators for carbon storage estimation.

[0076] The Enhanced Vegetation Index (EVI) is an improved vegetation index based on the National Distance Vibration Index (NDVI). By introducing the blue light band, it reduces the influence of atmospheric scattering and soil background, and outperforms NDVI in high biomass areas. The main purpose of EVI is to overcome the saturation effect of NDVI in high vegetation cover areas and more accurately reflect the growth and change process of vegetation.

[0077] Texture contrast based on gray-level co-occurrence matrices is used to quantify canopy roughness and structural complexity. Canopy structures differ significantly among different forest types; for example, coniferous forests typically have a more uniform canopy texture, while broadleaf mixed forests exhibit a coarser and more complex texture. By introducing texture features, vegetation types with similar spectral characteristics but different canopy structures can be distinguished, improving the accuracy of carbon layer delineation.

[0078] The red edge position (REP) refers to the wavelength position corresponding to the maximum value of the first derivative of the vegetation reflectance spectrum in the red edge region (680–750 nm). The REP is an important indicator parameter for vegetation physiological state and functional type; the red edge position varies systematically among different vegetation functional types (e.g., broadleaf vs. coniferous trees, healthy vs. stressed vegetation). The main purpose of the REP is to effectively distinguish different vegetation functional types, providing a classification-level basis for carbon layer delineation.

[0079] Blue light reflectance is directly taken from the blue light band (Band2) reflectance value in the preprocessed reflectance tensor. Blue light reflectance is related to leaf thickness, canopy shading ratio, and atmospheric scattering. In carbon layer delineation, the main purpose of blue light reflectance is to provide supplementary spectral information independent of vegetation indices, and it has unique value, especially in distinguishing different canopy structures and leaf morphologies.

[0080] The near-infrared to red light reflectance ratio is one of the most fundamental ratio-based indices in the vegetation index family. This ratio is strongly correlated with the biomass density of the vegetation canopy; the denser the vegetation and the higher the biomass, the stronger the near-infrared reflectance and the stronger the red light absorption, resulting in a larger ratio. Its main purpose is to quantify the biomass density of the vegetation canopy in the most direct way.

[0081] The neighborhood standard deviation (NDVI standard deviation) of the Normalized Difference Vegetation Index (NDVI) calculates the dispersion of NDVI values ​​within a 3×3 neighborhood and is used to quantify the spatial heterogeneity of vegetation. In forest ecosystems, spatial heterogeneity reflects the complexity of stand structure. Pure stands (such as monoculture plantations) exhibit relatively small spatial NDVI variation, while mixed forests or forests at different successional stages show greater spatial NDVI variation. The primary purpose of the NDVI standard deviation is to characterize the spatial structure and diversity of vegetation, providing a basis for distinguishing different stand structure types.

[0082] The difference between Enhanced Vegetation Index (EVI) and Normalized Difference Vegetation Index (NDVI) reflects the degree of difference in the responses of the two vegetation indices to the same vegetation target. Since NDVI is more sensitive to chlorophyll, while EVI is more sensitive to canopy structure, the difference between the two canopy indices can characterize the complexity of the canopy structure. In canopies with simple structures (such as monoculture forests), the difference between EVI and NDVI is relatively small; in canopies with complex structures (such as multi-layered mixed forests of different ages), the difference between EVI and NDVI is larger due to canopy shading and multiple scattering effects.

[0083] In one embodiment, obtaining the multidimensional feature vector specifically includes: The process is performed pixel by pixel, and the multidimensional feature vector is obtained based on the relationship between the feature vector of each dimension and the reflectivity tensor.

[0084] NDVI is calculated pixel-by-pixel based on the preprocessed reflectivity tensor: NDVI=(Band5-Band4) / (Band5+Band4); In this design, Band 5 represents the near-infrared reflectance, and Band 4 represents the red light reflectance. This utilizes the stark contrast between the strong absorption characteristics of green vegetation in the red light band and its strong reflectance characteristics in the near-infrared band. The NDVI value ranges from -1 to 1; healthy, dense vegetation typically has a high NDVI value (close to 1), while areas without vegetation (such as bare soil or water bodies) have a low or negative NDVI value.

[0085] It should be noted that before calculating NDVI, the reflectivity tensor can be processed with a cloud mask to remove cloud-covered pixels and avoid interference from clouds on the NDVI value. For pixels with abnormally high or low NDVI values, a threshold filtering mechanism can be set (such as removing pixels with NDVI < 0 or NDVI > 1) to ensure the quality of the input features.

[0086] NDVI is simple to calculate and has a clear physical meaning. It can quickly and effectively reflect the growth status and biomass level of vegetation, providing the most basic spectral discrimination basis for carbon layer division.

[0087] Calculate EVI pixel-by-pixel based on the preprocessed reflectivity tensor: EVI=2.5×(Band5-Band4) / (Band5+6×Band4-7.5×Band2+1); Where Band2 represents the reflectance in the blue light band. Compared to NDVI, EVI introduces the blue light band into the formula for atmospheric correction and uses different gain coefficients (2.5) and constant terms (1) to give it a better linear response in high biomass regions.

[0088] It should be noted that for data from different sensors (such as Landsat 8 and Sentinel-2), the differences in the band response functions of each sensor should be taken into account, and band matching calibration should be performed before calculating EVI.

[0089] NDVI is sensitive to chlorophyll, while EVI is more sensitive to canopy structure. Using them together can more comprehensively characterize vegetation features: NDVI provides reliable greenness information in low to medium biomass areas, while EVI overcomes the saturation effect and provides more accurate vegetation signals in high biomass areas. The two complement each other, improving the ability of the feature system to distinguish different carbon storage levels.

[0090] Construct a gray-level co-occurrence matrix within a 5×5 pixel window and calculate texture contrast: ; in, The elements of the gray-level co-occurrence matrix represent gray values. and The probability of two objects appearing simultaneously in a specified direction and at a given distance. Contrast reflects the degree of drastic change in grayscale in an image; the coarser the texture and the greater the difference in grayscale, the higher the contrast value.

[0091] During calculation, four directions—0°, 45°, 90°, and 135°—are typically selected, and the average contrast of each direction is taken as the final feature value. The window size (5×5) is chosen based on texture feature extraction in forestry remote sensing.

[0092] For different bands, texture features can be calculated separately and compared. The near-infrared band is preferred, as its texture features are usually most sensitive to the response of the canopy structure.

[0093] The introduction of texture features allows carbon layer delineation to no longer rely solely on spectral information, but incorporates spatial structural information. This design can effectively distinguish vegetation types with similar spectral characteristics but different canopy structures (such as shrubs and trees, coniferous forests and broad-leaved forests), significantly improving the sensitivity and discriminative power of carbon layer delineation to differences in vegetation types.

[0094] Third-order polynomial fitting was performed on the spectral range of 680nm to 750nm: ; in, Indicates wavelength Spectral reflectance at that location These are the fitting coefficients. The red-edge position is defined as the wavelength corresponding to the maximum value of the first derivative: ; For multispectral data, the spectral range between the red band (Band4, 630-680nm) and the near-infrared band (Band5, 845-885nm) is relatively wide. Directly calculating the REP requires the use of the red edge band or an approximate estimation through interpolation.

[0095] Specifically, linear interpolation or Lagrange interpolation can be used to interpolate dense spectral points between the red and near-infrared bands, and then the REP can be calculated.

[0096] If the data source is Sentinel-2 (which includes dedicated red-edge bands: Band5 (705nm), Band6 (740nm), and Band7 (783nm)), more accurate REP calculations can be performed directly using the red-edge bands.

[0097] The Residual Expansion Plate (REP) is an important indicator for distinguishing vegetation functional types. In carbon layer delineation, different carbon layers often correspond to different vegetation functional types. For example, the REP of tropical broadleaf forests is usually located at 720-740 nm, while the REP of cold-climate coniferous forests is usually located at 685-710 nm. The introduction of REP allows carbon layer delineation to reflect the essential differences in vegetation functional types, rather than just quantitative differences in biomass, providing a crucial basis for vegetation type identification for accurate matching of subsequent carbon accounting models.

[0098] The atmospherically corrected reflectance value of the blue light band (Band2) in the reflectance tensor is directly used as this feature: ; This feature requires no additional calculation and can be extracted directly from the preprocessed reflectivity tensor.

[0099] It should be noted that for areas heavily influenced by the atmosphere, further thin cloud correction or water vapor correction can be added after extracting the blue light reflectance.

[0100] Blue light reflectance, as an independent spectral dimension, supplements the spectral information not covered by vegetation indices. In carbon layer delineation, blue light reflectance can reflect structural parameters indirectly related to carbon storage, such as leaf thickness and canopy shading ratio, increasing the spectral coverage of the feature system and making the information basis for carbon layer delineation more comprehensive.

[0101] Based on the preprocessed reflectance tensor, the reflectance ratio of the near-infrared band to the red band is calculated pixel by pixel: NIR / R = Band5 / Band4; This feature is essentially the simplest form of vegetation index, without involving gain coefficients or constant terms, and directly reflects the relative strength of near-infrared and red light reflectance.

[0102] It should be noted that, to prevent abnormal ratios caused by a denominator of zero or a very small value, a minimum denominator protection mechanism can be set (e.g., when...). season ,in (It is a small positive number, such as 0.001).

[0103] The NIR / R ratio captures the core information of vegetation canopy biomass density in its simplest form. Compared with NDVI and EVI, the NIR / R ratio does not undergo nonlinear transformation, maintaining a more linear response relationship and exhibiting a relatively mild saturation effect in high biomass regions. These three features based on the red-near-infrared combination (NDVI, EVI, and NIR / R ratio) characterize vegetation biomass information from different perspectives, forming a multi-level biomass representation system.

[0104] First, calculate the NDVI value for each pixel using the method described above, and then calculate the standard deviation of the NDVI within a 3×3 pixel window: ; in, (Number of pixels within a 3x3 window) This represents the average NDVI value within the window.

[0105] The choice of window size (3×3) is based on the correlation between NDVI standard deviation and forest biodiversity in forestry remote sensing. The standard deviation of NDVI at the 3×3 pixel scale has the highest correlation with the biodiversity index.

[0106] It should be noted that the window size can be set as a configurable parameter. For high-resolution images, a larger window (such as 5×5 or 7×7) can be used, and for low-resolution images, a smaller window (such as 3×3) can be used. There is no restriction on this.

[0107] The introduction of the NDVI standard deviation incorporates the spatial heterogeneity dimension into carbon layer partitioning. In carbon layer partitioning, regions with high spatial heterogeneity may require more refined carbon layer divisions, while spatially homogeneous regions can be appropriately simplified. This feature provides important spatial structure information for subsequent adaptive clustering, improving the ecological rationality of the partitioning results.

[0108] Calculate the difference between EVI and NDVI pixel by pixel: ; The calculation of this feature relies on the previously calculated EVI and NDVI values ​​and requires no additional data source.

[0109] It is the only one of the eight features that directly characterizes the complexity of canopy structure. In carbon layer partitioning, canopy structure complexity directly affects the vertical distribution of carbon storage and the applicability of the carbon accounting model. Complex canopies (such as tropical rainforests) usually require more complex carbon accounting models, while simple canopies (such as plantations) can use simpler models. This feature provides crucial canopy structure information for the intelligent matching of subsequent carbon accounting models.

[0110] Traditional methods often use a single vegetation index (such as NDVI) for carbon layer delineation, which makes it difficult to distinguish the essential differences in carbon storage among different vegetation types. The 8-dimensional feature system of this invention comprehensively characterizes vegetation features from four dimensions: spectrum, texture, red edge, and spatial heterogeneity, significantly enhancing the feature differentiation between different vegetation types and carbon storage levels.

[0111] Eight features characterize vegetation attributes from different perspectives, and there is no simple linear dependency between them. NDVI and EVI are complementary (chlorophyll sensitivity vs. canopy structure sensitivity), spectral features and texture features are complementary (spectral information vs. spatial structure information), and red-edge features provide vegetation function type information that cannot be replaced by other features. This low-redundancy, highly complementary feature design provides a high-quality input foundation for subsequent PCA dimensionality reduction and clustering.

[0112] When a carbon layer exhibits an anomaly in a certain characteristic, it can be traced back to the corresponding vegetation attribute (e.g., an abnormally high NDVI indicates vigorous vegetation growth, and a REP shift indicates a change in vegetation functional type, etc.), which facilitates forestry experts in understanding and reviewing the classification results.

[0113] The 8-dimensional feature has a moderate dimensionality, containing enough information to characterize carbon storage-related attributes without causing excessive computational burden due to high dimensionality. While ensuring information integrity, it provides a computationally feasible input size for subsequent PCA dimensionality reduction and clustering.

[0114] In another embodiment, obtaining the multidimensional feature vector specifically includes: The reflectivity tensor is input as 4-channel input data into a convolutional neural network, which includes three convolutional layers and a global average pooling layer. The first convolutional layer yields local spectral features, the second convolutional layer yields combined low-level features, the third convolutional layer yields high-level abstract features, and the multi-dimensional feature vector is output after passing through a global average pooling layer. The number of convolution kernels in the third convolutional layer is greater than the number of convolution kernels in the second convolutional layer, and the number of convolution kernels in the second convolutional layer is greater than the number of convolution kernels in the first convolutional layer.

[0115] Multidimensional feature vectors are automatically learned and extracted from reflectivity tensors using convolutional neural networks (CNNs). This automates and automates feature extraction, enabling feature representations to adaptively capture complex vegetation patterns in different forest scenarios.

[0116] In areas with complex vegetation types and a prevalence of mixed pixels, physical formulas may be insufficient to fully characterize vegetation features, while CNNs can automatically discover better feature combinations from the data.

[0117] The CNN network designed in this embodiment consists of three convolutional layers and one global average pooling layer. The first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 and padding of 1. This layer directly operates on the input 4-channel reflectance tensor (blue, green, red, and near-infrared), extracting local spectral features through convolution operations, such as local variation patterns of reflectance in each band, simple edge and texture primitives, etc. A small number of convolutional kernels (32) is sufficient to capture basic spectral spatial features at this stage.

[0118] The second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1 and padding of 1. This layer receives the feature map output from the first layer and combines low-level features into mid-level features through convolution operations, such as correlation patterns between bands and more complex texture structures. The number of convolutional kernels is increased to 64 to accommodate richer feature combinations.

[0119] The third convolutional layer uses 128 3×3 convolutional kernels with a stride of 1 and padding of 1. This layer further extracts high-level abstract features based on the feature maps of the second layer, such as semantic information closely related to carbon storage, including vegetation type discrimination patterns and canopy structure patterns. The number of convolutional kernels is increased to 128 to capture higher-level and more diverse feature patterns.

[0120] Global average pooling layer: Located after the third convolutional layer, it compresses each feature map into a scalar value using global average pooling. Specifically, for the 128 feature maps output from the third layer (each feature map is H×W), global average pooling calculates the mean of all pixels in each feature map, converting the 128 feature maps into a 128-dimensional feature vector. This feature vector is the multi-dimensional feature vector output in this embodiment.

[0121] Furthermore, batch normalization layers can be added after each convolutional layer and before the activation function to accelerate network convergence and improve training stability. Residual connections can be introduced to add or concatenate shallow and deep feature maps, alleviating the vanishing gradient problem and promoting feature reuse. Different sized convolutional kernels (such as 3×3 and 5×5) can be used in the same layer to capture feature patterns at different spatial scales.

[0122] In this invention, the feature vector dimension of the output of the global average pooling layer is equal to the number of convolution kernels in the third convolution layer (i.e., 128 dimensions). The number of convolution kernels in the third layer can be adjusted according to actual needs to change the output feature dimension.

[0123] The number of convolutional kernels in the third layer is greater than the number of convolutional kernels in the second layer, which in turn is greater than the number of convolutional kernels in the first layer. Shallow networks process local, basic features (such as edges and spectral responses), requiring fewer feature types, thus necessitating fewer convolutional kernels. Deep networks process global, abstract features (such as vegetation type and canopy structure), needing to capture more diverse feature patterns, thus requiring more convolutional kernels. The number of convolutional kernels increases with network depth, achieving hierarchical evolution of features.

[0124] Each convolutional layer combines and abstracts features from the previous layer. As the network deepens, the spatial resolution of the feature maps gradually decreases (due to convolution and pooling operations), but the number of feature channels (i.e., the number of convolutional kernels) gradually increases. This allows the network to expand its semantic information capacity while simplifying spatial information, maintaining its overall information expressive power without diminishing.

[0125] Moreover, shallow feature maps are larger in size (close to the original input size), so fewer convolutional kernels can be used to control the computational load; deep feature maps are smaller in size, so using more convolutional kernels will not significantly increase the computational burden, thus achieving a balance between computational efficiency and expressive power.

[0126] It should be noted that the CNN in this embodiment is a pre-trained network, that is, it has been trained on a large-scale public remote sensing dataset before being applied to the carbon layer partitioning task of this invention.

[0127] Specific pre-training strategies include training datasets: using publicly available remote sensing image libraries that include vegetation type annotations (such as annotated data containing multiple forest types and different geographical regions) to ensure that the training data covers the diversity of the target application scenarios.

[0128] Auxiliary tasks: The network is pre-trained using vegetation type classification or feature reconstruction as auxiliary tasks. Through the vegetation type classification task, the network learns discriminative features to distinguish different vegetation functional types; through the feature reconstruction task, the network learns the representational ability to maintain the integrity of the input information.

[0129] Loss function and optimizer: The loss function can be mean squared error (MSE, suitable for feature reconstruction tasks) or cross-entropy loss (suitable for classification tasks). The optimizer is Adam, with an initial learning rate of 0.001. Training stops when the loss on the validation set converges.

[0130] After pre-training is completed, the network parameters are frozen or used as initialization parameters and directly used for the feature extraction task of this invention, without the need for retraining for each target forest region.

[0131] Compared to methods that calculate 8-dimensional features pixel-by-pixel based on physical formulas, CNNs eliminate the need for manually designed feature calculation formulas; the network automatically learns the optimal feature representation through pre-training. In areas with complex vegetation types and widespread mixed pixels, CNNs can capture non-linear feature patterns from the data that are difficult to fully characterize with manual formulas, thus improving the adaptive capability of feature extraction.

[0132] Moreover, the three-layer convolutional structure enables hierarchical feature extraction. The shallow layer captures basic information such as spectral response, the middle layer combines inter-band correlation patterns, and the deep layer extracts semantic information such as vegetation type and canopy structure. The three layers of features complement each other, forming a complete characterization of vegetation carbon sink-related attributes.

[0133] Global average pooling layers replace traditional fully connected layers, directly compressing the 128 feature maps output from the third convolutional layer into a 128-dimensional feature vector, significantly reducing the number of network parameters. Fully connected layers typically require millions of parameters, while global average pooling layers require no trainable parameters. This reduction in parameters lowers the risk of overfitting, allowing the network to maintain good generalization performance even in scenarios with small sample sizes.

[0134] This embodiment provides a dual implementation approach, combining the CNN extraction path with the pixel-by-pixel calculation path based on physical formulas. The physical formula path offers advantages such as simple computation and strong physical interpretability, making it suitable for fast processing in conventional scenarios. The CNN path, on the other hand, boasts high automation and strong adaptability, making it suitable for refined feature extraction in complex scenarios. The two paths can be flexibly selected or combined according to the specific application scenario.

[0135] In a preferred embodiment of the present invention, the full-image feature matrix is ​​subjected to feature dimensionality reduction based on the cumulative variance contribution rate through principal component analysis, including: The full-image feature matrix is ​​centered by subtracting the global mean of the corresponding dimension from each feature dimension; Calculate the covariance matrix between each feature dimension; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and corresponding eigenvectors. The variance contribution rates are accumulated sequentially from largest to smallest eigenvalues, and the top M principal components whose cumulative variance contribution rates reach a preset threshold are selected. The original multi-dimensional features are mapped to M-dimensional features by constructing a projection matrix using the feature vectors corresponding to the selected M principal components, and the reduced feature matrix is ​​output. Where M takes the value 3.

[0136] In step S2, this invention expands each pixel from a 4-band reflectance value into a standardized 8-dimensional feature vector, constructing a full-image feature matrix with dimensions H×W×8. The 8-dimensional features characterize the carbon sink-related attributes of vegetation from four dimensions: spectrum, texture, red edge, and spatial heterogeneity. However, directly clustering with high feature dimensions can lead to two problems: increased computational complexity, as the distance calculation in high-dimensional space increases linearly with dimension, affecting clustering efficiency; and feature redundancy and collinearity, as there is a certain degree of correlation among the 8 features (e.g., NDVI and EVI are both calculated based on the red-near-infrared bands), and redundant information can interfere with the clustering effect.

[0137] The main purpose of this step is to perform linear dimensionality reduction on the 8-dimensional full-image feature matrix through principal component analysis, projecting the original features onto several orthogonal directions with the largest variance. While preserving as much of the main information of the original features as possible, the feature dimension is compressed from 8 dimensions to a lower dimension, thereby reducing the computational complexity of subsequent clustering calculations. At the same time, it eliminates the linear correlation between features, providing a more compact and stable input for clustering.

[0138] Principal Component Analysis (PCA), a classic unsupervised linear dimensionality reduction method, works by constructing all linear combinations of the original features as principal components. Simultaneously, by selecting weights in these linear combinations, the principal components are chosen to have the maximum possible variance and are orthogonal to all previous principal components. This approach significantly reduces the feature dimensionality of the data while preserving its main variability.

[0139] The full-image feature matrix (H×W×8) is centered. It is then expanded into a two-dimensional matrix of (H×W)×8 (where H×W is the total number of valid pixels). For each feature dimension... ( ), calculate the global mean of this dimension across all cells. .

[0140] Then, the mean of the corresponding dimension is subtracted from the 8-dimensional feature vector of each pixel. The purpose of centering is to shift the origin of the data coordinates to the data center, so that the subsequent calculation of the covariance matrix is ​​not affected by the feature dimensions and absolute position, and ensures that the PCA transformation only reflects the relative variation structure of the data.

[0141] Based on the centered data, calculate the 8×8 covariance matrix. The elements of the covariance matrix C are... This represents the linear correlation between the p-th feature and the q-th feature. The covariance matrix reflects the correlation structure among the eight feature dimensions. If two features are highly positively correlated (such as NDVI and EVI), their corresponding covariance values ​​in the covariance matrix will be large, indicating information redundancy, which needs to be eliminated through PCA.

[0142] For covariance matrix Perform eigenvalue decomposition: ; Where Λ is a diagonal matrix, and the elements on the diagonal are eigenvalues. (arranged from largest to smallest); Q is the eigenvector matrix, with each column being an 8-dimensional unit eigenvector, representing a projection direction of the data in the original 8-dimensional space.

[0143] Eigenvalues This represents the variance of the data along the corresponding principal component direction. A larger eigenvalue indicates that the direction contains more data variation information. The eigenvector defines the projection direction from the original 8-dimensional feature space to the principal component space.

[0144] The contribution rate of the k-th principal component is defined as the proportion of the eigenvalue corresponding to that principal component to the sum of all eigenvalues. Cumulative variance contribution rate: the cumulative variance contribution rate of the first m principal components.

[0145] The variance contribution rates are accumulated sequentially from largest to smallest eigenvalues, and the smallest m value that satisfies the condition that the cumulative variance contribution rate is greater than or equal to a preset threshold is selected as the number of principal components to be retained.

[0146] In this preferred embodiment, the preset threshold is 95%. This threshold determines the number of principal components to retain, ensuring that information loss after dimensionality reduction is controllable. In remote sensing image processing, in most cases, the first three or four principal components can describe more than 95% of the variance.

[0147] Using the eigenvectors corresponding to the selected M principal components, a projection matrix P (8×M) is constructed, which transforms the original 8-dimensional feature matrix. Multiply by the projection matrix. Output the dimensionality-reduced feature matrix with dimensions H×W×M. Each pixel is projected from the original 8-dimensional space into an M-dimensional subspace spanned by the directions of the first M principal components.

[0148] In this preferred embodiment, M is set to 3. In the field of remote sensing image processing, the first three or four grids in the multi-band grid generated by PCA can typically describe more than 95% of the variance.

[0149] The 3D features after dimensionality reduction retain more than 95% of the information in the original 8D features, with information loss controlled within 5%. In the application scenario of carbon layer partitioning, this information retention ratio is sufficient to ensure that the accuracy of subsequent clustering and partitioning is not significantly affected.

[0150] Reducing the feature dimension from 8 to 3 dimensions lowers the computational complexity of distance calculations in subsequent clustering operations. For large forest areas (H×W can reach millions or even tens of millions of pixels), this efficiency improvement has significant engineering value.

[0151] Moreover, after dimensionality reduction to 3D, the distribution of each pixel in the feature space can be visually displayed in a three-dimensional coordinate system, which makes it easier for forestry experts to visually check the rationality of the clustering results and the separation degree of the feature space.

[0152] In a preferred embodiment of the present invention, the optimal number of clusters K is determined based on the dimensionality-reduced feature matrix, including: Traverse the range of cluster numbers [3, 8]; For each cluster value, multiple cluster centers are randomly initialized, and clustering is performed separately for each cluster. Calculate the average silhouette coefficient for each clustering result to obtain the silhouette coefficient curve; If the slope change of adjacent points is greater than or equal to the slope threshold, then the profile coefficient curve is determined to have an inflection point. The point corresponding to the maximum slope change of adjacent points is taken as the inflection point, and the clustering value corresponding to the inflection point of the profile coefficient curve is selected as the optimal number of clusters K. Otherwise, select the optimal number of clusters K, where the silhouette coefficient is greater than the coefficient threshold and is a local maximum.

[0153] In step S3, the present invention has reduced the full-image feature matrix from 8 dimensions to 3 dimensions (typical value), obtaining a compact feature matrix with sufficient information retention after dimensionality reduction.

[0154] However, K-means clustering faces a key problem: the number of clusters K needs to be determined in advance, but different forest types (such as tropical rainforests, cold-temperate coniferous forests, and temperate mixed forests) exhibit significant differences in spatial heterogeneity and carbon layer differentiation, making a fixed number of clusters unsuitable for all scenarios. If the value of K is too small, the division is too coarse and cannot effectively reflect the spatial differences in carbon storage within the region; if the value of K is too large, it will result in too few pixels per carbon layer, insufficient statistical representativeness, and a sharp increase in the complexity of subsequent model management.

[0155] The main purpose of this step is to dynamically evaluate the clustering quality under different K values ​​using the silhouette coefficient method, so that the granularity of carbon layer division can be automatically adjusted according to the actual characteristics of the target forest area.

[0156] Based on the determined optimal number of clusters, the pixels in the feature space are divided into K homogeneous subsets through iterative clustering, generating a spatially continuous carbon layer label map, which provides a spatial partitioning basis for model matching in step S5.

[0157] The candidate cluster number is set to a range of [3, 8]. This range is determined based on the following: The lower limit K≥3: In forestry carbon sequestration projects, the purpose of carbon layer division is to divide forest areas into several relatively homogeneous carbon layer units. From an ecological and carbon sequestration practice perspective, at least three carbon storage levels—high, medium, and low—need to be distinguished to effectively reflect the spatial differences in carbon storage within the region. If K<3 (i.e., K=1 or 2), the division is too coarse and cannot meet the basic requirement of carbon sequestration for characterizing spatial heterogeneity.

[0158] Upper limit K≤8: In forestry carbon sink project development, the number of carbon layers should generally not be too large. Too many carbon layers will result in too few pixels per carbon layer, insufficient statistical representativeness, and a sharp increase in model management complexity when matching carbon accounting models for each carbon layer. From the perspective of clustering effectiveness, for most forest remote sensing scenarios, eight or fewer carbon layers can adequately represent the main forest stand types and carbon storage gradients.

[0159] In practical applications, this range can be adjusted as a configurable parameter based on the complexity of the target area.

[0160] For each candidate K value (from 3 to 8), the following sub-process is executed: multiple sets of random initializations are performed, randomly initializing multiple sets (e.g., 10 sets) of different cluster centers. Since the K-means algorithm is sensitive to the selection of initial centers, multiple initializations can avoid getting trapped in local optima and improve the reliability of the evaluation.

[0161] Perform K-means clustering: For each group initialization, perform K-means clustering on the dimensionality-reduced 3-dimensional feature space (the maximum number of iterations is set to 50, and the early termination threshold is set to center shift <0.01).

[0162] Select the optimal result: Select the group with the highest average profile coefficient from multiple results as the optimal clustering result under this K value.

[0163] Calculate the average profile coefficient: for sample points Its profile coefficient is defined as: ; Where a(i) is the average distance from sample point i to other samples in the same cluster (intra-class compactness), and b(i) is the average distance from sample point i to its nearest neighbor cluster (inter-class separation). The silhouette coefficient ranges from [-1, 1], with a larger value indicating better clustering. The average silhouette coefficient at this K value is obtained by averaging the silhouette coefficients of all sample points.

[0164] Plot the profile coefficient curve: Plot the profile coefficient as a function of the K value on the x-axis and the average profile coefficient on the y-axis.

[0165] The inflection point of the silhouette coefficient curve means that the improvement in cluster quality from continuing to increase the number of clusters decreases significantly, and it is the optimal balance point between the number of clusters and cluster quality.

[0166] The determination logic is as follows: calculate the slope change between adjacent points on the profile coefficient curve.

[0167] If the slope change of adjacent points is greater than or equal to the preset slope threshold (e.g., 0.05), then the contour coefficient curve is determined to have an inflection point.

[0168] The point corresponding to the maximum change in slope between adjacent points is taken as the inflection point. The cluster value corresponding to the inflection point is selected as the optimal number of clusters K.

[0169] When the improvement in the silhouette coefficient brought about by increasing the number of clusters suddenly decreases, it indicates that continuing to increase the number of clusters can no longer significantly improve the clustering quality, and the K value at this time is the optimal choice.

[0170] If the profile coefficient curve does not have a clear inflection point (i.e., the slope change of all adjacent points is less than the slope threshold), then the backup judgment strategy is activated.

[0171] Select the K value that has a contour coefficient greater than a preset coefficient threshold (e.g., 0.5) and is a local maximum as the optimal number of clusters.

[0172] A silhouette coefficient greater than 0.5 indicates that the clustering results have good cohesion and separation; local maxima ensure that the selected K value is the one with the best clustering quality within that interval.

[0173] If the silhouette coefficients of all candidate K values ​​are not greater than the coefficient threshold, then the K value with the highest silhouette coefficient is selected, and the system can also mark and suggest manual review of the clustering quality.

[0174] Output the determined optimal number of clusters K (integer, 3≤K≤8) for use in subsequent iterative clustering.

[0175] Furthermore, feature matrix clustering is performed to obtain cluster center coordinates and generate carbon layer label maps, specifically: Iterative clustering is performed based on the determined optimal number of clusters K, and K pixels are selected as the initial cluster centers; Calculate the Euclidean distance from each pixel in the dimensionality-reduced feature matrix to the K cluster centers, and assign each pixel to the category of the nearest cluster center; Calculate the mean of all pixels in each cluster in the reduced feature space, and update the cluster center coordinates; When the distance the cluster center coordinates move in this round is less than the distance threshold, or the number of iterations reaches the upper limit, the iteration stops, and a carbon layer label map is generated based on the updated cluster center coordinates.

[0176] Initialize cluster centers (K-means++ strategy), selecting K pixels as initial cluster centers: Randomly select the first initial cluster center from all pixels. For each subsequent initial center, calculate the squared distance from each pixel to the selected nearest center. With probability Select the next center. Repeat until K initial cluster centers are selected. Each is a 3-dimensional vector.

[0177] The K-means++ strategy effectively avoids the local optimum problem caused by traditional random initialization by dispersing the initial centers as much as possible, thus improving the stability and quality of clustering results.

[0178] In each iteration, the following two sub-steps are executed sequentially: Assignment step: Calculate the Euclidean distance from each pixel in the dimensionality-reduced feature matrix to the K cluster centers, and assign each pixel to the category of the nearest cluster center. Calculate the mean of all pixels within each cluster in the dimensionality-reduced feature space, and update the cluster center coordinates.

[0179] After each iteration, check the following two termination conditions: center movement distance condition: calculate the movement distance of all cluster centers in this round. .like If the distance is less than a preset distance threshold (e.g., 0.01), the algorithm is considered to have converged and the iteration is terminated.

[0180] Iteration count condition: If the number of iterations reaches the preset limit (e.g., 200 times), the iteration will be forcibly terminated to avoid infinite loops.

[0181] The iteration stops when any of the above conditions are met, and a carbon layer label map is generated based on the updated cluster center coordinates.

[0182] Based on the final clustering results, each pixel is assigned a carbon layer label (an integer from 1 to K), and a carbon layer label map (dimension H×W, where each pixel value is an integer from 1 to K, representing the carbon layer to which it belongs) is generated. At the same time, the cluster center coordinates (a K×3 matrix, representing the center position of each carbon layer in the 3D feature space after dimensionality reduction) are output.

[0183] This invention abandons the traditional methods of fixing the number of clusters or manually setting thresholds for segmentation. Instead, it uses the silhouette coefficient method to dynamically determine the optimal number of clusters, allowing the granularity of carbon layer segmentation to adaptively adjust according to the characteristic distribution of the image itself. This invention achieves reasonable segmentation results in different forest types (tropical rainforest, cold-climate coniferous forest, temperate mixed forest, etc.), significantly improving the generalization ability and segmentation rationality of the method in different ecological and geographical regions.

[0184] This invention employs the K-means++ strategy to select initial cluster centers in iterative clustering, ensuring that the initial centers are as dispersed as possible and effectively avoiding the local optima problem caused by traditional random initialization. Furthermore, by selecting the result with the highest silhouette coefficient from multiple random initializations, the stability and repeatability of the clustering results are further guaranteed, ensuring consistent carbon layer partitioning results for the same region under different operating conditions.

[0185] Clustering is performed in the reduced 3D feature space, resulting in significantly lower computational complexity compared to the original 8D space. Furthermore, the silhouette coefficient method's computational overhead only involves pre-clustering within the candidate K value range (K=3 to 8, 10 initializations per group), keeping the total computational load manageable. This invention achieves high computational efficiency while maintaining partitioning accuracy, meeting the engineering application requirements for carbon layer partitioning in large-scale forest areas.

[0186] In a preferred embodiment of the present invention, the multidimensional feature vectors of each carbon layer in the carbon layer label image before dimensionality reduction are extracted to obtain the feature mean vector. This vector is then matched with the feature fingerprint database of a pre-stored carbon accounting model to bind a carbon accounting model to each carbon layer. Specifically, this includes: The position of the pixels contained in each carbon layer in the reflectivity tensor or the full image feature matrix is ​​inverted based on the carbon layer label map. Based on the indexed position, extract the multidimensional feature vectors of the pixels contained in each carbon layer before dimensionality reduction, calculate the average value, and obtain the feature mean vector of each carbon layer. The similarity between the feature mean vector of each carbon layer and the feature prototype vector corresponding to each carbon accounting model in the feature fingerprint database is calculated one by one. The carbon accounting model with the highest similarity is selected and bound to the corresponding carbon layer.

[0187] After the carbon layers are delineated, the selection of the carbon accounting model usually relies on human experience and judgment. Forestry personnel subjectively select the applicable accounting model (such as the IPCC default method, BIOME-BGC, CENTURY, etc.) based on information such as vegetation type and forest stand characteristics of the carbon layers.

[0188] This process has the following problems: different people have different judgment criteria, the model selection results vary from person to person, and the subjectivity is strong; when a large area involves dozens or even hundreds of carbon layers, the workload of manual matching is huge and the efficiency is low; the model selection is based solely on qualitative judgment based on experience and lacks quantitative evidence to support it.

[0189] This implementation uses a reverse indexing mechanism to trace back from the carbon layer label map to the pixel group of each carbon layer in the original image, and re-associates the spatial partitioning results of the carbon layers with the original feature information, thereby realizing the automatic conversion of carbon layer partitioning results to model matching input.

[0190] By extracting the original 8-dimensional feature mean vector of each carbon layer and performing cosine similarity matching with a pre-built feature fingerprint database, the optimal carbon accounting model is quantitatively and automatically selected for each carbon layer without the need for manual intervention in the matching process between carbon layers and models.

[0191] In one embodiment, a pre-built feature fingerprint library is shown in Table 1.

[0192] Table 1 Pre-built Feature Fingerprint Database

[0193] Based on the generated carbon layer label map (dimension H×W, each pixel value is an integer from 1 to K), for each carbon layer label... ( ), find all tag values ​​equal to in the carbon layer label map set of pixel positions .

[0194] Clustering is performed in the 3D feature space after dimensionality reduction, and the cluster center coordinates are also 3D. However, through the carbon layer label map, it is possible to back-map to the spatial location of all pixels contained in each carbon layer in the original image. This allows subsequent steps to extract information from the original feature space before dimensionality reduction, achieving efficient clustering in low-dimensional space and ensuring information integrity by returning to high-dimensional space for feature extraction.

[0195] Based on the set of indexed pixel positions Extract all 8-dimensional feature vectors at corresponding positions from the full-image feature matrix (8-dimensional) before dimensionality reduction.

[0196] Understandably, while dimensionality reduction improves computational efficiency, it inevitably results in the loss of some information (approximately 5%). In the model matching phase, it is necessary to preserve as much of the vegetation physiological, structural, and spatial characteristics of each carbon layer as possible to ensure matching accuracy.

[0197] The original mean feature vector of the carbon layer is obtained by averaging the 8-dimensional feature vectors of all pixels in the carbon layer dimension by dimension. (8-dimensional). This represents the central location of the carbon layer in the original feature space, and is a comprehensive representation of the physiological, structural, and spatial characteristics of the vegetation within that carbon layer. It is compared with the feature prototype vector in the fingerprint database. They have the same dimensions and physical meaning, and are therefore comparable.

[0198] Furthermore, the similarity between the feature mean vector of each carbon layer and the feature prototype vector corresponding to each carbon accounting model in the feature fingerprint database is calculated one by one, including: Calculate the cosine similarity between the feature mean vector and the feature prototype vectors in the feature fingerprint database: The carbon nucleus model with the highest cosine similarity was selected as the binding model for this carbon layer. If the highest cosine similarity is lower than the preset threshold, the carbon layer is marked as awaiting manual review or bound to a general accounting model.

[0199] The original feature mean vector of each carbon layer Compared with the feature prototype vectors of each model in the feature fingerprint database Calculate the cosine similarity for each vector. Cosine similarity measures the consistency of two vectors in direction, rather than their absolute numerical value. Its value ranges from -1 to 1. The closer the value is to 1, the more consistent the directions of the two vectors are, meaning that the feature structure of the carbon layer matches the vegetation features of the corresponding model's applicable scenario more closely.

[0200] Because biomass levels may differ across carbon layers (manifested as different feature vector magnitudes), but model matching is more concerned with the similarity of feature structures (i.e., the relative proportions between feature dimensions), cosine similarity can precisely eliminate the influence of magnitude and focus on the consistency of direction.

[0201] For each carbon layer k, the carbon nucleus model with the highest cosine similarity is selected. This serves as the binding model for the carbon layer, while simultaneously recording the highest similarity score. .

[0202] like (Preset threshold, default 0.6): This model is formally bound to the carbon layer. The confidence level of the binding relationship is high, and it can be directly used for subsequent automated carbon accounting.

[0203] like The system marks the carbon layer as requiring manual review and selects a preset general carbon accounting model (such as the IPCC default parameter model) as a temporary binding scheme. Simultaneously, a prompt message is generated in the output report, pushing the original image slices, feature mean vectors, and candidate model list of the carbon layer to the expert review interface for subsequent manual confirmation by forestry experts.

[0204] The final output includes a carbon layer distribution map and a model mapping table (e.g., carbon layer 1 → IPCC Tropical, similarity 0.87; carbon layer 2 → BIOME-BGC, similarity 0.79; carbon layer 3 → pending manual review, similarity 0.52). This result can be directly used for subsequent automated carbon storage calculation.

[0205] It should be noted that, in addition to cosine similarity, Euclidean distance and Pearson correlation coefficient can also be introduced as auxiliary similarity measures. When different measures give inconsistent optimal models, a weighted voting mechanism can be used to make a comprehensive judgment, thereby improving the robustness of matching.

[0206] This invention reduces the features from 8 dimensions to 3 dimensions using PCA, and performs clustering in the 3-dimensional space to ensure computational efficiency. It then backtracks to the original 8-dimensional feature space before dimensionality reduction to extract the feature mean vector of each carbon layer. This approach leverages the computational efficiency advantage of dimensionality reduction while preserving the information integrity advantage of the original high-dimensional feature space, achieving a balance between efficiency and accuracy.

[0207] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described herein.

[0208] Therefore, it can achieve any effect of carbon accounting methods based on carbon layer partitioning, which will not be elaborated here.

[0209] For any parts not mentioned in this invention, existing technologies can be used or referenced.

[0210] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0211] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A carbon accounting method based on carbon layer partitioning, characterized in that, include: Acquire multispectral image data of the target forest area, including at least blue, green, red, and near-infrared bands. After preprocessing, output a reflectance tensor with dimensions H×W×4, where H is the number of pixels in the image height direction, W is the number of pixels in the image width direction, and 4 corresponds to the reflectance values ​​of the four bands: blue, green, red, and near-infrared. The reflectance tensor is calculated pixel by pixel. Based on the reflectance values ​​of the four bands, a multidimensional feature vector characterizing the vegetation characteristics is obtained. The multidimensional feature vector is fused to obtain a full-image feature matrix of H×W×N, where N is the number of dimensions of the multidimensional feature vector. For the full-image feature matrix, principal component analysis is used to reduce the feature dimension based on the cumulative variance contribution rate, and the reduced feature matrix H×W×M is output, where M is the number of dimensions after feature dimensionality reduction, 1<M<N. Based on the reduced-dimensional feature matrix, the optimal number of clusters K is determined, feature matrix clustering is performed, the coordinates of the cluster centers are obtained, and a carbon layer label map is generated, where 3≤K≤8. Extract the multidimensional feature vectors of each carbon layer in the carbon layer label image before dimensionality reduction to obtain the feature mean vector. Perform feature matching with the feature fingerprint database of pre-stored carbon accounting models to bind a carbon accounting model to each carbon layer, including... The position of the pixels contained in each carbon layer in the reflectivity tensor or the full image feature matrix is ​​inverted based on the carbon layer label map. Based on the indexed position, extract the multidimensional feature vectors of the pixels contained in each carbon layer before dimensionality reduction, calculate the average value, and obtain the feature mean vector of each carbon layer. The similarity between the feature mean vector of each carbon layer and the feature prototype vector corresponding to each carbon accounting model in the feature fingerprint database is calculated one by one. The carbon accounting model with the highest similarity is then bound to the corresponding carbon layer. Specifically, Calculate the cosine similarity between the feature mean vector and each feature prototype vector in the feature fingerprint database. The carbon accounting model with the highest cosine similarity was selected as the binding model for this carbon layer. If the highest cosine similarity is lower than the preset threshold, the carbon layer is marked as awaiting manual review or bound to a general accounting model.

2. The method according to claim 1, characterized in that, The dimension N of the multidimensional feature vector is 8, including: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Texture Contrast Based on Gray-Level Co-occurrence Matrix (GLCM), Red Edge Position, Blue Band Reflectance, Near-Infrared Band to Red Band Reflectance Ratio, Neighborhood Standard Deviation of NDVI, and Difference Between EVI and NDVI.

3. The method according to claim 2, characterized in that, The process of obtaining the multidimensional feature vector specifically includes: The reflectivity tensor is input as 4-channel input data into a convolutional neural network, which includes three convolutional layers and a global average pooling layer. The first convolutional layer yields local spectral features, the second convolutional layer yields combined low-level features, the third convolutional layer yields high-level abstract features, and the multi-dimensional feature vector is output after passing through a global average pooling layer. The number of convolution kernels in the third convolutional layer is greater than the number of convolution kernels in the second convolutional layer, and the number of convolution kernels in the second convolutional layer is greater than the number of convolution kernels in the first convolutional layer.

4. The method according to claim 2, characterized in that, The process of obtaining the multidimensional feature vector specifically includes: The process is performed pixel by pixel, and the multidimensional feature vector is obtained based on the relationship between the feature vector of each dimension and the reflectivity tensor.

5. The method according to claim 1, characterized in that, For the full-image feature matrix, principal component analysis is used to perform feature dimensionality reduction based on the cumulative variance contribution rate, including: The full-image feature matrix is ​​centered by subtracting the global mean of the corresponding dimension from each feature dimension; Calculate the covariance matrix between each feature dimension; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and corresponding eigenvectors. The variance contribution rates are accumulated sequentially from largest to smallest eigenvalues, and the top M principal components whose cumulative variance contribution rates reach a preset threshold are selected. The original multi-dimensional features are mapped to M-dimensional features by constructing a projection matrix using the feature vectors corresponding to the selected M principal components, and the reduced feature matrix is ​​output. Where M takes the value 3.

6. The method according to claim 1, characterized in that, Based on the dimensionality-reduced feature matrix, the optimal number of clusters K is determined, including: Traverse the range of cluster numbers [3, 8]; For each cluster value, multiple cluster centers are randomly initialized, and clustering is performed separately for each cluster. Calculate the average silhouette coefficient for each clustering result to obtain the silhouette coefficient curve; If the slope change of adjacent points is greater than or equal to the slope threshold, then the profile coefficient curve is determined to have an inflection point. The point corresponding to the maximum slope change of adjacent points is taken as the inflection point, and the clustering value corresponding to the inflection point of the profile coefficient curve is selected as the optimal number of clusters K. Otherwise, select the optimal number of clusters K, where the silhouette coefficient is greater than the coefficient threshold and is a local maximum.

7. The method according to claim 6, characterized in that, Perform feature matrix clustering to obtain cluster center coordinates and generate carbon layer label maps, specifically: Iterative clustering is performed based on the determined optimal number of clusters K, and K pixels are selected as the initial cluster centers; Calculate the Euclidean distance from each pixel in the dimensionality-reduced feature matrix to the K cluster centers, and assign each pixel to the category of the nearest cluster center; Calculate the mean of all pixels in each cluster in the reduced feature space, and update the cluster center coordinates; When the distance the cluster center coordinates move in this round is less than the distance threshold, or the number of iterations reaches the upper limit, the iteration stops, and a carbon layer label map is generated based on the updated cluster center coordinates.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • A method and system for determining forest carbon storage by fusing multi-source data of forest land

    CN116433748B

  • Identification method of soil organic carbon multi-level management unit and related device

    CN120298904A

  • Carbon sink calculation method and device based on multi-dimensional carbon layer division

    CN122366879A