Methods and Systems for Predicting Carbon Sequestration Potential of Mangrove Vegetation

By constructing a dynamic coefficient mapping model set and combining multi-source remote sensing data with field sampling data, the uncertainty problem in mangrove carbon storage assessment was solved, achieving high-precision prediction of carbon sequestration potential and large-scale continuous monitoring, and supporting blue carbon asset accounting.

CN122864829BActive Publication Date: 2026-11-10SOUTH CHINA NORMAL UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611348042.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-10
Estimated Expiration
2046-09-02

AI Technical Summary

Technical Problem

When assessing carbon storage in mangroves, existing technologies are limited by traditional ecological surveys, which are subject to the constraints of accessibility and static models that ignore differences in environmental factors. This results in uncertain carbon storage assessments, making it impossible to accurately extrapolate carbon sequestration potential and meet the needs of large-scale blue carbon asset accounting.

Method used

By integrating multi-source Earth observation data with field vegetation sampling data, a dynamic coefficient mapping model group was constructed to generate spatial distribution results of aboveground biomass carbon density of vegetation. Combined with regional canopy inversion data, the carbon sequestration potential of mangrove vegetation was inferred.

Benefits of technology

It improves the accuracy of carbon density measurement, realizes the transformation from sparse static observation to large-scale continuous dynamic monitoring, reduces manpower and time costs, provides high-precision carbon sequestration potential prediction results, and supports blue carbon asset accounting and routine monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122864829B_ABST
    Figure CN122864829B_ABST
Patent Text Reader

Abstract

This application relates to a method and system for predicting the carbon sequestration potential of mangrove vegetation, belonging to the field of multi-source remote sensing monitoring technology. It acquires on-site data from sampling units in the target area, as well as Earth observation data from UAV lidar and optical radar remote sensing images, forming a cross-modal fusion representation system of refined vegetation scene 3D structural features, radar scattering features, and spectral features. By constructing a set of coefficient mapping models for the dynamic equation of vegetation growth, it outputs dynamic proportional coefficients and dynamic growth indices adapted to local site habitat conditions to replace fixed parameters. Combined with large-scale remote sensing inversion of diameter at breast height (DBH) and tree height, it achieves carbon density estimation of the dynamic vegetation growth equation. The method introduces a multi-source remote sensing feature enhancement strategy with adaptive weighting based on species similarity to improve model robustness, and combines forest age and climate factors to quantify future carbon sequestration potential, solving the systematic bias problem caused by the neglect of tidal flat habitat heterogeneity in traditional allometric parameter models.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of cross-integration technology of multi-source remote sensing monitoring and machine learning, and in particular to a method and system for predicting the carbon sequestration potential of mangrove vegetation. Background Technology

[0002] Aboveground biomass carbon density is a core indicator for assessing mangrove carbon storage and extrapolating long-term carbon sequestration potential. For a long time, research and operational work in this field have relied primarily on traditional ecological survey methods. Because mangroves grow in complex intertidal environments, field surveys are difficult and risky. Traditional plot surveys are often limited by accessibility, allowing only a small number of sampling points in localized areas, resulting in sparse data that fails to objectively reflect the actual vegetation growth across the entire region. Meanwhile, existing carbon density retrieval models often employ allometric growth equations based on a few static parameters. These models treat mangroves as homogeneous, ignoring the significant spatial differences in environmental factors such as tides, salinity, and sedimentary physicochemical properties that affect tree growth. This calculation method, based on limited samples and static models, often leads to significant uncertainty in mangrove carbon storage assessments, failing to support accurate extrapolation of regional carbon sequestration potential and providing continuous, detailed spatial data support for large-scale blue carbon asset accounting. This severely restricts the assessment of mangrove ecological value, the evaluation of long-term carbon sequestration potential, and the implementation of routine monitoring operations. Summary of the Invention

[0003] Based on this, the purpose of this application is to provide a method for predicting the carbon sequestration potential of mangrove vegetation. By integrating multi-source Earth observation data and field vegetation sampling data, a dynamic coefficient mapping model group is trained to generate the spatial distribution of aboveground biomass carbon density of vegetation in the target area at a reference time. Furthermore, the method extrapolates the increase in aboveground biomass carbon density of vegetation between a future preset time period and the reference time, thereby generating the predicted carbon sequestration potential of mangrove vegetation in the target area.

[0004] The method for predicting the carbon sequestration potential of mangrove vegetation described in this application includes the following steps: The study acquires historical characteristic data of multi-source Earth observation of the target mangrove area, as well as field vegetation sampling data of several sampling units within the target area. The field vegetation sampling data includes measured diameter at breast height (DBH), measured tree height, and measured aboveground biomass carbon density of each sampling unit. Based on the historical feature data of multi-source Earth observation corresponding to each sampling unit, a training feature vector corresponding to each sampling unit is constructed; using the training feature vector as input, the corresponding dynamic coefficients are output through the dynamic coefficient mapping model group to be trained; based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height and measured tree height, the predicted data of vegetation aboveground biomass carbon density is calculated through the dynamic vegetation growth estimation equation; based on the predicted data of vegetation aboveground biomass carbon density and the measured data of vegetation aboveground biomass carbon density, a loss function is constructed, and the model parameters of the dynamic coefficient mapping model group are iteratively optimized according to the loss function to obtain the trained dynamic coefficient mapping model group; The multi-source Earth observation historical feature data of each grid cell in the target area at the reference time are input into the trained dynamic coefficient mapping model group to obtain the dynamic coefficients corresponding to each grid cell; based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height and regional canopy inversion tree height, the vegetation aboveground biomass carbon density of each grid cell at the reference time is calculated. Based on the aboveground biomass carbon density of vegetation in each grid cell at the baseline time, the regional carbon sequestration increment is extrapolated, and the prediction results of the vegetation carbon sequestration potential of the target area are generated.

[0005] This application's embodiments construct a dynamic coefficient mapping model set by introducing historical characteristic data from multi-source Earth observations. This breaks through the limitations of the traditional allometric growth equation with globally fixed parameters, enabling the model to adaptively generate matching dynamic coefficients based on the complex habitat differences in tidal flat areas. This significantly reduces systematic estimation bias caused by ignoring spatial heterogeneity, improves the accuracy of carbon density measurement, and provides high-precision basic data for subsequent regional carbon sequestration potential projection, meeting the high standards required for blue carbon asset accounting and carbon sequestration potential assessment. Simultaneously, by combining the trained dynamic coefficient mapping model set with regional canopy inversion of diameter at breast height (DBH) and tree height data, the estimation model, which could only be verified at limited sampling points, is successfully extended to every grid cell in the target area. This effectively overcomes the shortcomings of insufficient spatial coverage in manual plot surveys, realizing the transformation from sparse static observation to large-scale continuous dynamic monitoring. It generates complete and seamless spatial distribution results of carbon density, and based on these distribution results, complete regional component calculations of carbon sequestration potential at the global scale are completed, outputting complete vegetation carbon sequestration potential prediction results. Furthermore, because the model training process fully utilizes the fusion advantages of multi-source remote sensing features and a small amount of field sampling data, it significantly reduces the reliance on high-density manual field sampling, lowers manpower and time costs, and improves the efficiency and timeliness of mangrove carbon sink monitoring. This provides reliable data support and technical assurance for the routine management of large-scale mangrove resources, the prediction and analysis of long-term carbon sequestration potential, and blue carbon trading.

[0006] To better understand and implement this application, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the method for predicting the carbon sequestration potential of mangrove vegetation according to an embodiment of this application. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. Wherein, when the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements.

[0009] It should be understood that the embodiments described below do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0010] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, in the description of this application, unless otherwise stated, “a plurality” means two or more. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items, for example, A and / or B, which can represent: A alone, A and B together, and B alone; the character “ / ” generally indicates that the preceding and following objects are in an “or” relationship.

[0011] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, this information should not be limited to these terms, and these terms are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. Depending on the context, the word "if" as used in this application can be interpreted as "when," "when," or "in response to determination."

[0012] This application relates to the scenario of detecting aboveground biomass carbon density in mangrove vegetation. As a unique blue carbon ecosystem in the intertidal zone, mangroves face several unique bottlenecks compared to terrestrial forests: First, the tidal and muddy environment restricts comprehensive ground sampling, making traditional fixed-parameter allometric equations difficult to adapt to habitat heterogeneity caused by salinity and tidal levels; second, coastal areas often experience cloud and rain combined with tidal dynamics, resulting in missing optical time-series data and significant radar saturation issues; third, the mixed growth of multiple mangrove species and severe community fragmentation hinder the cross-regional generalization ability of common forest models; fourth, carbon accumulation is simultaneously driven by temperature, precipitation, and sea-level rise, and is subject to unique disturbances such as aquaculture encroachment and coastal erosion, making existing terrestrial vegetation estimation systems unable to accurately quantify its blue carbon dynamics and long-term carbon sequestration potential. To address these issues, this application provides several embodiments.

[0013] Please refer to Figure 1 The method for predicting the carbon sequestration potential of mangrove vegetation as described in this application includes the following steps: S101: Obtain historical characteristic data of multi-source Earth observation of the target area of ​​mangrove forest, as well as field vegetation sampling data of several sampling units within the target area; the field vegetation sampling data includes the measured diameter at breast height, measured tree height and measured data of aboveground biomass carbon density of each sampling unit. S102: Based on the historical feature data of multi-source Earth observation corresponding to each sampling unit, construct the training feature vector corresponding to each sampling unit; using the training feature vector as input, output the corresponding dynamic coefficients through the dynamic coefficient mapping model group to be trained; based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height and measured tree height, calculate the predicted data of vegetation aboveground biomass carbon density through the dynamic vegetation growth estimation equation; construct the predicted data of vegetation aboveground biomass carbon density and the measured data of vegetation aboveground biomass carbon density, and iteratively optimize the model parameters of the dynamic coefficient mapping model group according to the loss function to obtain the trained dynamic coefficient mapping model group; S103: Input the multi-source Earth observation historical feature data of each grid cell in the target area at the reference time into the trained dynamic coefficient mapping model group to obtain the dynamic coefficients corresponding to each grid cell; based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height and regional canopy inversion tree height, calculate the vegetation aboveground biomass carbon density of each grid cell at the reference time. S104: Based on the aboveground biomass carbon density of vegetation in each grid cell at the baseline time, extrapolate the regional carbon sequestration increment and generate the prediction results of the vegetation carbon sequestration potential of the target area.

[0014] This application's embodiments construct a dynamic coefficient mapping model set by introducing historical characteristic data from multi-source Earth observations. This breaks through the limitations of the traditional allometric growth equation with globally fixed parameters, enabling the model to adaptively generate matching dynamic coefficients based on the complex habitat differences in tidal flat areas. This significantly reduces systematic estimation bias caused by ignoring spatial heterogeneity, improves the accuracy of carbon density measurement, and provides high-precision basic data for subsequent regional carbon sequestration potential projection, meeting the high standards required for blue carbon asset accounting and carbon sequestration potential assessment. Simultaneously, by combining the trained dynamic coefficient mapping model set with regional canopy inversion of diameter at breast height (DBH) and tree height data, the estimation model, which could only be verified at limited sampling points, is successfully extended to every grid cell in the target area. This effectively overcomes the shortcomings of insufficient spatial coverage in manual plot surveys, realizing the transformation from sparse static observation to large-scale continuous dynamic monitoring. It generates complete and seamless spatial distribution results of carbon density, and based on these distribution results, complete regional component calculations of carbon sequestration potential at the global scale are completed, outputting complete vegetation carbon sequestration potential prediction results. Furthermore, because the model training process fully utilizes the fusion advantages of multi-source remote sensing features and a small amount of field sampling data, it significantly reduces the reliance on high-density manual field sampling, lowers manpower and time costs, and improves the efficiency and timeliness of mangrove carbon sink monitoring. This provides reliable data support and technical assurance for the routine management of large-scale mangrove resources, the prediction and analysis of long-term carbon sequestration potential, and blue carbon trading.

[0015] The method for predicting the carbon sequestration potential of mangrove vegetation described in this application uses a computer as the execution subject. The following provides a detailed description of each step.

[0016] For step S101, acquire multi-source Earth observation historical characteristic data of the mangrove target area, as well as field vegetation sampling data of several sampling units within the target area; the field vegetation sampling data includes measured diameter at breast height, measured tree height, and measured aboveground biomass carbon density of each sampling unit.

[0017] Among them, multi-source Earth observation historical characteristic data refers to a multi-dimensional dataset that can reflect the growth environment and vegetation status of mangroves, obtained through satellite remote sensing, airborne remote sensing and near-ground sensing equipment. This includes, but is not limited to, the spectral reflectance characteristics of optical images, the backscattering coefficient characteristics of synthetic aperture radar images, the vegetation vertical structure characteristics extracted from lidar point clouds, and the topographic and hydrological characteristics represented by digital elevation models.

[0018] A sampling unit refers to a field survey unit established within the target mangrove area using stratified or systematic sampling methods. These units typically take the form of a fixed-area quadrat, transect, or single-tree survey point. Within each sampling unit, measured data on diameter at breast height (DBH), tree height, and aboveground biomass carbon density are obtained through manual field measurements. These data serve as training labels and accuracy verification benchmarks for supervised learning models. The number of sampling units is usually limited by manpower, resources, and accessibility, resulting in a sparse and discrete spatial distribution that cannot cover every location within the target area.

[0019] This step first acquires historical multi-source Earth observation data of the target mangrove area, as well as in-situ vegetation sampling data from several sampling units within the target area. Specifically, historical multi-source Earth observation data can be obtained by extracting mangrove vegetation indices (such as NDVI and EVI) from Sentinel-2 multispectral images, acquiring canopy roughness and water content information using Sentinel-1 synthetic aperture radar images, and obtaining average stand height and canopy volume using airborne LiDAR. In-situ vegetation sampling data is obtained during the mangrove growing season by establishing several representative plots within the target area using stratified sampling. The diameter at breast height (DBH) and tree height of individual mangrove trees within each plot are measured, and their dry weight is determined using the whole-tree harvest method or the standard tree method, thereby converting the measured data of aboveground biomass carbon density. These measured data serve as labels for supervised learning, used for subsequent model training and calibration.

[0020] In one embodiment, the multi-source Earth observation historical feature data includes three-dimensional structural features and remote sensing environmental features; the remote sensing environmental features include radar scattering features and spectral features.

[0021] Three-dimensional structural features refer to the characteristic parameters extracted from LiDAR point cloud data that can reflect the vertical structure and canopy morphology of mangroves, including but not limited to average stand height, canopy height model (CHM), canopy volume, leaf area index, and vertical profile distribution, which are used to characterize the biophysical structural information of vegetation.

[0022] Remote sensing environmental characteristics refer to a comprehensive set of features that reflect the background of mangrove growth environment obtained through remote sensing methods such as microwave remote sensing and optical remote sensing. These features include radar scattering characteristics and spectral characteristics, and are used to characterize the physical habitat and physiological state of the vegetation.

[0023] Radar scattering characteristics refer to the characteristic parameters obtained by radiometric calibration and polarization decomposition of synthetic aperture radar (SAR) image data, including backscattering coefficients (VV, VH, HH, HV polarization), coherence coefficient, polarization entropy, and polarization angle, which are used to reflect the geometric structure, water content, and surface roughness information of vegetation canopy.

[0024] Spectral characteristics refer to the electromagnetic spectrum response characteristics calculated using multispectral or hyperspectral optical image data, including reflectance in the visible, near-infrared and short-wave infrared bands, as well as various vegetation indices obtained based on multi-band combination calculations, which are used to reflect the pigment content, water stress and biomass level of vegetation.

[0025] Step S101, which involves obtaining historical feature data from multi-source Earth observations of the mangrove target area, includes: Step S1011: Acquire UAV lidar point cloud data, radar satellite time-series image data, and optical satellite time-series image data of the mangrove target area.

[0026] Among them, UAV lidar point cloud data refers to a three-dimensional spatial coordinate dataset collected by emitting laser pulses and receiving echo signals through lidar sensors mounted on UAV platforms. Each point contains X, Y, Z coordinates and intensity information, which can be used to accurately reconstruct the three-dimensional structural model of mangroves after processing.

[0027] Radar satellite time-series image data refers to a collection of synthetic aperture radar images that are repeatedly observed over a long time series of the same mangrove area. By stacking and analyzing multi-temporal data, the effects of cloud and fog obstruction can be effectively removed and the dynamic characteristics of vegetation changes over time can be captured.

[0028] Optical satellite time-series image data refers to multi-temporal optical remote sensing image data acquired continuously during the growing season. Time series analysis can eliminate the random errors of a single observation and extract vegetation phenological characteristics.

[0029] In this step, UAV lidar point cloud data, radar satellite time-series image data, and optical satellite time-series image data of the mangrove target area are acquired. Considering that mangroves grow in the intertidal zone and have a complex canopy structure, a single data source is insufficient to fully characterize their growth status. Therefore, this embodiment preferably adopts a multi-platform collaborative observation strategy. Specifically, the UAV lidar acquires high-density point clouds through terrain-following flight; the radar satellite time-series image data preferably uses Sentinel-1 C-band dual-polarization data; and the optical satellite time-series image data preferably uses Sentinel-2 MSI multispectral data, selecting continuous time-series images covering the peak growing season.

[0030] Step S1012: Feature extraction is performed on the UAV lidar point cloud data, the radar satellite time-series image data, and the optical satellite time-series image data to obtain three-dimensional structural features, radar scattering features, and spectral features.

[0031] In one specific embodiment, for UAV lidar point cloud data, denoising, filtering, and classification are first performed to separate ground points from vegetation points. Then, three-dimensional structural features, such as average tree height, maximum tree height, and canopy coverage, are extracted based on the canopy height model. For radar satellite time-series image data, after orbit correction, thermal noise removal, radiometric calibration, and terrain correction, the backscattering coefficients of VH and VV polarizations and their ratio are extracted as radar scattering features. For optical satellite time-series image data, after atmospheric correction and cloud masking, the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Red Edge Position Index (REP) are calculated as spectral features.

[0032] Step S1013: Based on the three-dimensional structural features, the radar scattering features, and the spectral features, obtain the multi-source Earth observation historical feature data of the mangrove target area.

[0033] The extracted three-dimensional structural features, radar scattering features, and spectral features are spatially registered and fused, and a unified projection coordinate system is used and resampled to the same spatial resolution (e.g., 5m×5m). Finally, a multi-source Earth observation historical feature dataset containing multi-dimensional attributes is constructed.

[0034] This embodiment constructs a comprehensive mangrove habitat characterization system by integrating the three-dimensional structural features of UAV lidar, the scattering characteristics of radar satellites, and the spectral features of optical satellites, effectively solving the problem of information loss from a single data source in complex tidal flat environments. Utilizing the high-precision structural information from UAV lidar to constrain the dynamic coefficients of the allometric growth equation significantly improves the model's adaptability to the unique growth forms of mangroves, thereby greatly reducing systematic biases caused by habitat differences. Simultaneously, multi-dimensional feature fusion based on time-series imagery enhances the model's ability to perceive vegetation physiological states, ensuring that the generated carbon density spatial distribution results are not only numerically accurate but also spatially consistent with the distribution patterns of mangroves along tidal gradients. This provides a practical and feasible technical path for the routine and high-precision monitoring of large-scale mangrove resources.

[0035] In one embodiment, step S1012 involves extracting features from the UAV lidar point cloud data, the radar satellite time-series image data, and the optical satellite time-series image data to obtain three-dimensional structural features, radar scattering features, and spectral features, respectively. Step S10121: Perform point cloud filtering and canopy height model calculation on the UAV lidar point cloud data to extract canopy complexity and tree height-to-canopy ratio; obtain three-dimensional structural data based on canopy complexity and tree height-to-canopy ratio.

[0036] Point cloud filtering refers to the process of classifying and removing noise from raw lidar point cloud data using mathematical morphology or slope-based algorithms. It aims to separate ground points from non-ground points (such as vegetation and buildings) and is the basis for generating digital elevation models (DEMs) and digital surface models (DSMs).

[0037] The canopy height model (CHM), also known as the normalized digital surface model, is obtained by subtracting the digital elevation model (DEM) from the digital surface model (DSM). It is used to represent the actual vertical height of vegetation from the ground and is a key intermediate product for extracting the height structure of forest stands.

[0038] Canopy complexity refers to a statistical measure that describes the heterogeneity of the vertical structure and the distribution characteristics of voids in mangrove canopies. It is usually obtained by calculating the height percentiles (such as P25, P50, P75, P95) or variance of lidar point clouds, and reflects the vertical stratification of the stand and the degree of biomass accumulation.

[0039] The tree height-to-canopy ratio refers to the ratio of the average tree height of a single tree or stand in a mangrove forest to its canopy projection radius (or canopy diameter). It is used to characterize the compactness or openness of the plant's morphology and is an effective indicator for distinguishing different types of mangrove communities.

[0040] This step first performs point cloud filtering on the UAV lidar point cloud data, using a cloth-based simulated filtering algorithm to separate ground points from vegetation points, generating a digital elevation model (DEM) and a digital surface model (DSM). Then, the canopy height model (CHM) is calculated through interpolation. Subsequently, the canopy complexity is quantified based on the height percentiles (e.g., P95 represents the maximum tree height) and height standard deviation within the CHM statistical units, and the ratio of average tree height to canopy projection area is calculated to obtain the tree height-to-canopy ratio. Finally, the canopy complexity and tree height-to-canopy ratio are integrated into three-dimensional structural data, which effectively preserves the vertical structural heterogeneity information of mangroves.

[0041] Step S10122: Perform speckle filtering and terrain correction processing on the radar satellite time-series image data, extract VV polarization backscattering values ​​and VH polarization backscattering values, and calculate the derived radar index based on the two types of backscattering values; obtain radar scattering data based on VV polarization backscattering values, VH polarization backscattering values ​​and derived radar index.

[0042] Speckle filtering is a process to suppress the inherent speckle noise in synthetic aperture radar (SAR) images. It typically employs algorithms such as Lee filtering, Frost filtering, or Gamma MAP filtering to smooth noise while preserving edge texture information.

[0043] Terrain correction processing is a process of correcting the radar backscattering coefficient due to the influence of terrain slope. By introducing digital elevation model (DEM) data, it eliminates the changes in radar cross section (RCS) caused by changes in slope aspect and slope, ensuring that the backscattering characteristics truly reflect vegetation attributes.

[0044] The derived radar index refers to the ratio or difference index obtained by mathematical calculation based on the backscattering coefficients of different polarization modes, such as the VH / VV polarization ratio, which is used to enhance the sensitivity to vegetation water content, canopy structure or surface roughness.

[0045] This step performs speckle filtering and terrain correction on radar satellite time-series image data. The Gamma MAP filtering algorithm is preferred to suppress speckle noise, and the terrain effect is eliminated using SRTM DEM data based on a cosine correction model. On this basis, VV polarization backscattering values ​​and VH polarization backscattering values ​​are extracted, and the ratio of VH to VV is further calculated as a derived radar index (i.e., cross-polarization ratio). This index has high sensitivity to the moisture content and structural complexity of the vegetation canopy. The VV polarization backscattering values, VH polarization backscattering values, and derived radar index are combined to form radar scattering data, comprehensively characterizing the interaction mechanism between microwaves and mangrove canopy.

[0046] Step S10123: Perform cloud masking on the optical satellite time-series image data, filter clear sky observation data within the specified key phenological period, extract multi-band reflectance values ​​by band, calculate the normalized vegetation index based on the multi-band reflectance values, and obtain clear sky spectral data based on the multi-band reflectance and the normalized vegetation index.

[0047] Cloud masking refers to the process of using the quality assessment band (QA Band) inherent in optical images or cloud detection algorithms based on spectral features to identify and mask cloud layers and cloud shadow coverage areas in images, ensuring that the extracted spectral features are not affected by atmospheric interference.

[0048] The critical phenological period refers to the time when mangrove vegetation grows most vigorously, leaf biomass is at its maximum, and spectral signals are most stable (usually the growing season). Image data acquired during this period best represents the annual carbon storage level of mangroves. In this embodiment, the critical phenological period refers to DOY180 to DOY300.

[0049] This step performs cloud masking on the time-series optical satellite imagery data, using the SceneClassification Layer band of Sentinel-2 imagery to identify and remove cloud, cloud shadow, and water body pixels, and to select clear-sky observation data within a specified key phenological period (such as the autumn low tide period). Subsequently, multi-band reflectance values, including blue, green, red, near-infrared, and short-wave infrared, are extracted by band, and the Normalized Difference Vegetation Index (NDVI) is calculated based on the sum of the reflectance differences between the near-infrared and red bands. The multi-band reflectance values ​​and the NDVI are then integrated into clear-sky spectral data to ensure that the spectral characteristics truly reflect the chlorophyll content and biophysical state of the vegetation.

[0050] Step S10124: Spatial alignment is performed on the three-dimensional structural data, the radar scattering data, and the clear-sky spectral data; using the specified key phenological period as a time filtering window, the time dimension of the spatially aligned three-dimensional structural data, the radar scattering data, and the clear-sky spectral data is matched to construct a spatiotemporally aligned data cube.

[0051] Spatiotemporally aligned data cubes refer to three-dimensional or multi-dimensional data arrays formed by stacking multidimensional datasets from different sources, with different resolutions and different acquisition times under a unified geographic coordinate system through spatial resampling and temporal window matching. This facilitates batch feature retrieval and fusion calculations by machine learning models.

[0052] This step employs a rigorous data registration process, spatially aligning the 3D structural data, radar scattering data, and clear-sky spectral data. All data are resampled to a uniform 10-meter resolution and transformed to the same projected coordinate system. Simultaneously, a specified key phenological period is used as a time filtering window to eliminate radar or optical time-series data outside this window, ensuring consistency of multi-source data across time and thus constructing a spatiotemporally aligned data cube. This data cube not only solves the challenge of fusing heterogeneous multi-source data but also eliminates seasonal biases introduced by differences in observation time.

[0053] Step S10125: Fuse the three-dimensional structural data, radar scattering data, and clear-sky spectral data within the spatiotemporal aligned data cube to generate the three-dimensional structural features, radar scattering features, and spectral features corresponding to each space in each time filtering window.

[0054] This step integrates 3D structural data, radar scattering data, and clear-sky spectral data within the spatiotemporally aligned data cube. Feature vectors from different spatial dimensions are concatenated or stacked to generate the final input model's 3D structural features, radar scattering features, and spectral features. This deeply fused feature set simultaneously encompasses the vegetation's physical framework (3D structure), microwave response (radar scattering), and optical response (spectrum), providing rich and robust input information for subsequent dynamic coefficient mapping model sets, significantly enhancing the model's ability to analyze complex tidal flat habitats.

[0055] This embodiment performs refined feature engineering on UAV lidar, radar satellite, and optical satellite data to extract key parameters such as canopy complexity, polarimetric scattering ratio, and phenological spectrum, and constructs a spatiotemporally aligned data cube, fundamentally solving the problem of inconsistent spatiotemporal references in multi-source remote sensing data. This multi-dimensional feature fusion strategy enables the dynamic coefficient mapping model set to simultaneously capture the comprehensive "morphological-structural-environmental" information of mangroves, effectively eliminating the saturation effect and estimation bias caused by traditional methods that rely solely on spectral or altitude information, and significantly improving the accuracy and stability of mangrove aboveground biomass carbon density detection.

[0056] In one embodiment, step S10125, which involves fusing 3D structural data, radar scattering data, and clear-sky spectral data within a spatiotemporally aligned data cube to generate 3D structural features, radar scattering features, and spectral features corresponding to each space within each time filtering window, includes: Step S101251: Extract the full-time radar scattering data and full-time clear-sky spectral data of each grid cell in the spatiotemporal aligned data cube.

[0057] Full-time radar backscattering data refers to a sequence of radar satellite images acquired continuously at fixed time intervals within a specific time window (such as a year or a growing season). It includes the microwave backscattering response of mangroves at multiple time points and is used to capture the dynamic characteristics of vegetation changes with tidal cycles and seasons. In this embodiment, full-time refers to data covering one year.

[0058] Full-time clear-sky spectral data refers to a series of multi-temporal optical remote sensing images acquired continuously during key phenological periods after cloud masking, reflecting the trend of spectral reflectance changes in mangrove vegetation throughout its entire growth cycle.

[0059] This step extracts the full-time radar scattering data and full-time clear-sky spectral data corresponding to each grid cell from the spatiotemporally aligned data cube. Unlike traditional methods that only use the average value of single or several periods of imagery, this embodiment utilizes a continuous time series covering the entire key phenological period, thereby preserving the signal fluctuation information of mangroves caused by tidal flooding, seasonal leaf change, and community succession.

[0060] Step S101252: Based on the full-time radar scattering data and full-time clear sky spectral data corresponding to each grid cell, extract several types of time-series statistical features and several types of time period features respectively to generate a time-series statistical feature set and a harmonic phenological parameter set; construct a spatiotemporal phenological spectrum feature according to the time-series statistical feature set and the harmonic phenological parameter set.

[0061] Temporal statistical features refer to aggregated features obtained by statistical calculations on time series data. They are used to summarize the distribution characteristics of data in the time dimension. Common features include mean, standard deviation, maximum value, minimum value, amplitude, and skewness, which can effectively suppress random noise from single-phase observations. In one embodiment, for the full-time radar scattering data and full-time clear-sky spectral data corresponding to each grid cell, the range, Q10 percentile, Q90 percentile, standard deviation, and skewness are calculated respectively to generate temporal statistical features.

[0062] Time-series characteristics refer to the features obtained by segmenting a time series according to a specific time interval (such as a month, quarter, or tidal cycle) and using them to reflect the differences in the response of vegetation at different growth stages or under environmental conditions.

[0063] Harmonic phenological parameter sets refer to the set of parameters extracted from time-series data using Fourier transform or wavelet transform to describe the phenological rhythms of vegetation. These parameters typically include fundamental amplitude, second harmonic amplitude, phase angle, and goodness-of-fit (R²), used to quantify the periodic fluctuations in vegetation growth. In one embodiment, Fourier series harmonic decomposition is performed on the full-time radar scattering data and full-time clear-sky spectral data of each grid cell to extract seasonal amplitude coefficients, semi-annual amplitude coefficients, and interannual trend terms, generating harmonic phenological parameters.

[0064] Spatiotemporal phenological spectrum features refer to composite feature vectors that integrate the statistical characteristics of time series and the characteristics of phenological rhythms. They contain both the instantaneous state information of vegetation and its long-term dynamic change patterns, serving as a bridge connecting remote sensing observation and the physiological and ecological processes of vegetation.

[0065] This step involves high-order feature mining based on full-time radar scattering data and full-time clear-sky spectral data for each grid cell. For time-series statistical features, the mean, standard deviation, coefficient of variation, and range of each sequence are calculated over time. These statistics effectively reflect the stability of mangrove canopy structure and the fluctuation range of biomass. For time-period features, the system divides the time series into dry and rainy seasons (or according to tidal cycles) and calculates the average backscattering coefficient and spectral index for different time periods to capture the impact of environmental stress on vegetation physiological state. Based on this, a harmonic analysis model is further used to fit the time series, extracting a set of harmonic phenological parameters, including the amplitude (reflecting the intensity of the annual main growth cycle) and phase (reflecting the time of growth peak occurrence) corresponding to the fundamental frequency. The time-series statistical feature set and the harmonic phenological parameter set are concatenated to form a spatiotemporal phenological spectrum feature. This feature set not only describes "how the mangroves are growing" but also reveals "how they change over time," greatly enriching the discriminative information of the model.

[0066] Step S101253: The spatiotemporal phenological spectrum features are fused with the three-dimensional structural data of the corresponding grid cell to obtain the three-dimensional structural features of the corresponding grid cell; the spatiotemporal phenological spectrum features are fused with the radar scattering data and clear sky spectral data of the corresponding grid cell to obtain the radar scattering features and spectral features of the corresponding grid cell.

[0067] Specifically, static three-dimensional structural data reflecting the vertical physical structure of vegetation (such as canopy height and complexity) are fused with time-varying characteristics—namely, spatiotemporal phenological spectral features (usually through feature vector concatenation) to generate the final three-dimensional structural features of the corresponding grid cell. Simultaneously, the spatiotemporal phenological spectral features are fused with the radar scattering data (microwave features) and clear-sky spectral data (optical features) of the corresponding grid cell to generate the radar scattering and spectral features of the corresponding grid cell. This fusion strategy is not a simple stacking; rather, it uses static structure as a constraint and dynamic phenological spectra and scattering features as independent variables, constructing a comprehensive feature set that includes spatial structure, microwave physical response, and optical physiological response.

[0068] This embodiment constructs a spatiotemporal phenological spectrum feature by introducing full-time-series data analysis and harmonic phenological parameter extraction, and deeply integrates it with three-dimensional structural data. This effectively solves the estimation bias problem caused by weather, tides, and phenological changes in traditional remote sensing monitoring due to interference from single-phase observations. This feature engineering method significantly enhances the model's ability to perceive the complex growth rhythms of mangroves, enabling the dynamic coefficient mapping model group to distinguish mangrove communities with similar appearances but different physiological states or habitat conditions, thereby greatly improving the anti-interference ability and universality of carbon density estimation.

[0069] For step S102, based on the historical feature data of multi-source Earth observation corresponding to each sampling unit, a training feature vector corresponding to each sampling unit is constructed; using the training feature vector as input, the corresponding dynamic coefficients are output through the dynamic coefficient mapping model group to be trained; based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height and measured tree height, the predicted data of aboveground biomass carbon density of vegetation is calculated through the dynamic vegetation growth estimation equation; based on the predicted data of aboveground biomass carbon density of vegetation and the measured data of aboveground biomass carbon density of vegetation, a loss function is constructed, and the model parameters of the dynamic coefficient mapping model group are iteratively optimized according to the loss function to obtain the trained dynamic coefficient mapping model group.

[0070] The dynamic coefficient mapping model set refers to a collection of at least one machine learning model used to establish a nonlinear mapping relationship between historical feature data of multi-source Earth observation and coefficients of allometric growth equations. This model set can adaptively output growth parameter correction coefficients for specific spatial units based on the input habitat characteristics.

[0071] The dynamic vegetation growth estimation equation is an improved computational model based on the traditional allometric growth equation. It replaces the original fixed coefficients with variable coefficients output by the dynamic coefficient mapping model group, so that the biomass estimation process can respond to the spatial differences of different tidal flat site conditions.

[0072] A grid cell refers to the smallest computational unit formed by dividing the target area of ​​a mangrove forest according to a preset spatial resolution (such as 5m×5m or 10m×10m). It is used to carry remote sensing inversion parameters and model prediction results, realizing spatial extrapolation from discrete points to a continuous surface. In this embodiment, the grid cell is divided into 10m×10m units.

[0073] The loss function is a mathematical expression used to quantify the difference between the model's predicted values ​​and the ground-based measured values. It is usually expressed in the form of mean squared error (MSE), mean absolute error (MAE), or Huber loss, and is used as the basis for backpropagation to optimize the model parameters.

[0074] Convergence criteria refer to the stopping criteria set during model training, which typically include one or more combinations of the following: the validation set loss no longer decreases, the maximum number of preset iterations is reached, or the gradient change is less than a set threshold.

[0075] This step generates training feature vectors based on historical feature data from multi-source Earth observations corresponding to each sampling unit. These training feature vectors are then input into the dynamic coefficient mapping model set to be trained. In this embodiment, the dynamic coefficient mapping model set can be constructed using a random forest regression model or a deep neural network model. The model takes the aforementioned multi-source features as input and outputs dynamic coefficients (e.g., the power coefficient or proportional coefficient of the diameter at breast height (DBH) term) corresponding to at least one parameter in the allometric growth equation. Next, the measured DBH and measured tree height of the sampling unit are substituted into the dynamic vegetation growth estimation equation, and the predicted aboveground biomass carbon density of the sampling point is calculated by combining the dynamic coefficients output by the model. To optimize model performance, a loss function is constructed, the deviation between the predicted data and the measured data is calculated, and backpropagation iteration is performed using stochastic gradient descent or the Adam optimizer to continuously adjust the weight parameters of the dynamic coefficient mapping model set until the loss meets the convergence condition, thereby solidifying the model parameters and obtaining the trained dynamic coefficient mapping model set. This process essentially involves the model learning the effects of different tidal flat habitats (such as flooding frequency, soil salinity, and nutrient status) on mangrove growth patterns, thereby correcting the inherent biases of traditional empirical formulas.

[0076] In one embodiment, the dynamic coefficient mapping model group is used to learn the dynamic coefficients of the dynamic vegetation growth estimation equation, the dynamic coefficients including dynamic proportionality coefficients and dynamic growth indexes; the dynamic vegetation growth estimation equation uses measured diameter at breast height (DBH) and measured tree height as basic input variables; the dynamic coefficient mapping model group is used to receive three-dimensional structural features and remote sensing environmental features as joint inputs; based on the first mapping relationship corresponding to the three-dimensional structural features and the second mapping relationship corresponding to the remote sensing environmental features, it learns together to output dynamic proportionality coefficients and dynamic growth indexes adapted to local site habitat conditions; wherein, the remote sensing environmental features include radar scattering features and spectral features.

[0077] The dynamic scaling factor refers to the basic scaling factor used in the allometric growth equation to control the proportional relationship between biomass and diameter at breast height (DBH) and tree height. In traditional fixed-parameter models, this factor is a constant. However, in this invention, the dynamic scaling factor is output in real time by the dynamic scaling factor mapping model group based on habitat characteristics, and is used to correct for differences in carbon sequestration efficiency of mangroves under different tidal flat elevations, soil salinity, or nutrient conditions.

[0078] The dynamic growth index refers to the power term in the allometric growth equation that acts on the diameter at breast height (DBH) or tree height. This index determines the rate sensitivity of biomass growth with plant size, and its dynamic adjustment can adapt to the significant differences in vertical growth strategies among different mangrove communities (such as shrub-like and tree-like).

[0079] The first mapping relationship refers to the nonlinear functional relationship specifically established in the dynamic coefficient mapping model group, used to analyze the intrinsic connection between three-dimensional structural features (such as canopy complexity and tree height-to-canopy ratio) and vegetation growth parameters. This relationship reflects the constraint effect of vegetation physical morphology on allometric growth patterns.

[0080] The second mapping relationship refers to the nonlinear functional relationship specifically established in the dynamic coefficient mapping model group, used to analyze the intrinsic connection between remote sensing environmental characteristics (radar scattering characteristics and spectral characteristics) and vegetation growth parameters. This relationship reflects the driving effect of external environmental factors such as intertidal flooding frequency, light conditions, and water stress on mangrove biomass accumulation.

[0081] Joint input refers to fusing three-dimensional structural features with remote sensing environmental features in the feature space dimension, serving as a unified input source for the dynamic coefficient mapping model group. This strategy enables the model to simultaneously consider both "what the vegetation looks like" (structure) and "what the environment is like" (environment), thereby outputting the most suitable local growth coefficient.

[0082] Local site habitat conditions refer to the comprehensive environmental constraints within a mangrove target area, with a single grid unit or sampling unit as the spatial boundary, that directly and microscopically affect vegetation growth. These conditions are not a macro-climate overview at the regional scale, but rather a local physical habitat characterized and quantified by three-dimensional structural features and remote sensing environmental features at the unit scale. Specifically, three-dimensional structural features (such as canopy complexity and tree height-to-canopy ratio) reflect the structural constraints of the mangrove community within the unit, including vertical stratification, canopy openness, and competition. Radar scattering features in the remote sensing environmental characteristics (such as VV and VH polarization backscattering coefficients and their ratios) reflect the microwave physical responses of the unit, such as canopy water content, surface roughness, and tidal flooding frequency. Spectral features (such as multi-band reflectance and normalized difference vegetation index) reflect the optical physiological responses of the unit, such as chlorophyll content, water stress, and phenological states. The combination of these multidimensional features accurately characterizes habitat heterogeneity, such as the varying duration of flooding due to differences in tidal flat elevation, the spatial differentiation of soil salinity and nutrients, and local hydrological changes caused by micro-topography. The dynamic coefficient mapping model learns the comprehensive expression of these local site habitat conditions to output dynamic proportional coefficients and dynamic growth indices adapted to the specific unit. This enables the dynamic vegetation growth estimation equation to respond to micro-habitat differences and eliminates the systematic bias caused by traditional global fixed-parameter models that ignore spatial heterogeneity.

[0083] In this embodiment, the first mapping relationship is mainly handled by the sub-network that processes three-dimensional structural features. This sub-network identifies the community type and vertical stratification structure of mangroves by analyzing the canopy complexity and the tree height-to-canopy ratio, thereby inferring the upper limit of the reasonable growth scale and morphological constraints of mangroves in the area, and outputs a preliminary coefficient adjustment signal accordingly.

[0084] The second mapping relationship is mainly handled by the sub-network that processes environmental characteristics. This sub-network identifies the site stress level of mangroves (such as growth restriction caused by frequent flooding in low-lying areas or slow growth caused by barrenness in high-lying areas) by analyzing radar scattering characteristics (reflecting canopy water content and surface roughness) and spectral characteristics (reflecting chlorophyll activity and phenological state). Based on this, it outputs environmental correction signals.

[0085] The model group fuses the two signals at the feature level and, through nonlinear transformation of fully connected layers and activation functions, finally outputs dynamic scaling coefficients and dynamic growth indices that adapt to the local site habitat conditions of the current sampling unit. This mechanism ensures that mangroves growing in low-lying, high-salinity, and heavily flooded environments are not subjected to the same growth equations as those in high-altitude, fresh-soil, and stable environments, thus eliminating the root cause of systematic bias.

[0086] Specifically, the parameters of the allometric growth equation are expressed as a function of UAV-LiDAR fine structure variables and remote sensing environmental variables:

[0087] in, This represents a 3D structural feature vector extracted from a LiDAR point cloud. Represents the remote sensing environment feature vector extracted from Sentinel1 or 2; parameter mapping relationship , The coefficients of the allometric growth equation are learned from sample data through a machine learning model. These coefficients are dynamically determined based on the spatial variations in local canopy structure complexity and remote sensing environmental characteristics. This not only effectively reflects the spatial heterogeneity within complex canopies but also provides a physically constrained dynamic physical benchmark for subsequent large-scale satellite remote sensing inversion.

[0088] Step S102, based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height (DBH) and measured tree height, calculates the aboveground biomass carbon density of vegetation using the dynamic vegetation growth estimation equation. This step includes: Step S1021: Calculate the aboveground biomass carbon density of vegetation based on the dynamic proportional coefficient, dynamic growth index, corresponding measured diameter at breast height, and corresponding measured tree height of each sampling unit.

[0089] This embodiment fundamentally reconstructs the physical model for mangrove biomass estimation by introducing the concepts of dynamic scaling factor and dynamic growth index, and constructing a dual-channel mapping mechanism based on the joint input of three-dimensional structural features and environmental features. This scheme not only retains the concise and clear mathematical form of the allometric growth equation, but also endows it with an intelligent mechanism for perceiving habitat heterogeneity through machine learning. This improvement enables carbon density estimation results to accurately respond to subtle changes in tidal flat micro-topography, effectively solving the pain point of inaccurate estimation in mangrove transition zones or mixed community areas, and significantly enhancing the scientific rigor and credibility of blue carbon measurement.

[0090] In one embodiment, step S1021, which calculates the aboveground biomass carbon density of vegetation based on the dynamic scaling factor, dynamic growth index, corresponding measured diameter at breast height (DBH), and corresponding measured tree height of each sampling unit, includes: Step S10211: For any sampling unit, the measured diameter at breast height (DBH) and measured tree height are used as the basic variables of the standing tree. The aboveground biomass is calculated by substituting them into the dynamic allometric growth formula constructed by the dynamic proportional coefficient and the dynamic growth index. Based on the aboveground biomass, the carbon density of the aboveground biomass of the vegetation is obtained by converting it using a preset carbon conversion coefficient.

[0091] The basic variables of standing trees refer to the core metrics that describe the basic growth status of a single tree or a stand per unit area. In this embodiment, they specifically refer to the measured diameter at breast height (DBH) and the measured tree height (H). These two variables are the basic independent variables for constructing the allometric growth equation and directly determine the order of magnitude of biomass estimation.

[0092] The dynamic allometric growth formula refers to the calculation expression of the traditional static allometric growth model. Its core improvement lies in replacing the original fixed proportional coefficients and fixed power exponents with dynamic proportional coefficients and dynamic growth exponents output by the dynamic coefficient mapping model group, thereby enabling the formula to respond to environmental changes and self-correct.

[0093] Aboveground biomass refers to the total dry weight of the aboveground parts of vegetation (including trunks, branches, leaves, and reproductive organs) in a terrestrial or wetland ecosystem, usually measured in kilograms (kg) or tons (t). It is a key intermediate physical quantity connecting the physical form of vegetation with carbon storage.

[0094] The preset carbon conversion factor refers to the conversion ratio determined based on experimental measurements or literature surveys, used to convert the weight of organic dry matter into the weight of carbon. Typically, the carbon conversion factor for mangrove vegetation ranges from 0.45 to 0.50 (i.e., 45%-50% of the dry weight is carbon), and is used to convert the physical mass of biomass into the carbon mass required for ecometrics.

[0095] This embodiment constructs a physically interpretable dynamic estimation model by combining the parameters output by the dynamic coefficient mapping model with classical allometric growth theory. By using dynamic proportionality coefficients and dynamic growth indices to weighted calculate measured diameter at breast height (DBH) and tree height, it not only inherits the rigor of traditional forestry surveys but also incorporates artificial intelligence's ability to perceive environmental heterogeneity. Combined with standardized conversion using preset carbon conversion coefficients, it ensures the numerical accuracy and ecological rationality of the output results, effectively solving the problem of systematic biases easily generated by traditional fixed-parameter models in complex tidal flat environments. This provides a highly reliable metrological basis for mangrove blue carbon trading.

[0096] In one embodiment, the remote sensing environmental features further include multi-density training feature samples. Considering that in actual operational use, optical imagery is easily obscured by clouds and fog, and radar data may experience transmission frame drops, the historical feature data from multi-source Earth observations input to the model often suffers from uneven density or temporal discontinuities. To improve the robustness of the dynamic coefficient mapping model group and ensure it can still output stable dynamic coefficients despite fluctuations in data quality, this embodiment introduces a training enhancement mechanism based on multi-density training feature samples. In one embodiment, the remote sensing environmental features further include the original feature samples corresponding to the multi-density training feature samples.

[0097] Multi-density training feature samples refer to multiple sets of training feature samples extracted by performing random downsampling operations of different intensities on the original full-time-series observation data of the same sampling unit. Each set of training feature samples represents the vegetation response characteristics under different data densities or observation frequencies, and is used to simulate situations where observation data is missing or of uneven quality that may occur in practical applications, thereby enhancing the model's generalization ability and anti-interference ability.

[0098] The multi-density training feature samples are obtained through the following steps: Step S201: Extract the full-time radar scattering data and full-time clear sky spectral data corresponding to each sampling unit in the spatiotemporal aligned data cube.

[0099] Step S202: Perform several preset ratio gradient random downsampling on the full-time radar scattering data and full-time clear sky spectral data of each sampling unit to obtain several sets of downsampled observation subsets corresponding to each sampling unit; wherein, the downsampled observation subsets include full-time radar scattering data and full-time clear sky spectral data.

[0100] Preset-ratio gradient random downsampling refers to the operation of randomly sampling and removing samples from a time-series dataset along the time dimension, with the removal ratio varying according to a preset gradient sequence (e.g., 10%, 30%, 50%). This operation aims to artificially construct training samples with different information densities, forcing the model to learn robust feature representations that do not depend on specific observation densities. In this embodiment, several preset ratios are 20%, 40%, 60%, and 80%.

[0101] Downsampled observation subsets refer to new time-series datasets obtained after undergoing random downsampling processing with a preset gradient. This subset retains the structural characteristics of the original time-series data, but the total amount of data or temporal resolution is reduced. It is used to simulate scenarios where data is incomplete in remote sensing observations due to weather, orbital revisit cycles, or equipment failures.

[0102] Step S203: For any sampling unit, based on the full-time radar scattering data and full-time clear sky spectral data in each set of downsampled observation subsets, extract several types of time-series statistical features and several types of time period features to generate a time-series statistical feature set and a harmonic phenological parameter set; based on the time-series statistical feature set and harmonic phenological parameter set corresponding to each set of downsampled observation subsets, generate corresponding multi-density training feature samples.

[0103] Specifically, for the i-th downsampled observation subset, based on its incomplete full-time radar scattering data and full-time clear-sky spectral data, a time-series feature mining operation similar to step S101252 is performed: time-series statistical features and time-segment features are extracted, and a time-series statistical feature set and a harmonic phenological parameter set are generated through harmonic analysis. Due to the reduced data density, the feature values ​​extracted from these subsets will differ from the original complete data (e.g., the standard deviation increases or the amplitude decreases). All features extracted based on this subset are fused to generate corresponding multi-density training feature samples. Ultimately, one sampling unit will correspond to multiple multi-density training feature samples, and each subset represents the feature performance of that sampling point under certain specific observation conditions.

[0104] This embodiment effectively addresses the issues of data loss and uneven density caused by weather, orbit, and other factors during the actual acquisition of remote sensing data by constructing multi-density training feature samples. This data augmentation strategy significantly improves the tolerance of the dynamic coefficient mapping model group to fluctuations in input data quality, preventing the model from overfitting to the perfect data distribution achieved during training. This ensures that even in operational scenarios with severe cloud cover or poor radar data quality, the model can still output high-precision dynamic coefficients based on incomplete observation data. This not only enhances the practicality and stability of the mangrove carbon density detection system but also provides reliable technical support for routine blue carbon monitoring under complex meteorological conditions.

[0105] In one embodiment, step S102, which involves constructing a loss function based on predicted aboveground biomass carbon density data and measured aboveground biomass carbon density data, includes: Step S1022: Based on the deviation between the predicted data of vegetation aboveground biomass carbon density and the measured data of vegetation aboveground biomass carbon density, construct a sample fitting loss term.

[0106] The sample fitting loss term is a quantitative indicator that measures the degree of deviation between the model's predicted values ​​and the actual ground-based measured values. It is used to drive the model to learn the main patterns in the data. The smaller the value of this loss term, the higher the model's fitting accuracy to the training samples, meaning that the predicted carbon density is closer to the actual measurement results.

[0107] Step S1023: Generate a model complexity regularization term; fuse the sample fitting loss term and the model complexity regularization term to obtain a loss function for training the dynamic coefficient mapping model group; wherein, the model complexity regularization term includes a penalty term for the number of model leaf nodes and a leaf node weight magnitude constraint term.

[0108] The model complexity regularization term is an additional term added to the loss function to penalize the structural complexity of the model. Its purpose is to prevent the model from overfitting (i.e., learning too much random noise or outliers in the training data) by constraining the model's degrees of freedom, thereby improving its generalization ability on unseen data (global grid).

[0109] The penalty term for the number of leaf nodes specifically refers to a regularization technique for tree-based machine learning models (such as random forests and XGBoost). By directly adding the total number of leaf nodes as a penalty factor to the loss function, the disordered growth of the tree model is limited, encouraging the model to use fewer branches to summarize feature patterns, thereby reducing model complexity.

[0110] The leaf node weight magnitude constraint is a regularization method (such as L1 or L2 regularization) that constrains the absolute value or square of the output value (i.e., weight) of each leaf node in a regression tree model. This constraint limits the extreme influence of a single leaf node on the final prediction result, prevents the model from producing an excessively large response to certain combinations of features, and thus improves numerical stability.

[0111] This embodiment significantly improves the generalization ability and robustness of the dynamic coefficient mapping model group by introducing a composite regularization strategy that includes penalties for the number of leaf nodes and weight magnitude constraints into the loss function. This design effectively overcomes the risk of model overfitting caused by the limited number of field-measured samples and the potential presence of outliers, ensuring that the trained model can still output dynamic coefficients that conform to ecological principles based on multi-source Earth observation characteristics when migrating to unsampled areas. This not only improves the overall accuracy and reliability of carbon density spatial distribution results but also makes the model more adaptable to mangrove data from different years and tidal flat environments, providing stable and reliable technical support for long-term monitoring of blue carbon trading.

[0112] In one embodiment, step S102, which iteratively optimizes the model parameters of the dynamic coefficient mapping model group according to the loss function to obtain the trained dynamic coefficient mapping model group, includes: Step S1024: After each iteration, the SHAP feature contribution is calculated based on all training features of the model. The training features with the highest SHAP contribution ranking are determined as high contribution features, and the remaining training features are low contribution features. This ensures that in each iteration of the model, the high contribution features are given high weights and the low contribution features are given low weights.

[0113] Feature contribution optimization (SHAP) is a cooperative game theory-based method used to measure the fairness of predictions made by machine learning models. It quantifies the importance of a feature by calculating the expected marginal contribution of a particular feature to the model's predicted output across all possible combinations of features, thus revealing the model's internal decision-making logic.

[0114] High-contribution features refer to key input features that, through SHAP value analysis, have a decisive influence or strong correlation with the dynamic coefficient output during model training. These features typically correspond to core physical parameters such as canopy height, biomass-sensitive bands, or specific polarization scattering information. In this embodiment, the top 20% of training features are selected as high-contribution features.

[0115] Low-contribution features refer to features that have a weak impact on the model's output predictions, have a high noise ratio, or are not highly relevant to the target task under the current data distribution. If these features are over-involved in training, they can easily lead to model overfitting or introduce unnecessary computational overhead.

[0116] Feature weights refer to the participation intensity or learning rate adjustment factors assigned to different features during model iterative optimization. High weights mean that the feature plays a dominant role in gradient updates, while low weights mean that the sensitivity to the feature is reduced or temporarily suppressed.

[0117] Specifically, after each round of backpropagation iteration optimization, parameter updates are paused, and a global feature importance assessment is performed. All training features in the current model state (i.e., the fused feature set generated in step S10125) are retrieved, and the SHAP interpreter is used to calculate the SHAP feature contribution of each training feature to the current batch prediction results. This calculation process iterates through all features, quantifying the absolute value of each feature's contribution to either increasing or decreasing the predicted carbon density value. Subsequently, all features are sorted from highest to lowest SHAP contribution, and the top-ranked features (e.g., the top 20% or top 30%, which can be determined based on cross-validation) are defined as high-contribution features, while the remaining features are defined as low-contribution features. After determining the high and low contribution features, the data values ​​of the features themselves are not changed; instead, the feature focus of the model in the next iteration is adjusted. In subsequent gradient descent or tree splitting processes, high-contribution features are automatically assigned high weights, such as increasing their corresponding learning rate, reducing their subsampling drop probability in the random forest, or increasing their priority as splitting nodes in the gradient boosting tree; conversely, low-contribution features are assigned low weights, such as reducing their learning rate, increasing their drop probability, or limiting their splitting gain. The guided model prioritizes learning physical features strongly correlated with mangrove growth, such as canopy height and phenological phase, using limited model capacity, while filtering out interference information from atmospheric residual noise or irrelevant bands. As iterations proceed, the mapping relationship between high-contribution features (such as canopy complexity in 3D structural features) and dynamic coefficients is continuously strengthened and refined, while the noise impact of low-contribution features is continuously suppressed. When the loss meets the convergence condition, the final trained dynamic coefficient mapping model group not only mathematically optimally fits the training data but also physically focuses on the most critical ecological driving factors.

[0118] This embodiment effectively addresses the issues of feature redundancy and noise interference in high-dimensional remote sensing feature spaces by introducing a dynamic weight adjustment strategy based on SHAP feature contribution during training. This scheme significantly improves the training efficiency and convergence stability of the dynamic coefficient mapping model group, preventing the model from getting trapped in local optima or over-relying on physically irrelevant pseudo-features. More importantly, this mechanism endows the model with strong interpretability, allowing the output dynamic scaling coefficients and dynamic growth exponents to be traced back to specific habitat-driving factors, thereby verifying the scientific validity of the carbon density estimation results.

[0119] For step S103, the multi-source Earth observation historical feature data of each grid cell in the target area at the reference time are input into the trained dynamic coefficient mapping model group to obtain the dynamic coefficients corresponding to each grid cell; based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height and regional canopy inversion tree height, the aboveground biomass carbon density of the vegetation in each grid cell is calculated.

[0120] Among them, a grid cell refers to the smallest computational unit formed by regularly or irregularly dividing the spatial range of the mangrove target area according to a preset resolution. It is the basic carrier for remote sensing feature extraction, model inference, and spatial distribution mapping. The spatial resolution (i.e., cell size) of the grid cell can be flexibly set according to actual operational needs. It can either adopt the same resolution or spatial specifications as the field survey range of the sampling unit, thereby achieving direct correspondence and seamless connection between the sampling unit and the grid cell in spatial location; or it can re-divide the target area according to the original resolution of the remote sensing data, computing resource limitations, or the accuracy requirements of blue carbon accounting, using a grid size different from the sampling unit specifications.

[0121] This step utilizes a trained dynamic coefficient mapping model set for global extrapolation. Specifically, the target area is divided into several regular grid cells according to a preset resolution, and the historical feature data of multi-source Earth observation corresponding to each grid cell are input into the trained model to obtain the dynamic coefficients specific to each grid cell. Simultaneously, remote sensing inversion methods are used to obtain the regional canopy inversion diameter at breast height (DBH) and regional canopy inversion tree height corresponding to each grid cell, such as using a machine learning model to invert the average DBH and height of the forest stand based on spectral features. Substituting the dynamic coefficients of the grid cells with their corresponding regional canopy inversion DBH and regional canopy inversion tree height into the dynamic vegetation growth estimation equation allows for the parallel calculation of the aboveground biomass carbon density of vegetation in all grid cells within the target area. This step overcomes the physical limitations of the number of manual sampling points, achieving a continuous spatial mapping from point to area.

[0122] In one embodiment, the spectral features further include the canopy moisture index. The canopy moisture index is a physiological and biochemical indicator specifically used to characterize the water content of vegetation canopies and their water stress state, calculated based on multi-band reflectance from optical remote sensing. Typical examples include the Normalized Difference Water Index (NDWI) or the Land Surface Water Index (LSWI). Its value directly reflects the water saturation of mangrove leaves and is a key parameter for assessing vegetation vitality and resilience.

[0123] Step S103 involves inputting the multi-source Earth observation historical feature data of each grid cell in the target area at the reference time into the trained dynamic coefficient mapping model group to obtain the dynamic coefficients corresponding to each grid cell. This step includes: Step S1031: Obtain mangrove species data corresponding to each grid unit in the target area and mangrove species data corresponding to the sampling unit, and calculate the mangrove species similarity between each grid unit and the sampling unit.

[0124] Mangrove species data refers to taxonomic information describing the composition of mangrove communities, including the dominant tree species (such as Kandelia candel, Avicennia marina, and Avicennia marina) and their canopy coverage ratios within each grid cell or sampling cell of the target area. This data can be obtained through visual interpretation of high-resolution remote sensing images, field biodiversity surveys, or species distribution model predictions.

[0125] Mangrove species similarity is a metric that measures the closeness of species composition between a target grid cell and a known sampling cell. It is typically calculated using ensemble similarity algorithms or based on dominant species matching. The value ranges from 0 to 1, with higher values ​​indicating greater similarity in community structure and biological characteristics.

[0126] Step S1032: When the similarity of mangrove species between any grid cell and the sampling cell is greater than a preset threshold, the multi-source Earth observation historical feature data of the corresponding grid cell is input into the dynamic coefficient mapping model group, and the corresponding dynamic coefficient is output.

[0127] This step sets a preset threshold (e.g., 0.8) based on the actual situation. When the mangrove species similarity between any grid cell and at least one sampling cell is greater than this preset threshold, the grid cell is determined to belong to a "known ecotype". In this case, the original multi-source Earth observation historical characteristic data (including canopy moisture index) of the grid cell are directly input into the dynamic coefficient mapping model group. The model uses the general mapping relationship learned during the training phase to directly output the corresponding dynamic coefficients. This situation indicates that the model has sufficient prior knowledge to handle this type of habitat.

[0128] Step S1033: When the mangrove species similarity between any grid cell and the sampling cell is less than or equal to a preset threshold, based on the regional dominant mangrove species, adjust the feature weight of the canopy water index in the spectral features of the multi-source Earth observation historical feature data of the corresponding grid cell, obtain the adjusted multi-source Earth observation historical feature data of the grid cell, input it into the dynamic coefficient mapping model group, and output the corresponding dynamic coefficient.

[0129] Regionally dominant mangrove species refer to the mangrove plant species that are most widely distributed, have the highest biomass, or play a leading role in ecosystem function within the target area. When specific species data are lacking for a particular grid cell, the ecological characteristics of the dominant species are used as a proxy by default.

[0130] Feature weights refer to the participation intensity coefficients assigned to different feature dimensions during multi-source feature fusion or model inference. Adjusting the feature weights of the canopy water index changes the contribution ratio of this index in the calculation process of the dynamic coefficient mapping model group, in order to adapt to the physiological response mechanisms specific to different species.

[0131] In this step, when the similarity of mangrove species between any grid cell and all sampling cells is less than or equal to a preset threshold, the grid cell is determined to belong to an "unknown or rare ecotype" (e.g., the presence of invasive species or a rare stage of community succession). In this case, directly using the original features may lead to estimation distortion. Therefore, adjustments are made based on the ecological habits of the dominant mangrove species in the region: if the dominant species is drought-resistant but its canopy moisture index is abnormally high (possibly due to tidal intrusion), the feature weight of the canopy moisture index in the spectral features of that grid cell is reduced to prevent the model from misjudging it as being in a high-productivity state; conversely, if the dominant species is a moisture-loving species but its moisture index is low, the weight of this feature can be appropriately increased to emphasize its water stress signal. This weight adjustment can be achieved by modifying the numerical scaling factor of the feature vector or adjusting the connection weights of the model's input layer. Subsequently, the adjusted multi-source Earth observation historical feature data is input into the model, and dynamic coefficients adapted to this specific habitat are output.

[0132] This embodiment effectively addresses the problem of decreased model generalization ability caused by differences in species composition in cross-regional or large-scale applications by introducing an adaptive adjustment mechanism for mangrove species similarity discrimination and canopy moisture index feature weights. This scheme utilizes species similarity as prior knowledge to guide the model in intelligently correcting the confidence level of key physiological indicators (canopy moisture) when dealing with unseen vegetation types, thereby avoiding systematic estimation biases caused by differences in species physiological characteristics. This not only significantly improves the robustness of the dynamic coefficient mapping model group under complex community structures but also enhances the rationality of carbon density spatial distribution results in ecological zoning.

[0133] In one embodiment, before step S103, which calculates the aboveground biomass carbon density of each grid cell at a reference time based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height and regional canopy inversion tree height, the following steps are included: Step S1034: Using at least a portion of the historical feature data from multi-source Earth observations of each sampling unit as feature data and the corresponding measured diameter at breast height (DBH) as label data, a DBH prediction model is trained.

[0134] The diameter at breast height (DBH) prediction model refers to a regression model constructed using machine learning algorithms, with historical characteristic data from multiple sources of Earth observation as input variables and the measured DBH value of mangroves in the field as the output label. This model is used to capture the nonlinear mapping relationship between the spectral, texture, and structural information contained in remote sensing images and the diameter at the base of vegetation, thereby achieving non-destructive, large-scale DBH parameter inversion.

[0135] Specifically, historical feature data from multi-source Earth observations of the sampling units, such as the three-dimensional structural features, radar scattering features, and spectral features generated in step S10125, are selected as input feature data. Simultaneously, the measured diameter at breast height (DBH) of the corresponding sampling units is selected as the label data for supervised learning. Considering the complexity of mangrove canopy structure, the DBH prediction model preferably employs an ensemble learning algorithm, such as random forest or extreme gradient boosting (XGBoost) regression model. During training, the model learns the correlation between remote sensing features and DBH growth in different habitats (e.g., high and low shoals). The model's hyperparameters are optimized by minimizing the mean square error between predicted and measured values, ultimately resulting in a DBH prediction model with strong generalization ability.

[0136] Step S1035: Using at least a portion of the historical feature data of multi-source Earth observations from each sampling unit as feature data and the corresponding measured tree height as label data, a tree height prediction model is trained.

[0137] Tree height prediction models are regression models constructed using machine learning algorithms, with historical characteristic data from multi-source Earth observations as input variables and measured mangrove tree height as the output label. Given that tree height is an important representation of vegetation vertical structure, this model typically utilizes lidar point cloud features or radar interferometric phase information to establish a quantitative relationship with vegetation height.

[0138] A similar approach was used to construct a tree height prediction model. At least a portion of the historical feature data from multi-source Earth observations of the sampling units (especially the 3D structural features containing the canopy height model CHM) were used as input feature data, and the measured tree height of the corresponding sampling units was used as label data. Due to the strong physical correlation between tree height and lidar point cloud data, the tree height prediction model could employ support vector regression (SVR) or a deep learning model. Through training, the model mastered the ability to invert vegetation vertical height using remote sensing features, accurately identifying the canopy apex location and vertical distribution characteristics of mangroves.

[0139] Step S1036: Input at least some of the multi-source Earth observation historical feature data corresponding to each grid cell in the target area into the diameter at breast height (DBH) prediction model and the tree height prediction model respectively to obtain the predicted canopy DBH and predicted canopy tree height corresponding to each grid cell; determine the predicted canopy DBH and predicted canopy tree height of each grid cell as the corresponding regional canopy inversion DBH and regional canopy inversion tree height.

[0140] Predicted canopy diameter at breast height (DBH) refers to the theoretical DBH value of mangroves at a given location, calculated by inputting historical characteristic data from multi-source Earth observations of grid cells (without field sampling) into a trained DBH prediction model. This value serves as one of the input parameters to the dynamic vegetation growth estimation equation, replacing the need for comprehensive measured data that is difficult to obtain in traditional methods.

[0141] Predicted canopy tree height refers to the theoretical tree height of the mangrove forest at a given location, calculated by inputting historical feature data from multi-source Earth observations of grid cells (without field sampling) into a trained tree height prediction model. This value, along with the predicted canopy diameter at breast height (DBH), constitutes a key biophysical parameter for calculating the carbon density of the grid cells.

[0142] This embodiment completely eliminates the reliance on manually measured data across the entire area by adding a remote sensing inversion step for diameter at breast height (DBH) and tree height before calculating carbon density, achieving full-chain automation from "feature input—parameter inversion—carbon density calculation". The high-precision mapping relationship between remote sensing features and vegetation structure parameters established using a machine learning model ensures the spatial continuity and estimation accuracy of regional canopy inversion DBH and tree height data, thus avoiding distortion of carbon density calculation results caused by the propagation of input parameter errors. This not only significantly reduces the manpower and time costs of large-scale mangrove surveys but also allows the dynamic coefficient mapping model set to fully function at a high-resolution grid scale, ensuring that the final generated carbon density spatial distribution map has high geometric accuracy.

[0143] For step S104, the carbon sequestration increment in the region is extrapolated based on the aboveground biomass carbon density of each grid cell at the reference time, and the prediction results of the carbon sequestration potential of vegetation in the target area are generated.

[0144] The baseline time refers to the time section that serves as the starting reference for carbon density prediction and carbon sequestration potential assessment. It is usually the current survey year or the time point of mangrove resource inventory in a specific historical period. The spatial distribution of carbon density at this time serves as the baseline for subsequent incremental calculations.

[0145] In one embodiment, based on the aboveground biomass carbon density of each grid cell at a baseline time, and combined with Geographic Information System (GIS) technology, the spatial distribution of aboveground biomass carbon density in the target area is generated. This result is typically presented as a raster layer, with different colors or values ​​representing different carbon density levels. This distribution map not only visually demonstrates the spatial aggregation characteristics and edge effects of mangrove carbon storage, but can also be directly imported into carbon trading platforms or ecological management systems for calculating the total regional carbon storage and assessing the carbon sequestration effectiveness of ecological restoration projects.

[0146] In this embodiment of the application, after obtaining the full coverage and spatially continuous aboveground biomass carbon density of vegetation through the aforementioned steps, the carbon density dataset is the core basic spatial carrier for predicting the carbon sequestration potential of mangrove vegetation. Based on the gridded carbon density data, the complete generation of the carbon sequestration potential prediction results of the target area can be completed. Specifically, a variety of different implementation paths can be used to generate the carbon sequestration potential prediction results.

[0147] In one implementation, environmental time-series data on forest age, average annual temperature, average annual precipitation, and sea-level rise rate for each grid cell in the target area can be retrieved from the associated storage. The gridded carbon density is used as the baseline carbon storage and input into pre-constructed vegetation growth models of various types. These models can include one or more, such as the von Bertalanffy growth model, the Gompertz growth model, and the Chapman-Richards growth model, to perform a fitting calculation between carbon density and forest age. Simultaneously, one or more of the data, such as average annual temperature, average annual precipitation, and sea-level rise rate, are used as covariates. By embedding parameter systems from various growth models, the regulatory effects of different environmental conditions on the carbon accumulation rate of mangroves are depicted. Based on this, multiple simulation scenarios are defined, including those involving only changes in forest age, only changes in temperature, only changes in precipitation, only changes in sea-level rise, and synergistic changes involving multiple environmental factors (i.e., encompassing at least two of these changes). The carbon density increment compared to the baseline year is calculated for each grid cell within a predetermined future time period. The carbon increment values ​​under different scenarios for each grid cell are matched one-to-one with the spatial grid location, integrating them to form a spatial raster dataset of carbon sequestration potential covering the entire region. This raster dataset represents the predicted carbon sequestration potential of vegetation in the target area. Furthermore, the raster data can be rendered hierarchically and statistically analyzed by region according to business needs, outputting visualized thematic mapping results that fully convey the upper limit and incremental potential of carbon sink growth achievable by mangroves under different spatial locations and climate development scenarios within the region.

[0148] In another implementation, for short-term ecological restoration project assessment scenarios, there is no need to introduce multiple types of climate prediction covariates. Instead, the current gridded carbon density and the existing forest age structure of each grid unit are used as fixed benchmarks. Combined with the growth saturation carbon density threshold corresponding to the mangrove species in the region, the theoretical carbon sequestration increment of each grid unit under the condition of not reaching growth saturation is directly obtained through difference calculation. Simultaneously, the incremental values ​​of the gain coefficients of human interventions such as artificial tending and ecological restoration are superimposed to quickly generate the short-term carbon sequestration potential vector zoning results that only consider the dual conditions of natural forest age succession and artificial restoration intervention. This result can be directly matched with the boundary of the restoration plot to carry out zoning potential statistics.

[0149] In one embodiment, step S104, which extrapolates the regional carbon sequestration increment based on the aboveground biomass carbon density of each grid cell at a reference time and generates a predicted result of the mangrove vegetation carbon sequestration potential of the target area, includes: Step S1041: Obtain forest age data and environmental factor data for each grid cell at the reference time, as well as environmental factor data for a future preset time period; the environmental factor data includes annual average temperature data, annual average precipitation data, and sea level rise rate data.

[0150] Tree age data refers to the number of years mangrove vegetation has grown from seedling planting to a baseline time, and is a key temporal parameter characterizing the developmental stage and biomass accumulation level of the vegetation community. This data can be obtained through historical remote sensing image retrospection, logging record surveys, or plot inference based on dendrochronology.

[0151] Environmental factor data refers to the set of key climate and ocean dynamic indicators that drive mangrove vegetation growth and carbon accumulation. In this embodiment, it specifically includes annual average temperature data, annual average precipitation data, and sea level rise rate data, which are used to characterize the macro-ecological environment background of the vegetation.

[0152] The future preset time period refers to the forward-predicted time span set for assessing the carbon sequestration potential of mangroves, such as the next 10, 20, or 50 years. The length of this time period can be flexibly configured according to the blue carbon trading cycle or the needs of ecological management planning.

[0153] In this step, forest age data and environmental factor data for each grid cell at the baseline time, as well as environmental factor data for the corresponding future preset time period, are acquired. The forest age data at the baseline time can be obtained through historical remote sensing image retrospection or field surveys, while the environmental factor data can be extracted from the meteorological bureau's historical observation database or historical simulation grids of climate models. The environmental factor data for the corresponding future preset time period comes from the output results of climate models such as CMIP6 under different shared socio-economic pathways, ensuring that the predictions have a scientific basis.

[0154] Step S1042: Construct a carbon density-forest age prediction model; the carbon density-forest age prediction model uses forest age as the independent variable and aboveground biomass carbon density of vegetation as the dependent variable, and embeds at least one of the annual mean temperature data, annual mean precipitation data and sea level rise rate data as covariates into the model parameters of the carbon density-forest age prediction model.

[0155] The carbon density-forest age prediction model refers to a growth regression model that uses forest age as the independent variable and aboveground biomass carbon density of vegetation as the dependent variable. This model incorporates environmental factors as covariates into its internal parameter structure, enabling the carbon density growth trajectory with forest age to dynamically respond to the impacts of climate change and sea-level rise.

[0156] Covariate embedding refers to introducing external environmental variables (such as climate or hydrological factors) as independent variables into the parameter structure of a growth model, so that the growth limit, growth rate or inflection point of the model can be dynamically adjusted with environmental changes, rather than remaining a fixed constant.

[0157] In this step, a carbon density-forest age prediction model is constructed. This model uses forest age as the independent variable and aboveground biomass carbon density as the dependent variable. Models such as the von Bertalanffy growth model, the Gompertz growth model, and the Chapman-Richards growth model can be used. The key is to embed at least one of the following as covariates into the model parameters: annual mean temperature data, annual mean precipitation data, and sea-level rise rate data. For example, the growth rate parameter can be set as a linear function of the annual mean temperature data, or the carrying capacity parameter can be set as an exponentially decaying function of the sea-level rise rate data. This allows the model to express the regulatory mechanism of environmental change on the growth limit and rate of mangroves.

[0158] Step S1043: For each grid cell, the aboveground biomass carbon density, forest age data, and environmental factor data at the reference time, as well as the environmental factor data corresponding to the future preset time period, are input into the carbon density-forest age prediction model to obtain the aboveground biomass carbon density of each grid cell corresponding to the future preset time period.

[0159] In this step, for each grid cell within the target area, the aboveground biomass carbon density at the baseline time, the cell's age data, and environmental factor data, along with environmental factor data for a future preset time period, are input into the carbon density-age prediction model. The model first calibrates the specific growth parameters of the cell based on the environmental factors at the baseline time and the age of the cell. Then, it substitutes future environmental factors and extrapolates to obtain the predicted aboveground biomass carbon density value for that cell at the end of the future preset time period, achieving a grid-by-grid extrapolation from a static current state map to a dynamic future map.

[0160] Step S1044: Calculate the difference between the aboveground biomass carbon density of vegetation in each grid cell at the future preset time period and the aboveground biomass carbon density of vegetation at the reference time to obtain the aboveground biomass carbon density increment of vegetation in each grid cell; generate the predicted result of the carbon sequestration potential of mangrove vegetation in the target area based on the aboveground biomass carbon density increment of vegetation in each grid cell.

[0161] The increase in vegetation aboveground biomass carbon density refers to the difference between the predicted value of vegetation aboveground biomass carbon density at a future preset time period and the vegetation aboveground biomass carbon density at the baseline time within the same grid cell. It directly reflects the net carbon sequestration capacity of the cell per unit time.

[0162] The predicted carbon sequestration potential of mangrove vegetation refers to the spatial aggregation and visualization of the increase in aboveground biomass carbon density of all grid units in the target area over a future preset period. It is usually presented in the form of raster layers or statistical reports and is used to quantify the future carbon sink contribution of the mangrove ecosystem.

[0163] This step calculates the difference between the aboveground biomass carbon density of vegetation in each grid cell at a future preset time period and the aboveground biomass carbon density of vegetation at the baseline time, thus obtaining the increase in aboveground biomass carbon density of vegetation in each grid cell. Then, the increments of all grid cells are spatially superimposed and statistically analyzed to generate the prediction result of the carbon sequestration potential of mangrove vegetation in the target area. This result can intuitively identify high-potential restoration areas and vulnerable loss areas.

[0164] This embodiment constructs a carbon density-forest age prediction model embedded with environmental covariates, extending the static carbon density distribution map at a baseline time to the dynamic carbon sequestration potential prediction for a future preset period. This effectively overcomes the shortcomings of traditional carbon sink measurement, which only focuses on the current situation and ignores the evolutionary patterns driven by climate. By utilizing the coupling relationship between forest age and environmental factors to estimate carbon density increments, the prediction results can scientifically reflect the marginal impact of rising temperatures, changes in precipitation patterns, and sea-level rise on the carbon sequestration function of mangroves, significantly improving the reliability of long-term blue carbon asset accounting. The carbon sequestration potential prediction results generated by this scheme use grid cells as the precision unit, combining macro-level totals with spatial heterogeneity characteristics, providing a high-value data foundation for mangrove ecological restoration engineering planning, climate investment and financing project design, and protection strategies for adapting to climate change.

[0165] In one embodiment, the carbon density-forest age prediction model in step S1042 includes at least one of the von Bertalanffy growth model, the Gompertz growth model, and the Chapman-Richards growth model; the carbon density-forest age prediction model is obtained by fitting parameters using nonlinear least squares method based on the aboveground biomass carbon density data of vegetation corresponding to different forest age samples; at least one of the annual mean temperature data, annual mean precipitation data, and sea level rise rate data is set as a covariate of at least one of the growth limit parameter, growth rate parameter, and shape parameter of the carbon density-forest age prediction model.

[0166] The growth limit parameter refers to the model coefficient in the growth model that represents the maximum saturation value that the aboveground biomass carbon density of vegetation can reach in the theoretical succession stage, reflecting the upper limit of the mangrove carbon pool under the carrying capacity of a specific habitat.

[0167] The growth rate parameter refers to the model coefficient in the growth model that controls the rate at which carbon density accumulates with the advancement of forest age, and determines the steepness of the transition of the model curve from a rapid rise in the juvenile stage to a gradual transition in the mature stage.

[0168] Shape parameters refer to the model coefficients in the growth model that adjust the overall asymmetry of the curve and the inflection point, so that the model can flexibly adapt to the skewed or symmetrical growth trajectories unique to different mangrove species.

[0169] The von Bertalanffy growth model is a classic biological growth equation that assumes that the growth rate of an organism is proportional to a power function of its current body length. It is often used to describe the allometric growth process of organisms such as fish and trees, and can simulate the S-shaped curve of mangroves transitioning from rapid growth in the juvenile stage to growth saturation in the mature stage.

[0170] The Gomperts growth model is an asymmetric growth model often used to describe the growth process of biological populations or individuals. Its characteristics are that the growth rate is slow in the early stage, accelerates in the middle stage, and gradually approaches the limit value in the later stage. It is particularly suitable for simulating the biomass accumulation of ecosystems limited by environmental carrying capacity.

[0171] The Chapman-Richardt growth model is a generalized growth function model that flexibly fits various types of growth curves by introducing shape parameters. It is widely used in forestry science to describe the changes in tree diameter, height, or volume with age.

[0172] The carbon density-forest age prediction model constructed in this embodiment specifically includes at least one of the von Bertalanffy growth model, the Gompertz growth model, and the Chapman-Richards growth model. Based on the measured data of aboveground biomass carbon density corresponding to samples of different ages in long-term mangrove monitoring, the initial parameters of the selected model are fitted using nonlinear least squares or Bayesian estimation methods to make the model curve approximate the actual growth trajectory of the samples. Crucially, at least one environmental factor among the annual mean temperature data, annual mean precipitation data, and sea-level rise rate data is set as a covariate for at least one of the growth limit parameter, growth rate parameter, and shape parameter of the carbon density-forest age prediction model. For example, the growth limit parameter of the Chapman-Richards model can be set as a direct proportional function of the annual mean precipitation data to express that abundant precipitation promotes an increase in the upper limit of biomass; simultaneously, the growth rate parameter can be set as a negative exponential function of the sea-level rise rate data to characterize the increased inhibition of growth rate by inundation of low-lying areas; alternatively, a linear regression relationship can be established between the shape parameter and the annual mean temperature data. This covariate embedding mechanism breaks the limitation of constant parameters in traditional growth models, enabling the model to dynamically express environmental evolution.

[0173] In one embodiment, step S1043, under at least two of the following scenarios—only forest age change scenario, only annual mean temperature change scenario, only annual mean precipitation change scenario, only sea level rise rate change scenario, and multi-factor synergistic change scenario—uses the carbon density-forest age prediction model to predict the aboveground biomass carbon density of each grid cell in the future preset time period, and calculates the increase in aboveground biomass carbon density of each grid cell; wherein, the multi-factor synergistic change scenario refers to a scenario in which at least two of the following changes occur simultaneously: forest age change, annual mean temperature change, annual mean precipitation change, and sea level rise rate change.

[0174] Multifactor synergistic change scenario refers to a composite assumption that considers the changes of multiple environmental factors, such as natural forest age growth, annual mean temperature, annual mean precipitation, and sea level rise rate, when simulating future evolution. It is used to assess carbon sequestration response under integrated climate and ocean dynamic pressures.

[0175] This embodiment, based on the covariate embedded prediction model constructed in step S1042, performs simulations under at least two of the following scenarios: only forest age change scenario, only annual mean temperature change scenario, only annual mean precipitation change scenario, only sea-level rise rate change scenario, and multi-factor synergistic change scenario. Specifically, all inputs except the target variable are locked as baseline values. Only the variables for the corresponding scenario are released to move towards a predetermined future time period. The baseline carbon density, forest age, and environmental factors of each grid cell, along with the environmental factors for the future time period, are input into the model. The model outputs the predicted aboveground biomass carbon density of each grid cell under that scenario for the predetermined future time period. Subsequently, according to the difference calculation logic in step S1044, the increase in aboveground biomass carbon density of each grid cell under that scenario is obtained. Through parallel simulations of at least two scenarios, the system can separate the contribution and risk levels of natural forest age maturity, single climate factor stress, and multi-factor coupling effects on mangrove carbon sequestration, ultimately converging to generate prediction results of the mangrove vegetation carbon sequestration potential of the target area under multi-dimensional assumptions.

[0176] This embodiment ensures the biological fidelity of the carbon density-sustainability prediction model to the allometric growth patterns of mangroves by explicitly employing three classic growth models and fitting them to multi-age samples. By embedding annual mean temperature, precipitation, and sea-level rise rate data as covariates for growth limits, rates, or shape parameters, the model is upgraded from a static empirical formula to a climate-responsive dynamic equation, accurately characterizing the marginal regulation of carbon sink function by environmental change. Furthermore, by outputting the future carbon density and increment of grid cells in parallel under at least two scenarios, this scheme not only quantifies the basic carbon sequestration from natural growth but also separates the carbon sink gains and losses under single and combined environmental pressures. This significantly enhances the scientific rigor and interpretability of long-term blue carbon asset accounting and climate risk assessment, providing high-precision scenario-based decision support for mangrove ecological management and carbon trading contract design.

[0177] This application also provides a system for predicting the carbon sequestration potential of mangrove vegetation, including: The mangrove area data acquisition module is used to acquire multi-source Earth observation historical characteristic data of the mangrove target area, as well as field vegetation sampling data of several sampling units within the target area; the field vegetation sampling data includes the measured diameter at breast height, measured tree height, and measured data of aboveground biomass carbon density of each sampling unit. The dynamic coefficient mapping model training module is used to construct training feature vectors for each sampling unit based on the historical feature data of multi-source Earth observation corresponding to each sampling unit; using the training feature vectors as input, the corresponding dynamic coefficients are output through the dynamic coefficient mapping model group to be trained; based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height and measured tree height, the predicted data of aboveground biomass carbon density of vegetation is calculated through the dynamic vegetation growth estimation equation; based on the predicted data of aboveground biomass carbon density of vegetation and the measured data of aboveground biomass carbon density of vegetation, a loss function is constructed, and the model parameters of the dynamic coefficient mapping model group are iteratively optimized according to the loss function to obtain the trained dynamic coefficient mapping model group; The vegetation carbon sequestration data prediction module is used to input the multi-source Earth observation historical feature data of each grid cell in the target area at the reference time into the trained dynamic coefficient mapping model group to obtain the dynamic coefficients corresponding to each grid cell; based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height and regional canopy inversion tree height, the aboveground biomass carbon density of each grid cell is calculated. The prediction result generation module is used to extrapolate the regional carbon sequestration increment based on the aboveground biomass carbon density of each grid cell at the reference time, and generate the prediction result of the carbon sequestration potential of mangrove vegetation in the target area.

[0178] It should be noted that the mangrove vegetation carbon sequestration potential prediction system provided in the above embodiments is only illustrated by the division of the above functional modules when executing the mangrove vegetation carbon sequestration potential prediction method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the mangrove vegetation carbon sequestration potential prediction system and the mangrove vegetation carbon sequestration potential prediction method provided in the above embodiments belong to the same concept, and its implementation process is detailed in the above embodiments of the mangrove vegetation carbon sequestration potential prediction method, which will not be repeated here.

[0179] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and this application also intends to include these modifications and variations.

Claims

1. A method for predicting the carbon sequestration potential of mangrove vegetation, characterized in that, Includes the following steps: The study acquires historical characteristic data of multi-source Earth observation of the target mangrove area, as well as field vegetation sampling data of several sampling units within the target area. The field vegetation sampling data includes measured diameter at breast height (DBH), measured tree height, and measured aboveground biomass carbon density of each sampling unit. Based on the historical feature data of multi-source Earth observation corresponding to each sampling unit, a training feature vector corresponding to each sampling unit is constructed; using the training feature vector as input, the corresponding dynamic coefficients are output through the dynamic coefficient mapping model group to be trained; based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height and measured tree height, the predicted data of vegetation aboveground biomass carbon density is calculated through the dynamic vegetation growth estimation equation; based on the predicted data of vegetation aboveground biomass carbon density and the measured data of vegetation aboveground biomass carbon density, a loss function is constructed, and the model parameters of the dynamic coefficient mapping model group are iteratively optimized according to the loss function to obtain the trained dynamic coefficient mapping model group; The multi-source Earth observation historical feature data of each grid cell in the target area at the reference time are input into the trained dynamic coefficient mapping model group to obtain the dynamic coefficients corresponding to each grid cell; based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height and regional canopy inversion tree height, the vegetation aboveground biomass carbon density of each grid cell at the reference time is calculated. Based on the carbon density of aboveground biomass of vegetation in each grid cell at the baseline time, the carbon sequestration increment in the region is extrapolated, and the predicted carbon sequestration potential of mangrove vegetation in the target area is generated.

2. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 1, characterized in that, Before calculating the vegetation aboveground biomass carbon density of each grid cell at the reference time, based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height (DBH) and regional canopy inversion tree height, the following steps are included: A diameter prediction model is trained by using at least some of the historical feature data from multi-source Earth observations of each sampling unit as independent variable feature data and the corresponding measured diameter at breast height (DBH) as label data. A tree height prediction model is trained by using at least some of the historical feature data of multi-source Earth observations from each sampling unit as independent variable feature data and the corresponding measured tree height as label data. At least some of the historical feature data from multi-source Earth observations corresponding to each grid cell in the target area are input into the diameter at breast height (DBH) prediction model and the tree height prediction model, respectively, to obtain the predicted canopy DBH and predicted canopy tree height for each grid cell; the predicted canopy DBH and predicted canopy tree height for each grid cell are determined as the corresponding regional canopy inversion DBH and regional canopy inversion tree height.

3. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 1, characterized in that, The multi-source Earth observation historical feature data includes three-dimensional structural features and remote sensing environmental features; The remote sensing environmental features include radar scattering features and spectral features; The steps for obtaining historical characteristic data of multi-source Earth observation of mangrove target areas include: Acquire UAV lidar point cloud data, radar satellite time-series image data, and optical satellite time-series image data of the mangrove target area; Feature extraction is performed on the UAV lidar point cloud data, the radar satellite time-series image data, and the optical satellite time-series image data to obtain three-dimensional structural features, radar scattering features, and spectral features. Based on the three-dimensional structural features, the radar scattering features, and the spectral features, multi-source Earth observation historical feature data of the mangrove target area are obtained.

4. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 3, characterized in that, The steps of extracting features from the UAV lidar point cloud data, the radar satellite time-series image data, and the optical satellite time-series image data to obtain three-dimensional structural features, radar scattering features, and spectral features include: Point cloud filtering and canopy height model calculation are performed on the UAV lidar point cloud data to extract canopy complexity and tree height-to-canopy ratio; based on the canopy complexity and tree height-to-canopy ratio, three-dimensional structural data are obtained. Spot filtering and terrain correction are performed on radar satellite time-series image data to extract VV-polarized backscattering values ​​and VH-polarized backscattering values. Derived radar indices are calculated based on the two types of backscattering values. Radar scattering data are obtained based on VV-polarized backscattering values, VH-polarized backscattering values, and derived radar indices. Cloud masking was performed on the time-series image data of optical satellites, clear sky observation data within a specified key phenological period were selected, multi-band reflectance values ​​were extracted by band, and normalized vegetation index was calculated based on the multi-band reflectance values; clear sky spectral data were obtained based on the multi-band reflectance and normalized vegetation index. Spatial alignment is performed on the three-dimensional structural data, the radar scattering data, and the clear-sky spectral data; using the specified key phenological period as a time filtering window, the time dimension of the spatially aligned three-dimensional structural data, radar scattering data, and clear-sky spectral data is matched to construct a spatiotemporally aligned data cube; By fusing 3D structural data, radar scattering data, and clear-sky spectral data within a spatiotemporally aligned data cube, 3D structural features, radar scattering features, and spectral features corresponding to each space at each time filtering window are generated.

5. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 4, characterized in that, The steps for fusing 3D structural data, radar scattering data, and clear-sky spectral data within a spatiotemporally aligned data cube to generate 3D structural features, radar scattering features, and spectral features corresponding to each space at each time filtering window include: Extract the full-time radar scattering data and full-time clear-sky spectral data of each grid cell in the spatiotemporal aligned data cube; Based on the full-time radar scattering data and full-time clear-sky spectral data corresponding to each grid cell, several types of time-series statistical features and several types of time-period features are extracted to generate a time-series statistical feature set and a harmonic phenological parameter set; according to the time-series statistical feature set and the harmonic phenological parameter set, a spatiotemporal phenological spectrum feature is constructed. The spatiotemporal phenological spectrum features are fused with the three-dimensional structural data of the corresponding grid cell to obtain the three-dimensional structural features of the corresponding grid cell; the spatiotemporal phenological spectrum features are fused with the radar scattering data and clear sky spectral data of the corresponding grid cell to obtain the radar scattering features and spectral features of the corresponding grid cell.

6. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 3, characterized in that, The dynamic coefficient mapping model group is used to learn the dynamic coefficients of the dynamic vegetation growth estimation equation, and the dynamic coefficients include dynamic proportional coefficients and dynamic growth index. The dynamic coefficient mapping model group is used to receive three-dimensional structural features and remote sensing environmental features as joint inputs; Based on the joint learning of the first mapping relationship corresponding to the three-dimensional structural features and the second mapping relationship corresponding to the remote sensing environment features, the adaptive dynamic scaling coefficient and dynamic growth index are output; wherein, the remote sensing environment features include radar scattering features and spectral features; The steps for calculating aboveground biomass carbon density of vegetation using the dynamic vegetation growth estimation equation, based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height (DBH) and measured tree height, include: Based on the dynamic proportional coefficient, dynamic growth index, corresponding measured diameter at breast height (DBH), and corresponding measured tree height of each sampling unit, the aboveground biomass carbon density of the vegetation is calculated.

7. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 1 or 6, characterized in that, The steps for constructing a loss function based on predicted and measured data of aboveground biomass carbon density include: Based on the deviation between the predicted data of vegetation aboveground biomass carbon density and the measured data of vegetation aboveground biomass carbon density, a sample fitting loss term is constructed. Generate a model complexity regularization term; fuse the sample fitting loss term and the model complexity regularization term to obtain a loss function for training the dynamic coefficient mapping model group; wherein, the model complexity regularization term includes a penalty term for the number of leaf nodes in the model and a leaf node weight magnitude constraint term.

8. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 1, characterized in that, The steps for generating predictions of vegetation carbon sequestration potential in a target area based on the vegetation aboveground biomass carbon density of each grid cell at a baseline time include: Acquire forest age data and environmental driving factor data for each grid cell at the reference time, as well as environmental factor data for a future preset time period; the environmental factor data includes annual average temperature data, annual average precipitation data, and sea level rise rate data. A carbon density-forest age theoretical growth prediction model is constructed. The carbon density-forest age theoretical growth prediction model uses forest age as the independent variable and aboveground biomass carbon density of vegetation as the dependent variable. At least one of the following data is embedded as a covariate into the model parameters of the carbon density-forest age prediction model: annual mean temperature data, annual mean precipitation data, and sea level rise rate data. For each grid cell, the aboveground biomass carbon density, stand age data, and environmental driving factor data at the reference time, as well as the environmental factor data corresponding to the future preset time period, are input into the carbon density-stand age theoretical growth prediction model to obtain the aboveground biomass carbon density of each grid cell in the future preset time period. The difference between the aboveground biomass carbon density of vegetation in each grid cell at the preset future time period and the aboveground biomass carbon density of vegetation at the reference time is calculated to obtain the aboveground biomass carbon density increment of vegetation in each grid cell; based on the aboveground biomass carbon density increment of vegetation in each grid cell, the predicted carbon sequestration potential of mangrove vegetation in the target area is generated.

9. The method for predicting the carbon sequestration potential of mangrove vegetation according to claim 8, characterized in that, The carbon density-forest age prediction model includes at least one of the von Bertalanffy growth model, the Gompertz growth model, and the Chapman-Richards growth model; the carbon density-forest age theoretical growth prediction model is obtained by parameter fitting based on the aboveground biomass carbon density data of vegetation corresponding to different forest age samples using the nonlinear least squares method; at least one of the annual mean temperature data, the annual mean precipitation data, and the sea level rise rate data is set as a covariate of at least one of the growth limit parameter, growth rate parameter, and shape parameter of the carbon density-forest age theoretical growth prediction model. Under at least two of the following scenarios—namely, only forest age change, only annual mean temperature change, only annual mean precipitation change, only sea-level rise rate change, and multi-factor synergistic change—the carbon density-forest age theoretical growth prediction model is used to predict the aboveground biomass carbon density of each grid cell in the future preset time period, and to calculate the increase in aboveground biomass carbon density of each grid cell. The multi-factor synergistic change scenario refers to a scenario in which at least two of the following changes occur simultaneously: forest age change, annual mean temperature change, annual mean precipitation change, and sea-level rise rate change.

10. A system for predicting the carbon sequestration potential of mangrove vegetation, characterized in that, include: The mangrove area data acquisition module is used to acquire multi-source Earth observation historical characteristic data of the mangrove target area, as well as field vegetation sampling data of several sampling units within the target area; the field vegetation sampling data includes the measured diameter at breast height, measured tree height, and measured data of aboveground biomass carbon density of each sampling unit. The dynamic coefficient mapping model training module is used to construct training feature vectors for each sampling unit based on the historical feature data of multi-source Earth observation corresponding to each sampling unit; using the training feature vectors as input, the corresponding dynamic coefficients are output through the dynamic coefficient mapping model group to be trained; based on the dynamic coefficients of each sampling unit and the corresponding measured diameter at breast height and measured tree height, the predicted data of aboveground biomass carbon density of vegetation is calculated through the dynamic vegetation growth estimation equation; based on the predicted data of aboveground biomass carbon density of vegetation and the measured data of aboveground biomass carbon density of vegetation, a loss function is constructed, and the model parameters of the dynamic coefficient mapping model group are iteratively optimized according to the loss function to obtain the trained dynamic coefficient mapping model group; The vegetation carbon sequestration data prediction module is used to input the multi-source Earth observation historical feature data of each grid cell in the target area at the reference time into the trained dynamic coefficient mapping model group to obtain the dynamic coefficients corresponding to each grid cell; based on the dynamic coefficients of each grid cell and the corresponding regional canopy inversion diameter at breast height and regional canopy inversion tree height, the aboveground biomass carbon density of vegetation in each grid cell is calculated. The prediction result generation module is used to extrapolate the regional carbon sequestration increment based on the aboveground biomass carbon density of each grid cell at the reference time, and generate the prediction result of the carbon sequestration potential of mangrove vegetation in the target area.

Citation Information

Patent Citations

  • Mangrove forest vegetation carbon sink monitoring and metering method based on unmanned aerial vehicle and AI technology

    CN119540803A

  • Forest carbon sink accounting method based on nonlinear regression and stepwise regression combined model

    CN121723379A