Coastal salt marsh vegetation carbon sink estimation method and system based on multi-temporal phenological characteristics

Abstract

The embodiment of the invention relates to the technical field of artificial intelligence, and provides a coastal salt marsh vegetation carbon sink estimation method and system based on multi-temporal phenological characteristics. The method comprises the following steps: acquiring a remote sensing image of a target coastal salt marsh area in a key phenological period; extracting an initial vegetation contour of a vegetation coverage range in the target area from the remote sensing image by adopting an unsupervised classification method; constructing a phenological decision tree model by taking an NDVI threshold method as a core, and integrating a machine learning enhanced node division mechanism and a phenological period slope analysis method to enhance the phenological decision tree model; inputting the initial vegetation contour into a phenology decision tree model, and performing multi-level classification on vegetation types according to the initial vegetation contour through the phenology decision tree model to distinguish vegetation coverage ranges of various types of vegetation in the target coastal salt marsh area; and vegetation carbon density parameters are obtained, and the estimation of the total carbon sink amount of the target coastal salt marsh area is completed through spatial superposition calculation in combination with a vegetation coverage range, so that accurate classification of vegetation and efficient estimation of carbon sink are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and in particular to a coastal salt marsh vegetation carbon sink estimation method and system based on multi-temporal phenology characteristics. BACKGROUND

[0002] As an important blue carbon ecosystem, coastal salt marshes have extremely high carbon capture and storage capacity. Their carbon sink function plays a key role in addressing global climate change and maintaining ecological balance. Quantifying the carbon sink of coastal salt marsh vegetation is an important basis for ecological protection and carbon cycle research. Therefore, there is an increasing demand for precise identification of coastal salt marsh vegetation types and carbon sink estimation techniques.

[0003] Currently, the premise of coastal salt marsh vegetation carbon sink estimation is to accurately obtain the spatial distribution information of vegetation types. Existing technologies mainly rely on field investigation and remote sensing monitoring methods. Field investigation method obtains vegetation species, biomass and other data through field sampling, and then calculates carbon sink by combining carbon density parameters. However, it has limitations. Remote sensing technology has become the mainstream method for monitoring coastal salt marsh vegetation due to its wide coverage, short cycle, and convenient data acquisition. Traditional remote sensing methods are based on single phenological period images or single vegetation index for vegetation classification, and then for carbon sink estimation. However, in practical applications, there are many technical problems: first, the parameter setting is highly subjective and the classification stability is insufficient. Existing methods often rely on artificial experience to set classification thresholds or phenological period parameters, lack of independent filtering mechanism for key features, resulting in poor adaptability of the model to different regions and years, and insufficient stability of the classification results. Second, the utilization of phenological differences is not sufficient, and the precision of vegetation differentiation is low. The phenological characteristics of coastal salt marsh vegetation (such as Suaeda salsa, Phragmites australis, Spartina alterniflora, and Aeluropus japonicus) have subtle differences. However, traditional methods rely on static phenological indicators and fail to effectively quantify the dynamic changes of vegetation during the growth cycle, especially in key phenological stages such as the withering period. It is difficult to accurately distinguish similar vegetation types (such as Spartina alterniflora and Phragmites australis), resulting in large errors in identifying the coverage of vegetation. Third, the error of carbon sink estimation is obvious. The lack of precision in vegetation classification directly affects the matching accuracy of carbon density parameters, and then leads to the deviation of the total carbon sink estimation result, making it difficult to meet the actual demand of high-precision carbon sink monitoring.

[0004] In summary, related technologies in the estimation of coastal salt marsh vegetation carbon sink face the problems of high cost of field investigation, difficulty in large-scale dynamic monitoring, subjectivity of remote sensing classification, insufficient utilization of phenological differences, and low estimation accuracy. Therefore, how to achieve accurate vegetation classification and efficient carbon sink estimation has become an important research direction in the field of coastal salt marsh ecological monitoring. SUMMARY

[0005] The main objective of this application is to provide a method and system for estimating carbon sinks in coastal salt marsh vegetation based on multi-temporal phenological characteristics, aiming to solve at least one technical problem in related technologies, such as high cost of field surveys, difficulty in large-scale dynamic monitoring, strong subjectivity of remote sensing classification, insufficient utilization of phenological differences, and low estimation accuracy.

[0006] In a first aspect, embodiments of this application provide a method for estimating carbon sinks in coastal salt marsh vegetation based on multi-temporal phenological characteristics, including: Acquire remote sensing images of the target coastal salt marsh area during key phenological periods; The remote sensing image is processed using an unsupervised classification method to extract the initial vegetation outline of the vegetation cover area within the target region; A phenological decision tree model was constructed based on the NDVI threshold method. A machine learning-enhanced node partitioning mechanism was incorporated into the phenological decision tree model to optimize the node hierarchy and judgment priority of the decision tree. The phenological slope analysis method was used to enhance the ability of the phenological decision tree model to identify the phenological differences of various types of vegetation during the withering period. The initial vegetation outline is input into the phenological decision tree model. The phenological decision tree model performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation coverage of Suaeda salsa, Reed, Spartina alterniflora, and Ligusticum striatum in the target coastal salt marsh area. Obtain the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. Combine the vegetation cover ranges corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, and complete the estimation of the total carbon content of the target coastal salt marsh area through spatial overlay calculation.

[0007] Secondly, embodiments of this application provide a system for estimating carbon sequestration in coastal salt marsh vegetation based on multi-temporal phenological characteristics, including: The acquisition module is used to acquire remote sensing images of the target coastal salt marsh area during key phenological periods; The classification module is used to process the remote sensing image using an unsupervised classification method to extract the initial vegetation outline of the vegetation cover area within the target region. The module is used to construct a phenological decision tree model based on the NDVI threshold method. The model incorporates a machine learning-enhanced node partitioning mechanism to optimize the node hierarchy and judgment priority of the decision tree. The phenological slope analysis method is used to strengthen the ability of the phenological decision tree model to identify the phenological differences of various types of vegetation during the withering period. The decision module is used to input the initial vegetation outline into the phenological decision tree model, and the phenological decision tree model performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation coverage range of Suaeda salsa, Reed, Spartina alterniflora and Ligusticum striatum in the target coastal salt marsh area. The estimation module is used to obtain the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. Combined with the vegetation cover range corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, the total carbon content of the target coastal salt marsh area is estimated by spatial overlay calculation.

[0008] Thirdly, embodiments of this application also provide an electronic device, which includes a processor and a memory for storing a computer program; the processor is used to execute the computer program and, when executing the computer program, implement the method for estimating carbon sinks of coastal salt marsh vegetation based on multi-temporal phenological characteristics as described in the first aspect or any embodiment of this application.

[0009] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer software program, which, when executed by a processor, implements the method for estimating carbon sinks of coastal salt marsh vegetation based on multi-temporal phenological characteristics as described in the first aspect or any embodiment of this application.

[0010] This application provides a method and system for estimating carbon sinks in coastal salt marshes based on multi-temporal phenological characteristics. First, remote sensing images of the target coastal salt marsh area during key phenological periods are acquired. Then, an unsupervised classification method is used to process the remote sensing images, extracting the initial vegetation outline of the vegetation cover within the target area. A phenological decision tree model is constructed using the NDVI thresholding method as its core. A machine learning-enhanced node partitioning mechanism is incorporated into the phenological decision tree model to optimize the node hierarchy and decision priority of the decision tree. Furthermore, a phenological slope analysis method is used to strengthen the phenological decision tree model's ability to identify the phenological differences of various vegetation types during the withering period. Then, the initial vegetation outline is input into the phenological decision tree model. The model performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation cover of Suaeda salsa, Phragmites communis, Spartina alterniflora, and Ligusticum striatum within the target coastal salt marsh area. Finally, the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum are obtained. Combined with the corresponding vegetation cover areas of Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, the total carbon content of the target coastal salt marsh area is estimated through spatial overlay calculation. This application's embodiment achieves greater accuracy and comprehensiveness in carbon sink assessment, completing the entire process optimization from fine vegetation classification to accurate carbon content estimation. It effectively solves the technical problems in related technologies, such as high cost of field surveys, difficulty in large-scale dynamic monitoring, strong subjectivity in remote sensing classification, insufficient utilization of phenological differences, and low estimation accuracy. Attached Figure Description

[0011] Figure 1 A flowchart illustrating a method for estimating carbon sequestration in coastal salt marsh vegetation based on multi-temporal phenological characteristics, provided for an embodiment of this application; Figure 2 A schematic diagram of the module structure of a coastal salt marsh vegetation carbon sink estimation system based on multi-temporal phenological characteristics provided in an embodiment of this application; Figure 3 This is a schematic block diagram of a terminal device provided in an embodiment of this application. Detailed Implementation

[0012] This application provides a method and system for estimating carbon sequestration in coastal salt marsh vegetation based on multi-temporal phenological characteristics. The method for estimating carbon sequestration in coastal salt marsh vegetation based on multi-temporal phenological characteristics can be applied to terminal devices, such as mobile terminals, mobile phones, virtual reality devices, tablets, laptops, desktop computers, wearable devices, and other electronic devices. The terminal device can be a server connected to a cloud service system or a server cluster. The connection can be implemented through hardware circuitry or through a communication module.

[0013] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating a method for estimating carbon sinks in coastal salt marsh vegetation based on multi-temporal phenological characteristics, provided in an embodiment of this application.

[0014] like Figure 1 As shown, the method for estimating carbon sinks in coastal salt marsh vegetation based on multi-temporal phenological characteristics includes the following steps: Step S101: Acquire remote sensing images of the target coastal salt marsh area during key phenological periods.

[0015] In this embodiment of the application, the target coastal salt marsh area refers to the coastal salt marsh area where vegetation carbon sequestration needs to be estimated. This area can be selected by researchers or technical implementers. This area has typical coastal salt marsh ecological characteristics, with characteristic vegetation such as Suaeda salsa, Reed, Spartina alterniflora, and Ligusticum striatum. It is also greatly affected by environmental factors such as tides and salinity. Its specific scope can be defined according to research or application needs, and it usually includes a complete vegetation growth environment gradient and ecosystem unit.

[0016] Critical phenological periods are those during vegetation growth when significant differences in spectral characteristics are observed. For example, critical phenological periods include at least one of the following: June, August, and November. Specifically, June marks the initial stage of vigorous growth for coastal salt marsh vegetation, with its growth pattern gradually becoming apparent and various vegetation types beginning to exhibit distinct growth characteristics. August is the peak growth period, with vegetation biomass reaching a high level, and its spectral characteristics are more prominent in remote sensing imagery, making it an important period for distinguishing different vegetation types. November marks the beginning of the withering period, with differences in the withering process among different vegetation types. These differences can be reflected in remote sensing imagery through spectral information, which is particularly crucial for distinguishing easily confused vegetation such as Spartina alterniflora and Phragmites communis. Selecting these three periods as critical phenological periods allows for a comprehensive capture of the phenological characteristics of vegetation at different stages of its growth cycle, providing rich and crucial information support for subsequent vegetation classification.

[0017] In acquiring remote sensing images of the target coastal salt marsh area during key phenological periods, data can be collected using remote sensing satellites or airborne remote sensing equipment. For example, in step S101, remote sensing image data of the target coastal salt marsh area during at least one key phenological period in June, August, and November is acquired using remote sensing satellites or airborne remote sensing equipment. The remote sensing image data includes spectral information used to extract the NDVI vegetation index.

[0018] Remote sensing satellites, with their wide coverage and stable data acquisition cycles, are capable of monitoring large-scale coastal salt marsh areas. Airborne remote sensing equipment, on the other hand, offers higher spatial resolution, enabling the acquisition of more detailed remote sensing image data in specific areas. Combining both methods can meet monitoring needs at different scales and with varying precision. The acquired remote sensing image data must contain spectral information from which vegetation indices such as NDVI can be extracted. NDVI is a crucial indicator reflecting vegetation growth, and its value is closely related to vegetation cover and growth vigor. Remote sensing images containing this spectral information form the basis for subsequent vegetation contour extraction. Unsupervised classification methods can quickly delineate the approximate extent of vegetation based on this spectral information. Simultaneously, this spectral information is also core data for constructing phenological decision tree models, providing the initial basis for the model's machine learning-enhanced node partitioning mechanism to select key phenological periods and index combinations, and for the phenological period slope analysis method to calculate the NDVI change slope. Furthermore, during vegetation classification, the spectral information in remote sensing images helps the model accurately distinguish different vegetation types, ensuring the accuracy of vegetation cover identification, and thus laying a reliable data foundation for carbon totality estimation.

[0019] Step S102: Process the remote sensing image using an unsupervised classification method to extract the initial vegetation outline of the vegetation coverage area within the target region.

[0020] In this embodiment, the initial vegetation outline is the spatial distribution boundary of vegetation cover initially extracted from remote sensing images using an unsupervised classification method. The initial vegetation outline is the result of preliminary differentiation between vegetated and non-vegetated areas based on the spectral characteristics of remote sensing images (such as the NDVI vegetation index).

[0021] As an optional embodiment, in step S102, the remote sensing image is preprocessed specifically to address the image noise characteristics caused by tidal interference and water reflection in the target coastal salt marsh area. An adaptive filtering algorithm is used to remove salt-and-pepper noise and stripe noise from the remote sensing image. Radiometric bias correction and geometric fine correction are then performed by combining regional topographic data and a tidal correction model to ensure the accuracy of the image's spectral information and spatial location. Furthermore, for the preprocessed remote sensing image, an improved K-means unsupervised classification algorithm is used. By dynamically adjusting the cluster centers and iterative thresholds, hierarchical automatic clustering analysis is performed on the land cover types in the preprocessed remote sensing image. Based on the differences in the spectral response of the NDVI vegetation index, vegetated and non-vegetated areas are distinguished. Finally, from the clustering results, a set of pixels corresponding to the vegetated areas is selected by setting a threshold range for vegetation spectral characteristics. Morphological filtering is then used to remove discrete noise pixels, and the spatial distribution range of the pixel set is defined as the initial vegetation outline of the vegetation coverage area within the target coastal salt marsh area.

[0022] Specifically, the initial vegetation contour extraction process in step S102 revolves around image denoising and accurate clustering under tidal interference. Addressing the issues of water reflection and mudflat exposure caused by tidal fluctuations in the target coastal salt marsh area, as well as salt-and-pepper noise and stripe noise from sensor noise, targeted preprocessing of the remote sensing image is first performed. An adaptive filtering algorithm is used to scan the image pixel by pixel. For salt-and-pepper noise with abrupt grayscale changes, a median filtering window dynamic adjustment strategy is employed, with the window size adaptively switching between 3×3 and 5×5 pixels based on the noise density. For regularly distributed stripe noise, Fourier transform is used to convert the image to the frequency domain, identifying the spectral peaks corresponding to the stripe noise and filtering them to suppress them, effectively preserving the spectral details of the vegetation. The preprocessing stage also combines regional 1:5000 topographic data with a tidal correction model. Based on the tidal level data at the time of image acquisition, each pixel is corrected for radiometric deviation to eliminate spectral reflectance deviation caused by differences in tidal inundation depth. At the same time, geometric fine correction is performed through ground control points to control the spatial position error of the image to within one pixel, ensuring the dual accuracy of spectral information and spatial position in subsequent analysis.

[0023] After preprocessing, an improved K-means unsupervised classification algorithm is applied to cluster land cover in the images. The algorithm first determines the optimal number of clusters (6 classes) based on the statistical distribution characteristics of the NDVI vegetation index in the images using the elbow method, covering typical land cover types such as high-coverage vegetation, medium-to-low-coverage vegetation, water bodies, bare beaches, tidal channels, and artificial features. A dynamic adjustment mechanism is introduced during the clustering process. Initial cluster centers are automatically generated based on the NDVI mean and the mean band reflectance. After each iteration, the intra-class variance and inter-class distance for each class are calculated. When the intra-class variance of a class exceeds a preset threshold (1.2 times the standard deviation of spectral reflectance), the cluster center is automatically split and pixels are redistributed. When the inter-class distance between two classes is less than the merging threshold, they are merged into one class. Hierarchical automatic clustering analysis is achieved by dynamically adjusting the cluster centers and iteration thresholds. The clustering process mainly relies on the spectral response differences of the NDVI vegetation index. Areas with high NDVI values ​​(≥0.3) correspond to areas with vigorous vegetation photosynthesis, while areas with low NDVI values ​​(<0.1) correspond to non-vegetated areas. This is used to distinguish the preliminary clustering results between vegetation-covered and non-vegetated areas.

[0024] Finally, vegetation cover areas were selected from the clustering results, and precise extraction was achieved by setting threshold ranges for vegetation spectral features. Based on the spectral characteristics of salt marsh vegetation, clusters with NDVI ≥ 0.2, red band reflectance ≤ 0.25, and near-infrared band reflectance ≥ 0.4 were labeled as vegetation cover classes, and the corresponding pixel sets were extracted. To eliminate discrete noise pixels generated during clustering, morphological filtering was used for post-processing. First, isolated noise points were removed by erosion of 3×3 structuring elements, and then the complete morphology of the vegetation area was restored by dilation. Finally, the spatial distribution range of the filtered pixel set was defined as the initial vegetation outline of the vegetation cover range within the target coastal salt marsh area. This outline can clearly distinguish the boundary between vegetated and non-vegetated areas, providing a reliable foundation for the subsequent fine classification of the phenological decision tree model.

[0025] Besides the improved K-means unsupervised classification algorithm in the above embodiments, step S102 can also be implemented using other unsupervised classification methods. In practical applications, the method can be selected based on scene features and actual image processing requirements.

[0026] For coastal salt marsh images with significant spectral fluctuations due to tidal interference, the ISODATA algorithm is used to extract initial vegetation contours. First, adaptive filtering and denoising are applied to the remote sensing image to remove salt-and-pepper noise and stripe noise. Then, radiometric bias correction and geometric fine correction are performed by combining regional topographic data and a tidal correction model, preserving spectral information used for NDVI vegetation index extraction. The preprocessed image data is input, with an initial cluster size of 6 to cover typical land features such as vegetation, water bodies, and sandbars. The maximum number of iterations is 50, the intra-cluster standard deviation threshold is 0.05, and the minimum intra-cluster pixel count is 30. The mean spectral features of pixels are calculated iteratively, and cluster merging and splitting operations are automatically performed. For example, when the number of pixels in a certain cluster is less than the minimum threshold, it is merged into a neighboring cluster; when the intra-cluster standard deviation exceeds the threshold, it is split into two clusters, dynamically optimizing the number of clusters. The mean NDVI of each cluster is calculated, and clusters with an NDVI greater than 0.15 are selected as vegetation candidate clusters. The pixel sets of adjacent vegetation candidate clusters are merged to obtain the preliminary vegetation coverage area. A 3×3 morphological filter is used to remove discrete noise pixels such as isolated patches with an area of ​​less than 5 pixels. After smoothing the boundaries, a continuous vegetation cover is output, which is the initial vegetation outline.

[0027] For scenarios with high real-time requirements, an improved K-means algorithm with dynamic threshold optimization is used to quickly generate initial vegetation outlines. The preprocessing workflow is simplified, retaining only radiometric and geometric corrections, and the NDVI index is calculated and used as the core feature. The elbow method is used to determine the optimal number of clusters as 5. Cluster centers are dynamically initialized based on the NDVI distribution (vegetation centers are set to NDVI 0.4, and non-vegetation centers are set according to the spectral distribution of water bodies and light beaches). The similarity between pixels and cluster centers is calculated using Euclidean distance, and the iteration threshold is dynamically adjusted (iteration stops when the intra-cluster variance change rate is less than 1%) to quickly separate vegetation and non-vegetation clusters. Pixels in clusters with NDVI greater than 0.15 are directly selected, merged, and then subjected to fast morphological filtering (1×3 window) to remove small noise. The vegetation cover boundary is generated as quickly as possible, outputting an initial vegetation outline that meets real-time requirements.

[0028] To address the complex spectral scene of mixed Suaeda salsa and pristine beaches in high-salinity areas, this study utilizes the spatial-spectral joint features of spectral clustering to enhance discriminative power. First, the NDVI vegetation index is calculated on the preprocessed image. Then, a 3D feature vector (NDVI, red band reflectance, and near-infrared band reflectance) is constructed by combining red and near-infrared reflectance to enhance the spectral differences between vegetation and non-vegetation areas. Using pixels as nodes, similarity between pixels is calculated based on a Gaussian kernel function, with spatial distance accounting for 30% and spectral distance accounting for 70%, constructing a similarity matrix that includes spatial adjacency relationships. The similarity matrix undergoes eigenvalue decomposition, and the feature vectors corresponding to the top 5 largest eigenvalues ​​are selected to form a low-dimensional embedding space. K-means clustering is used to cluster the embedding vectors, with the number of clusters set to 5. The mean NDVI and spatial continuity of each cluster are calculated. Clusters with NDVI greater than 0.2 and spatial connectivity greater than 80% are labeled as vegetation classes, and their pixel sets are extracted. The initial vegetation outline is output by filling small cavities inside the vegetation area through morphological closing operations and then smoothing the boundaries through erosion-dilation operations.

[0029] For vegetation phenological dynamics scenes in multi-temporal images, when extracting initial vegetation contours using superpixels and density peak clustering, the SLIC algorithm is used to generate superpixels from multi-temporal preprocessed images (e.g., June, August, November). The superpixel size is set to 20×20 pixels, and the compactness parameter is set to 10 to ensure consistency of spectral and spatial features within each superpixel. The mean NDVI and NDVI change slope (June-August, August-November) of each superpixel are calculated during key phenological periods, constructing a 5-dimensional temporal feature vector (NDVI for 3 periods + 2 slopes). The local density (cutoff distance set to 1.5 times the standard deviation of the feature space) and distance peak of the superpixel feature vector are calculated, and cluster centers are automatically identified, i.e., superpixels with high local density and far from other peaks are selected. Based on the temporal characteristics of the cluster centers, a set of superpixels with a mean NDVI of not less than 0.18 and an NDVI greater than that during the peak growth period (August) than during the withering period (November) are selected and marked as vegetation superpixels. The pixels corresponding to the vegetation superpixels are merged, and the jagged edges of the superpixels are removed by boundary smoothing, resulting in an initial vegetation outline that is consistent across time.

[0030] For high-dimensional data, such as hyperspectral imagery, when using the DEC algorithm to automatically learn optimal features for vegetation contour extraction, the hyperspectral imagery is first radiometrically normalized, and 15 sensitive bands related to vegetation growth (such as red, near-infrared, and short-wave infrared bands) are selected. After dimensionality reduction, the core spectral information is preserved. A neural network containing an encoder (e.g., composed of an input layer, a 32-dimensional hidden layer, and a 10-dimensional embedding layer) and a decoder is constructed. Using the hyperspectral pixel spectrum as input, the network is pre-trained by reconstructing the loss function to learn deep feature representations of the data. K-means clustering (initially with 4 clusters) is performed on the 10-dimensional embedding features output by the encoder to generate pseudo-labels. The clustering loss is calculated using the pseudo-labels, and the autoencoder parameters are fine-tuned to enhance feature discriminativeness. The average NDVI value of each cluster (based on red and near-infrared bands) is calculated, and clusters with NDVI not less than 0.2 are classified as vegetation classes, and their corresponding pixel sets are extracted. Morphological filtering is used to remove noisy pixels, and spatial connectivity analysis is used to fill gaps in vegetation areas, outputting a high-resolution initial vegetation contour.

[0031] Step S103: Construct a phenological decision tree model based on the NDVI threshold method. Integrate a machine learning-enhanced node partitioning mechanism into the phenological decision tree model to optimize the node hierarchy and judgment priority of the decision tree. Use the phenological slope analysis method to strengthen the ability of the phenological decision tree model to identify the phenological differences of various types of vegetation during the withering period.

[0032] In this embodiment, the phenological decision tree model is a decision tree model specifically optimized for the classification of coastal salt marsh vegetation. It uses vegetation phenological characteristics as the core discrimination criterion and integrates multi-source data and algorithm mechanisms to achieve accurate classification.

[0033] A decision-making framework was constructed based on the NDVI threshold method. Combining the growth characteristics of coastal salt marsh vegetation (Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum), dynamic NDVI threshold ranges were set for different key phenological stages. During the peak growth period (August), the NDVI thresholds for Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum were set to 0.6-0.8, 0.4-0.6, and 0.2-0.4, respectively. During the withering period (November), the thresholds were adjusted accordingly to 0.3-0.5, 0.2-0.4, and 0.1-0.3, forming a basic decision-making node system to achieve preliminary differentiation of vegetation types. A random forest algorithm was embedded as the node optimization mechanism, with vegetation index combination data consisting of NDVI, EVI2, and MSAVI from June, August, and November as input. Core classification features were autonomously mined through feature importance ranking; for example, EVI2 contributed 45% to the classification of Spartina alterniflora in August, and NDVI contributed 38% to the classification of Phragmites australis in June. The node hierarchy is dynamically adjusted based on weight ranking, prioritizing nodes with high contribution features at higher levels. For example, "August EVI2 > 0.5" is used as the first-level discrimination node for Spartina alterniflora, optimizing the decision logic priority. The slope of NDVI change during the withering period from August to November is introduced as a dynamic verification indicator, utilizing the differences in the decay rates of different vegetation to improve classification accuracy. The absolute value of the slope during the withering period of Spartina alterniflora is relatively large (-0.02 to -0.05 / month), and the threshold is set to ≤-0.03 / month. The slope of Reed is moderate (-0.01 to -0.03 / month), forming a differentiated threshold system. The slope threshold is embedded in the decision tree as a key verification node. When the initial classification result conflicts with the slope feature (e.g., classified as Reed but with a slope ≤-0.03 / month), a secondary discrimination mechanism is triggered for reclassification, achieving secondary verification of the feature optimization results.

[0034] In this way, the NDVI thresholding method ensures the intuitiveness and interpretability of the classification logic, while the random forest enhancement mechanism improves the scientific rigor of feature mining and node optimization, adapting to the spectral complexity of salt marsh vegetation. Phenological slope analysis strengthens the sensitivity to vegetation decay characteristics during the withering period, effectively addressing the confusion caused by similar vegetation types. Through multi-mechanism fusion, the model can accurately distinguish the spatial distribution of Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, providing high-quality vegetation classification data support for carbon sink estimation in coastal salt marsh vegetation, and offering a referable technical framework for fine vegetation classification in other ecological monitoring scenarios.

[0035] Step S103 focuses on constructing a phenological decision tree model based on the NDVI threshold method. By incorporating machine learning enhancement mechanisms and phenological slope analysis, it achieves accurate identification of differences in the withering period of salt marsh vegetation. Specifically, firstly, the model framework is constructed using the NDVI vegetation index extracted from preprocessed remote sensing images as the core parameter. Combining the growth characteristics of coastal salt marsh vegetation, dynamic initial NDVI threshold ranges are set for the phenological differences of Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum: during the peak growth period (August), the NDVI threshold for Spartina alterniflora is set at 0.6-0.8, for Phragmites australis at 0.4-0.6, and for Suaeda salsa at 0.2-0.4; during the withering period (November), the NDVI threshold for Spartina alterniflora is lowered to 0.3-0.5, for Phragmites australis to 0.2-0.4, and for Suaeda salsa to remain at 0.1-0.3, forming a basic decision node system adapted to the salt marsh environment.

[0036] Building upon this foundation, a random forest algorithm is integrated as a machine learning enhancement mechanism for node partitioning. Input is a combination of vegetation indices (NDVI, EVI2, and the modified soil-regulated vegetation index, MSAVI) after radiometric and geometric correction for each key phenological period (June, August, and November). The input features are weighted using a feature importance ranking algorithm to autonomously identify core classification features: EVI2 contributes 45% to the classification of Spartina alterniflora in August, NDVI contributes 38% to the classification of Phragmites australis in June, and NDVI contributes 32% to the classification of Suaeda salsa in November. The decision tree node hierarchy is dynamically adjusted based on the weight ranking, placing nodes corresponding to high-contribution features at the upper levels for priority judgment. For example, EVI2 > 0.5 in August is used as a primary discriminant node for Spartina alterniflora, and NDVI between 0.3 and 0.5 in June is used as a secondary discriminant node for Phragmites australis, thus optimizing node judgment priority.

[0037] Meanwhile, phenological slope analysis was employed to enhance the identification of vegetation differences during the withering period. Based on NDVI data from key phenological periods, the slope of NDVI changes during the growth phase (June-August) and the withering phase (August-November) was calculated using a linear regression algorithm, with the slope value from August to November serving as the core dynamic indicator. Comparison of vegetation slope distribution characteristics in the sample database revealed that: Spartina alterniflora had a relatively large absolute slope value during the withering period (-0.02 to -0.05 / month), Phragmites australis had a moderate absolute slope value (-0.01 to -0.03 / month), and Suaeda salsa had a relatively gentle slope value (-0.005 to -0.02 / month). Based on this, a differential adaptive threshold was set, with the slope threshold for Spartina alterniflora set to ≤-0.03 / month and for Phragmites australis set to -0.01 to -0.03 / month, and embedded in a decision tree as a key validation node. When a pixel is initially classified as reed by the NDVI threshold but the slope during the withering period is ≤-0.03 / month, a secondary discrimination mechanism is triggered, and it is reclassified as Spartina alterniflora, forming a secondary verification of the feature optimization results, thereby improving the recognition accuracy of vegetation types during the withering period.

[0038] Through the above steps, the constructed phenological decision tree model combines the intuitiveness of NDVI threshold, the feature optimization capability of machine learning, and the dynamic discrimination advantage of phenological slope. It can effectively distinguish the spectral differences of different vegetation types during key phenological periods, especially strengthening the sensitive identification of vegetation decay characteristics during the withering period, laying the model foundation for subsequent accurate vegetation classification.

[0039] As an optional embodiment, in step S103, firstly, using the NDVI vegetation index extracted from the preprocessed remote sensing image as the core parameter, and combining it with the growth characteristics of coastal salt marsh vegetation, a dynamic initial NDVI threshold range is set to construct a model framework for a phenological decision tree model adapted to the salt marsh environment. Secondly, a random forest algorithm is embedded in the model framework as a machine learning enhancement node partitioning mechanism. Vegetation index combination data for each key phenological period, after radiometric bias correction and geometric fine correction, is input. The vegetation index combination data includes NDVI and EVI2. A feature importance ranking algorithm is used to autonomously mine the weight relationship between different key phenological periods and index combinations on vegetation classification, identifying the core features that contribute the most to the classification of Suaeda salsa, Phragmites communis, Spartina alterniflora, and Ligusticum striatum. The node hierarchy and judgment priority of the decision tree are dynamically adjusted based on the weight ranking of the core features. Furthermore, based on NDVI data from key phenological periods, a linear regression algorithm was used to calculate the NDVI change slopes from June to August and August to November. The slope value from August to November was used as the core dynamic indicator of the phenological period slope analysis method. An adaptive threshold was set by comparing the difference in NDVI decay rates between Spartina alterniflora and Reed during the withering period. The adaptive threshold was embedded in the decision tree as a key verification node to form a secondary verification of the feature optimization results, thereby enhancing the sensitivity and recognition accuracy of the phenological decision tree model to vegetation phenological differences during the withering period, thus completing the construction of the phenological decision tree model.

[0040] Specifically, step S103, constructing the phenological decision tree model, begins with building a basic framework. Using the NDVI vegetation index extracted from preprocessed remote sensing images as the core parameter, and combining it with the growth characteristics of vegetation in the coastal salt marsh region, an initial NDVI threshold range that varies with phenological stages is set to build a basic framework adapted to the salt marsh environment, providing initial discrimination criteria for subsequent vegetation classification. Based on this, to optimize the model's node partitioning logic, a random forest algorithm is embedded in the constructed model framework as a machine learning-enhanced node partitioning mechanism. Input is vegetation index combinations for each key phenological stage, after radiometric bias correction and geometric fine correction. This data includes NDVI and EVI2. A feature importance ranking algorithm is used to autonomously analyze the weight relationship between different key phenological stages and index combinations in vegetation classification, identifying the core features that contribute most to the classification of plants such as Suaeda salsa, Phragmites communis, Spartina alterniflora, and Ligusticum striatum. Based on the weight ranking results of these core features, the hierarchical order and judgment priority of each node in the decision tree are dynamically adjusted, enabling the model to prioritize discrimination based on more discriminative features. To further enhance the model's ability to identify differences in vegetation phenology during the withering period, based on NDVI data from key phenological periods, a linear regression algorithm was used to calculate the slope of NDVI changes during the growth phase (June-August) and the withering phase (August-November). The slope value from August to November was used as the core dynamic indicator of the phenological slope analysis method. By comparing the differences in NDVI decay rates between Spartina alterniflora and Phragmites australis during the withering period, corresponding adaptive thresholds were set. These thresholds were embedded in the decision tree as key validation nodes, and the preliminary classification results obtained through core feature optimization were validated a second time. This enhanced the sensitivity and accuracy of the phenological decision tree model in identifying differences in vegetation phenology during the withering period, and finally completed the construction of the entire phenological decision tree model.

[0041] For example, in the above steps, based on NDVI data from key phenological periods, a linear regression algorithm is used to calculate the NDVI change slopes from June to August and August to November. The slope value from August to November is used as the core dynamic indicator of the phenological period slope analysis method. An adaptive threshold is set by comparing the difference in NDVI decay rates between Spartina alterniflora and Phragmites australis during the withering period. The adaptive threshold is embedded in the decision tree as a key verification node to form a secondary verification of the feature optimization results, including: The NDVI data after preprocessing the key phenological periods were smoothed over time to eliminate short-term fluctuations. A weighted linear regression algorithm was used to calculate the NDVI slope changes during the peak growth period (June-August) and the withering period (August-November). The withering period slope was assigned the highest weight to the November NDVI data to highlight the late-stage decay characteristics. Next, based on the withering period slope distribution characteristics of *Spartina alterniflora* and *Phragmites australis* in a historical vegetation classification sample database, a dynamic clustering algorithm was used to divide the slope intervals. Differentiated adaptive thresholds were set for different vegetation types; for example, the withering period slope threshold for *Spartina alterniflora* was set to -0.02 / month to -0.05 / month, and for *Phragmites australis*, it was set to -0.01 / month to -0.03 / month, forming a threshold system suitable for each vegetation type. Finally, the threshold is embedded in the decision tree model as an independent validation node to cross-validate the preliminary classification results after feature optimization. When the sample slope value exceeds the threshold range of the corresponding vegetation type, a secondary discrimination mechanism is triggered. Based on the validated withering slope value, vegetation carbon density parameters are dynamically allocated to construct a correlation model between slope and biomass decline. For herbaceous vegetation, such as Spartina alterniflora and Reed, the calculation method is the product between the carbon density as the base value and the 1.2 times standardized slope value. The larger the absolute value of the slope, the higher the decline coefficient and the lower the carbon density value. At the same time, an abnormal slope warning mechanism is established. When the withering slope of a certain area deviates from the mean of the same type of vegetation by more than 2 standard deviations, the area is marked as a suspicious area to be verified in the field.

[0042] The above steps revolve around the calculation and application of the NDVI change slope during phenological periods. The core principle lies in capturing the dynamic spectral characteristics of vegetation at different growth stages to enhance the accurate differentiation of vegetation types during the withering period. Specifically, firstly, the NDVI data after preprocessing for key phenological periods undergoes time-series smoothing to eliminate interference from short-term environmental fluctuations, making the data more reflective of the true trend of vegetation growth. When calculating the slope, a weighted linear regression algorithm is used to process the data from the peak growth period and the withering period separately. Especially in the slope calculation for the withering period, higher weights are assigned to the data from the final stage to highlight the spectral attenuation characteristics of the vegetation decline stage, making the slope values ​​more consistent with the actual changes in the vegetation withering process. Next, relying on the vegetation slope distribution characteristics accumulated in the historical vegetation classification sample database, a dynamic clustering algorithm is used to divide the slope intervals. Differentiated adaptive thresholds are set for the growth characteristics of different vegetation types, forming a threshold system adapted to the growth patterns of various vegetation types, providing a clear basis for subsequent classification verification. Finally, these thresholds are embedded as independent validation nodes into the decision tree model to cross-validate the preliminary classification results. When the slope value of a sample exceeds the threshold range of the corresponding type, a secondary discrimination mechanism is triggered to correct the classification result. Simultaneously, a correlation model between the validated slope value and biomass decline is constructed, making the allocation of vegetation carbon density parameters more closely reflect the actual growth state of vegetation. Furthermore, an abnormal slope warning mechanism is established to promptly mark areas with abnormal growth states, providing guidance for field verification. Thus, this example not only improves the sensitivity of the decision tree model to identifying differences in vegetation phenology during the withering period and reduces classification confusion among similar vegetation types, but also achieves dynamic optimization of vegetation carbon density parameters, providing more accurate basic data for carbon sink estimation. At the same time, the abnormal warning mechanism enhances the reliability and verifiability of the model results.

[0043] As an optional embodiment, after step S103, the tide table data of the target coastal salt marsh area, the real-time tide level data monitored by the hydrological station, and the salinity monitoring data of the target coastal salt marsh area can be preprocessed to remove outliers and fill in missing values. For example, an outlier detection algorithm can be used to identify and remove abnormal records caused by instrument malfunction or extreme weather. For missing values ​​that occur during data collection, time series interpolation methods can be used to fill in the missing values ​​to ensure the continuity and integrity of the data. Furthermore, based on the preprocessed tide table data and tide level data, the correlation between tide level and vegetation type distribution is analyzed. The tide level classification threshold is determined by combining the growth characteristics of Suaeda salsa and Spartina alterniflora. A tide level threshold node is added to the constructed phenological decision tree model. The tide level threshold node is set to prioritize the classification of Suaeda salsa in the low tide bare area and prioritize the classification of Spartina alterniflora in the high tide area. Specifically, based on preprocessed tide and tidal level data, the distribution patterns of vegetation types within different tidal level intervals are analyzed. Combining the characteristics that Suaeda salsa prefers to grow in the lower and middle parts of the intertidal zone and is more densely distributed in the bare areas at low tide, while Spartina alterniflora is more adapted to the growth characteristics of higher tidal levels, a tidal level classification threshold is determined. A tidal level threshold node is added to the constructed phenological decision tree model. When the model classifies the vegetation type of a certain area, if the area belongs to the bare area at low tide, it is preferentially classified as Suaeda salsa; if it is in the high tide area, it is preferentially classified as Spartina alterniflora. By incorporating the ecological relationship between tidal level and vegetation distribution, the rationality of the classification logic is improved. Based on this, and considering the influence of tidal height and salinity on the estimated net exchange volume of the ecosystem, combined with the difference analysis between measured vegetation distribution data and the preliminary classification results of the model under different salinity gradients, salinity correction factors and applicable salinity ranges were determined for each of the following species: Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. This resulted in a ternary correspondence table containing salinity ranges, vegetation types, and correction factors. For example, a positive correction was applied to the classification results of Suaeda salsa in high-salinity environments, and the classification weight of Phragmites australis was optimized in medium-salinity ranges. Finally, the salinity correction factors were integrated into the classification result output layer of the phenological decision tree model to optimize its classification accuracy. For instance, after the model completes the preliminary classification, the corresponding correction factors are used to fine-tune the classification results based on the salinity data of the region, further reducing the classification bias caused by salinity differences. Ultimately, through multi-dimensional optimization of tidal and salinity data, the classification accuracy of the phenological decision tree model is made to better reflect the actual vegetation distribution in the complex environment of coastal salt marshes.

[0044] It is worth noting that in the above embodiments, the acquired tidal and tidal level data were first systematically organized. Anomaly detection algorithms were used to remove abnormal fluctuations caused by instrument malfunctions or extreme weather, and time series interpolation was employed to fill in missing values ​​that occurred during monitoring, ensuring the continuity and reliability of the data. Based on this, and considering the topographic features of the target coastal salt marsh area, the tidal level data was divided into multiple continuous intervals according to spatial distribution. The distribution frequency of Suaeda salsa, Spartina alterniflora, and other vegetation types in historical vegetation survey data within each interval was statistically analyzed to determine the intrinsic correlation between different tidal environments and vegetation community distribution. According to the vegetation growth characteristics discovered in the ecological survey, Suaeda salsa, as a typical intertidal vegetation, depends on a certain duration of low-tide exposure, and its distribution is most dense in the exposed areas of the lower intertidal zone. Spartina alterniflora, on the other hand, has stronger flood tolerance and tends to form dominant communities in areas with relatively high tides and high flooding frequency. Based on these differences in ecological adaptability, the key threshold for tidal classification was determined by statistically analyzing the dominance ratio of the two vegetation types within different tidal ranges. Specifically, when the tide level is below a certain value, the area is defined as a low-tide bare zone; when the tide level is above another value, it is classified as a high-tide zone. This tidal threshold node was added to the constructed phenological decision tree model as an important supplement to the model's classification logic. When the model classifies the vegetation type of a region, it first calls the tidal attribute data for that region. If the data shows that the region belongs to a low-tide bare zone, the model will prioritize classifying it as *Suaeda salsa* based on NDVI threshold and phenological slope analysis, as this environmental condition highly matches the growth requirements of *Suaeda salsa*. If the region is in a high-tide zone, then considering the flood tolerance of *Spartina alterniflora*, the model will prioritize classifying it as *Spartina alterniflora* in the classification decision. The above embodiments fully incorporate the natural ecological relationship between tide level and vegetation distribution. When the model has ambiguity in NDVI features or phenological slope features, which leads to uncertainty in classification results, the tide level threshold node can provide clear ecological logic support, reduce misclassification caused by similar spectral features, and make the classification decision more in line with the natural distribution law of coastal salt marsh vegetation, thereby improving the accuracy and rationality of the entire phenological decision tree model in identifying vegetation types.

[0045] As an optional embodiment, after integrating the salinity correction factor into the classification result output layer of the phenological decision tree model in the above steps, a three-dimensional decision rule containing three dimensions—tide level, salinity, and vegetation type—can be constructed based on the added tide level threshold node, the determined salinity correction factor, and the original vegetation classification features in the phenological decision tree model. The original vegetation classification features include one of the following: NDVI threshold, phenological slope. The three-dimensional decision rule is used to represent the correspondence between tide level intervals, salinity ranges, and classification results for species such as *Suaeda salsa*, *Phragmites australis*, *Spartina alterniflora*, and *Ligusticum striatum*. Furthermore, a typical vegetation distribution area of ​​the target coastal salt marshland is selected as the validation dataset. The three-dimensional decision rule is applied to the model classification process, and the classification accuracy is calculated by comparing the measured vegetation types in the validation dataset with the output results of the phenological decision tree model. Finally, based on the accuracy calculation results, the division threshold of the tide level threshold node and the values ​​of the salinity correction factor and the applicable salinity range are iteratively optimized. If the vegetation classification error in a certain tide level range exceeds the preset value, the tide level threshold of the target coastal salt marsh area is adjusted. If the classification deviation is large in a certain salinity range, the salinity correction factor of the corresponding vegetation type is corrected until the model classification accuracy meets the preset requirements, so as to optimize the classification accuracy of the phenological decision tree model.

[0046] In the above embodiments, the classification accuracy of the phenological decision tree model is improved by constructing multi-dimensional decision rules and iteratively optimizing parameters. After integrating the salinity correction factor into the model output layer, based on the added tide threshold node, the determined salinity correction factor, and the vegetation classification features such as the original NDVI threshold and phenological slope in the model, a three-dimensional decision rule is constructed, including tide level, salinity, and vegetation type. This clarifies the classification relationships of Suaeda salsa, Phragmites australis, Spartina alterniflora, and Scutellaria baicalensis within different tide level and salinity ranges. For example, in areas with moderate tide levels and low salinity, the rule will preferentially match Phragmites australis; in areas with high tide levels and moderate salinity, it tends to match Spartina alterniflora. Furthermore, typical areas with known vegetation distribution within the target area are selected as validation datasets, such as the concentrated areas of Suaeda salsa in the low tide bare area and the Spartina alterniflora growth area in the high tide area, and the three-dimensional decision rules are applied to the model classification process. By comparing the vegetation types observed in the field in the validation dataset with the model output results, the overall classification accuracy and the confusion matrix of each type of vegetation are calculated. If the model misclassifies some reeds as Spartina alterniflora within a certain high tide range, and the error exceeds the preset value, it indicates that the tide threshold division for that range is unreasonable, and the upper limit of the tide threshold needs to be adjusted to narrow the high tide range. If the classification deviation of Suaeda salsa is large in high salinity areas, the salinity correction factor for Suaeda salsa in that salinity range should be corrected to enhance its classification weight in high salinity environments. After multiple rounds of parameter iteration and adjustment, until the model classification accuracy reaches the preset standard, the classification accuracy of the phenological decision tree model is optimized.

[0047] Understandably, the three-dimensional decision rule is a multi-dimensional classification system built during the optimization of the phenological decision tree model. Its core is to integrate the relationships between three key elements—tide level, salinity, and vegetation type—to form precise classification criteria. Based on existing vegetation classification features such as the NDVI threshold and phenological slope in the model, it integrates newly added tide level threshold nodes and salinity correction factors to construct a three-dimensional correspondence framework encompassing tide level ranges, salinity ranges, and vegetation types, clearly defining the priority classification logic for various vegetation types under different environmental conditions. In practical applications, the three-dimensional decision rule sets differentiated judgment logic based on the ecological adaptability characteristics of the vegetation. For example, considering the growth characteristics of Suaeda salsa, which prefers low tides and tolerates high salinity, the rule will prioritize Suaeda salsa in areas with low tides and high salinity. For Reeds, which are more adapted to low to medium salinity and medium tides, the rule will prioritize their classification in areas with medium tides and low salinity. However, Spartina alterniflora, adapted to higher tides, will tend to be classified as Spartina alterniflora in areas with high tides and medium salinity. This rule system is not fixed, but is closely linked to the ecological habits of vegetation, forming a dynamically adaptable classification guide.

[0048] For example, when the model classifies vegetation in a certain area, the 3D decision rule first determines the tidal range based on the tidal data of that area, then matches the corresponding salinity range with salinity monitoring data, and finally outputs the most likely vegetation type according to the preset 3D correspondence. Through this multi-dimensional cross-validation method, the 3D decision rule can effectively make up for the limitations of single-feature classification, reduce classification bias caused by environmental factors, provide the model with a more comprehensive and ecologically accurate classification basis, and thus improve the accuracy of the phenological decision tree model in distinguishing coastal salt marsh vegetation types.

[0049] Step S104: Input the initial vegetation outline into the phenological decision tree model. The phenological decision tree model performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation coverage of Suaeda salsa, Reed, Spartina alterniflora, and Ligusticum striatum within the target coastal salt marsh area.

[0050] As an optional embodiment, in step S104, the initial vegetation outline is converted into a model input format at the pixel scale, and combined with the preprocessed key phenological period NDVI data and vegetation index combination data, and input into the phenological decision tree model. Then, the model's multi-level classification process is initiated. The first-level classification is based on the dynamic initial NDVI threshold range for coarse classification, removing interfering pixels that do not conform to the basic vegetation spectral characteristics, and selecting a set of vegetation pixels including Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum as the first-level classification result. The second-level classification utilizes a machine learning-enhanced node partitioning mechanism. Based on the weighted ranking of core features, a random forest optimization algorithm incorporating dynamic weight allocation is used to select key phenological periods and index combinations. Combining the growth characteristics of coastal salt marsh vegetation at different phenological periods, dynamically changing weights are assigned to the vegetation indices for these key phenological periods. The selected key phenological periods and index combinations are then used to dynamically classify the first-level classification results, prioritizing the distinction between vegetation types whose NDVI feature differences meet set criteria. For example, *Spartina alterniflora* exhibits high NDVI values, while *Suaeda salsa* exhibits medium to low NDVI values, resulting in the second-level classification. The third-level classification employs a phenological slope verification mechanism. For easily confused vegetation types in the second-level classification results, the slope of NDVI changes during the withering period (August-November) is compared with an adaptive threshold. Pixels with slope values ​​matching the *Spartina alterniflora* threshold range are labeled as *Spartina alterniflora*, and those matching the *Phragmites australis* threshold range are labeled as *Phragmites australis*, resulting in the third-level classification. Next, a three-dimensional decision rule integrating tide level threshold nodes and salinity correction factors was used to spatially validate the classification results at each level. Vegetation pixels in the low-tide exposed area were optimized for Suaeda salsa classification boundaries using the salinity correction factor, while pixels in the high-tide area were used to strengthen the classification weight of Spartina alterniflora. Finally, morphological post-processing was used to remove isolated noise pixels from the classification results at each level, smoothing the vegetation cover boundaries and outputting the differentiated vegetation cover vector data for Suaeda salsa, Phragmites australis, Spartina alterniflora, and Spartina alterniflora.

[0051] For example, in an optional embodiment of step S104, the multi-level classification process of vegetation types achieves accurate differentiation through multi-level feature screening and cross-validation. First, the initial vegetation outline is converted into a model-recognizable input format at the pixel scale. Simultaneously, pre-processed key phenological period NDVI data and vegetation index combinations such as NDVI and EVI2 are imported to provide multi-dimensional basic features for model classification. The multi-level classification process begins with the first-level coarse classification. Based on the dynamic initial NDVI threshold range, non-vegetation pixels such as water bodies and bare beaches are removed, retaining only the set of pixels that conform to the spectral characteristics of vegetation. For example, pixels with NDVI < 0.2 during the peak growth period are judged as non-vegetation interference items, and vegetation pixels that may include Suaeda salsa, Reed, Spartina alterniflora, and Ligusticum striatum are selected to form the first-level classification result. The second-level classification calls a machine learning-enhanced node partitioning mechanism, assigning dynamic weights to vegetation indices at different phenological periods through a random forest algorithm. For example, the weight of EVI2 for Spartina alterniflora is increased in August, and the weight of NDVI for Reed is increased in June. Key feature combinations are selected using weight ranking. Based on these characteristics, the primary results are further classified, prioritizing types with significant spectral differences. For example, pixels in the high NDVI value range are initially labeled as *Spartina alterniflora*, and those in the medium-to-low NDVI value range are labeled as *Suaeda salsa*, resulting in secondary classification results. For the tertiary classification, phenological slope verification is used for easily confused types in the secondary results (such as reeds and *Spartina alterniflora*). By comparing the NDVI change slope during the withering period from August to November with an adaptive threshold, pixels with slopes matching the attenuation characteristics of *Spartina alterniflora* are confirmed, while those matching the characteristics of reeds are relabeled to reduce type confusion.

[0052] Subsequently, spatial verification is performed using integrated three-dimensional decision rules. Pixels in the low-tide exposed area have their boundaries refined using salinity correction factors, while pixels in the high-tide area have their classification weights for Spartina alterniflora strengthened using tidal thresholds, ensuring consistency between classification results and ecological adaptability characteristics. For example, in the spatial verification stage of multi-level vegetation classification, the core logic of integrating three-dimensional decision rules is to deeply bind the vegetation classification results with the ecological correlation of environmental factors such as tidal level and salinity, and to optimize the classification boundaries through targeted adjustments. For pixels in the low-tide exposed area, the low-tide interval to which they belong is first determined based on tidal data. These areas are naturally suited to the growth habits of Suaeda salsa due to their short flooding time and high exposure frequency. At this point, combining salinity monitoring data with the salinity correction factor, if a pixel is in a high-salinity environment and there is a blurred boundary between Suaeda salsa and other vegetation in the primary and secondary classifications, the classification signal of Suaeda salsa is strengthened by increasing the weight of the salinity correction factor. For example, in areas where salinity exceeds a certain threshold, pixels with previously blurred boundaries are automatically classified into the Suaeda salsa category, thereby refining the spatial distribution boundary of Suaeda salsa and ensuring that its coverage matches the typical habitat characteristics of "low tide plus high salinity". For pixels in the high tide area, the classification weight of Spartina alterniflora is strengthened based on the setting of the tide level threshold node. The high tide area has a longer tidal submersion time, and the water conditions are more suitable for the growth of Spartina alterniflora. When the model confuses the vegetation type of a pixel with Spartina alterniflora and Reed in the secondary or tertiary classification, the weight adjustment mechanism is triggered by calling the tide level attribute label to automatically increase the classification priority of Spartina alterniflora in that area. For example, if a pixel's NDVI feature and slope value simultaneously meet some of the criteria for both Spartina alterniflora and Phragmites australis, the system will prioritize classifying it as Spartina alterniflora based on its location in the high tide zone, avoiding classification bias caused by similar spectral features. Through this spatial verification combined with specific habitat conditions, the final vegetation classification results not only conform to spectral feature patterns but also highly align with the ecological adaptability characteristics of plants such as Suaeda salsa and Spartina alterniflora, further improving the accuracy and rationality of classification boundaries.

[0053] Finally, morphological post-processing is used to remove isolated noise pixels and smooth the boundaries, outputting the cover vector data of the three types of vegetation, achieving a step-by-step optimization from coarse to fine classification.

[0054] Step S105: Obtain the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. Combine the vegetation cover range corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, and complete the estimation of the total carbon content of the target coastal salt marsh area through spatial overlay calculation.

[0055] As an optional embodiment, in step S105, sampling points are set up in the target coastal salt marsh area according to a unit grid of a first preset size. Field sampling is conducted on four types of vegetation: Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. The number of samples for each type of vegetation is no less than a set number. By collecting vegetation samples and measuring the carbon content of each type of vegetation, a basic carbon density raw dataset containing the coordinates of the sampling points and carbon content values ​​is generated. Then, the basic carbon density raw dataset is spatially matched with regional-scale vegetation carbon density literature data. Kriging interpolation is used to spatially interpolate and calibrate the field sampling data. Cross-validation is used to adjust the interpolation parameters to expand the coverage of the sampling point data, so that the calibrated basic carbon density parameters are spatially compatible with the vegetation growth environment characteristics of the target area. For example, in the target coastal salt marsh area, sampling points are set up in a 100m × 100m grid, and sampling is conducted on the four types of vegetation: Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. At least 30 samples are collected for each type of vegetation. For example, in a Spartina alterniflora growing area, plant samples are collected from different locations and brought back to the laboratory to determine their carbon content. This generates basic data containing the coordinates of each sampling point and its corresponding carbon content. For instance, the carbon content of Spartina alterniflora corresponding to the coordinates (X1, Y1) of a sampling point is denoted as 'a'. Then, this basic data from the field sampling is correlated with existing literature data on vegetation carbon density in the area, and the sampling data is expanded using Kriging interpolation. Just as the carbon content of known sampling points is used as a reference, the carbon density of unsampled areas is calculated through interpolation. Then, parameters are adjusted through cross-validation to ensure that the adjusted carbon density distribution matches the local vegetation growth environment. For example, the carbon density parameters in areas with sufficient water will be adapted to the characteristics of more lush vegetation.

[0056] Then, based on the interpolated baseline carbon density parameters, a multi-maintenance positive matrix was established, incorporating the NDVI change slope during the August-November withering period, salinity monitoring data, tide data, and correction coefficients. Specifically, for the NDVI change slope during the August-November withering period in the phenological dynamics, a carbon density correction coefficient was adjusted down by 0.02 for every 0.01 / month increase in the absolute value of the slope, and this was associated with a pixel-level slope label. For salinity in the environmental gradient characteristics, a reed carbon density correction coefficient was adjusted down by 0.05 for every 5‰ increase in salinity, and this was mapped to a salinity interval label. For tide level, a tide level correction coefficient was set for Spartina alterniflora in the high tide zone, and a set multiple of the baseline value was assigned to Suaeda salsa based on its low tide exposure duration, with the baseline value increasing by 0.1 for every 2 hours of exposure duration, and this was bound to a tide level attribute label. For example, a multi-maintenance positive matrix was established based on the interpolated baseline carbon density parameters. If the absolute value of the NDVI change slope of Spartina alterniflora in a certain area increases by 0.01 per month from August to November, the corresponding carbon density correction factor will be lowered by 0.02 and associated with the slope label of the pixel in that area; if the salinity of a certain area increases by 5‰, the carbon density correction factor of Reed will be lowered by 0.05 and associated with the salinity range label; a tide level correction factor is set for Spartina alterniflora in the high tide area, such as 1.1, and for every 2 hours increase in the low tide exposure time in the Suaeda salsa area, the baseline value will be increased by 0.1. All of these are bound to the tide level attribute label.

[0057] Among them, the multi-correction positive matrix is ​​a parameter adjustment tool that integrates multiple influencing factors such as phenological dynamics and environmental gradients to accurately correct the basic carbon density parameters of different vegetation types. By establishing the correlation between phenological characteristics, environmental factors, and correction coefficients, it achieves refined calibration of carbon density. Specifically, this matrix covers correction logic in three key dimensions: First, phenological dynamics, using the slope of NDVI change during the withering period from August to November as the core indicator, binding the vegetation growth decline rate to the carbon density correction coefficient; the larger the absolute value of the slope, the faster the vegetation declines, corresponding to a downward adjustment of the carbon density correction coefficient. Second, environmental salinity gradient, based on the impact of different salinity ranges on vegetation growth, setting specific correction rules for salinity-sensitive vegetation; when salinity increases, the correction coefficient is adjusted downward accordingly. Third, tidal environmental characteristics, considering the differences in tidal adaptability between Spartina alterniflora and Suaeda salsa, setting a tidal correction coefficient for Spartina alterniflora in the high-tide zone, and associating the low-tide exposure time with the baseline value fluctuation rule for Suaeda salsa; the longer the exposure time, the greater the upward fluctuation of the correction coefficient. Through this multi-dimensional parameter correlation, the multi-dimensional positive matrix can closely integrate the basic carbon density parameters with the actual growth status of vegetation and environmental adaptability, making the carbon density assignment more consistent with the complex and ever-changing ecological environment of coastal salt marshes, and providing accurate parameter support for subsequent pixel-level carbon density calculation and total carbon sink estimation.

[0058] Furthermore, a differentiated parameter calling model is constructed based on a multi-maintenance positive matrix. This model includes models for *Spartina alterniflora*, *Phragmites australis*, and *Suaeda salsa*. The *Spartina alterniflora* model takes a base carbon density value as input, and calculates the output by weighting the standardized slope value taken from the pixel slope label with the tide level correction coefficient taken from the tide level attribute label. The *Phragmites australis* model multiplies the correction coefficient corresponding to the salinity interval label with the base value. The *Suaeda salsa* model determines the baseline value fluctuation multiple based on the low tide exposure duration label, forming a three-dimensional carbon density parameter library containing data association rules. The vector data of vegetation cover areas corresponding to *Suaeda salsa*, *Phragmites australis*, *Spartina alterniflora*, and *Ligusticum striatum* are converted into pixel rasters of a second preset size. Through spatial coordinate association, the phenological slope label, salinity interval label, and tide level attribute label of the pixel are matched with the three-dimensional carbon density parameter library. The corresponding model is then called to complete the pixel-level carbon density assignment, generating a carbon density raster dataset. Based on the aforementioned carbon density raster dataset, a hierarchical weighted summation algorithm is used to calculate carbon sink, where cell weights are determined by weighted fusion of classification confidence and August NDVI values. For example, when constructing differentiated models, the Spartina alterniflora model multiplies the base carbon density value by (1 plus 0.1 times the standardized slope value) and then by the tide level correction coefficient. For instance, if the base value is b, the standardized slope value is c, and the tide level correction coefficient is 1.1, the result is b × (1 + 0.1c) × 1.1. The Reed model multiplies the base value by the correction coefficient corresponding to the salinity range. If the salinity is in the ≤10‰ range, the correction coefficient is 1.05, and the base value is d, resulting in d × 1.05. The Suaeda salsa model determines the baseline value fluctuation factor based on the low tide exposure duration. For example, if the exposure is 10 hours, the baseline value increases by 0.1 × ((10-8) / 2) = 0.1, forming a three-dimensional carbon density parameter library. The vector data of the coverage of the three types of vegetation are converted into 30m×30m pixel grids. By matching the coordinates of the pixels with various labels and parameter libraries, a carbon density value is assigned to each pixel to generate a carbon density raster dataset.

[0059] Finally, a sampling area representing a predetermined proportion of the total area is selected. The measured values ​​of soil profile carbon storage in the sampling area are compared with the model estimates for the corresponding area. An error correction coefficient is calculated and fed back to the carbon density raster dataset for correction. The output includes the estimated total carbon sink of the target area, the carbon sink contribution rate of each vegetation type, and a carbon density spatial distribution heatmap. Continuing the example above, based on the carbon density raster dataset from the previous example, a hierarchical weighted summation algorithm is used to calculate the carbon sink. Pixel weights are combined with classification confidence and August NDVI values; for example, pixels with high classification confidence and large August NDVI have higher weights. Finally, a 5% area is selected as the sampling area. The measured values ​​of soil profile carbon storage in this area are compared with the model estimates to calculate the error correction coefficient. This coefficient is used to correct the carbon density raster dataset, ultimately outputting the total carbon sink, the carbon sink contribution rate of each vegetation type, and a carbon density heatmap.

[0060] Alternatively, in the process of estimating carbon sinks in coastal salt marshes, the accuracy and completeness of the model can be improved through multi-source data fusion and technological collaboration. Eddy covariance data from flux towers can be introduced, and the vegetation carbon sink model can be calibrated using the results of ecosystem net exchange observations, making the model output more closely match the actual ecosystem carbon exchange process. For example, using real-time ecosystem net exchange data captured by flux towers to reflect the dynamics of carbon exchange between the ecosystem and the atmosphere, these observation results can be compared and analyzed with the simulated output of the vegetation carbon sink model. By adjusting the carbon allocation parameters and phenological response coefficients in the model, the carbon exchange process output by the model can be made consistent with the diurnal and seasonal variation patterns of the actual ecosystem.

[0061] Simultaneously, underground biomass sampling can be combined with laboratory analysis of root carbon content to supplement blind spots in remote sensing technology for monitoring underground carbon storage, thereby improving the comprehensive coverage of carbon sink assessment. For example, addressing the limitation of remote sensing technology in penetrating the surface to monitor underground carbon storage, underground biomass sampling can be conducted. By collecting root samples at different depths in typical vegetation communities and obtaining root carbon density data through laboratory carbon content analysis, these measured data can be used as reference values ​​to supplement the model, correcting the bias in underground carbon storage retrieved by remote sensing, and achieving comprehensive carbon sink assessment coverage from aboveground to underground. Building on this, machine learning enhancement nodes can be applied to the assimilation process of flux data. Drawing on the idea of ​​using flux data to correct light energy utilization models, observational data can be used to dynamically optimize the classification rules of the decision tree, making vegetation type classification and carbon sink parameter assignment more adaptable to the actual dynamic changes of the ecosystem. Specifically, flux observation data is input into the decision tree model as dynamic constraints. The judgment threshold of vegetation classification nodes and the assignment rules of carbon sink parameters are adjusted in real time through the feature importance reordering algorithm, so that the vegetation type classification results are more in line with the actual community distribution and the carbon density parameter is accurately optimized with the dynamic changes of the ecosystem.

[0062] To construct a more comprehensive carbon sink assessment system, a sediment deposition rate inversion step can be optionally added. This involves using suspended sediment concentration data obtained through remote sensing technology to invert regional sedimentation conditions, and simultaneously establishing a soil carbon burial assessment module through laboratory calibration of soil samples' organic carbon content. Combining vegetation carbon storage and soil carbon burial creates a dual-channel assessment framework that includes both vegetation carbon sinks and soil carbon burial, more comprehensively reflecting the carbon sink function of coastal salt marsh ecosystems.

[0063] For example, remote sensing technology is used to conduct multi-period observations of the target coastal salt marsh area to obtain suspended sediment concentration data at different times. This data can reflect the distribution and migration of sediment in the water. Based on the spatial distribution characteristics and variation patterns of suspended sediment concentration, combined with a regional hydrodynamic model, the sediment deposition rate of the entire region is obtained. The principle is that suspended sediment settles under the action of water flow, and the deposition rate is intrinsically related to factors such as suspended sediment concentration and water flow velocity. This correlation is captured by remote sensing data and transformed into a quantitative deposition rate index, thereby clarifying the sediment accumulation status in the salt marsh area. Simultaneously, soil samples are collected. Representative soil profiles are selected in different vegetation cover areas and topographic locations of the salt marsh. Soil samples are collected layer by layer, and the organic carbon content is determined in the laboratory to determine the amount of organic carbon stored in the soil at different depths. This establishes a soil carbon burial assessment module. The core principle is that organic carbon in the soil is gradually buried and accumulated with sediment deposition. By calibrating the soil organic carbon content in different areas, the potential and current status of soil carbon burial can be quantified.

[0064] Furthermore, the vegetation carbon storage calculated using vegetation classification and carbon density parameters is integrated with the soil carbon sequestration obtained from sediment deposition rate and soil organic carbon content analysis, forming a dual-channel assessment framework that includes both vegetation carbon sinks and soil carbon sequestration. Vegetation carbon sinks reflect the process of carbon fixation by vegetation through photosynthesis, while soil carbon sequestration reflects the long-term storage of carbon carried by sediment deposition in the soil. Together, they constitute an important component of the carbon sink in coastal salt marsh ecosystems. This dual-channel framework can more comprehensively cover the carbon cycle process of the ecosystem, making up for the limitations of relying solely on vegetation carbon storage assessment, thus more completely reflecting the overall function of coastal salt marsh ecosystems in carbon sequestration and enhancement.

[0065] Furthermore, it integrates lidar data to improve the accuracy of vegetation canopy structure analysis, especially for low-coverage woody vegetation, addressing the underestimation of carbon storage by optical remote sensing in this type of vegetation. In conjunction with the needs of blue carbon trading scenarios, it generates a carbon sink spatial distribution heatmap based on the output carbon aggregate, and clarifies the uncertainty range of the assessment results, providing accurate spatial data support for carbon credit accounting and enhancing the practicality and reliability of carbon sink estimation results in real-world applications.

[0066] In this embodiment, from the data source perspective, remote sensing images were acquired during key phenological periods to capture the spectral differences of vegetation at different growth stages, laying a high-quality data foundation for subsequent analysis. Unsupervised classification methods, combined with targeted preprocessing, effectively eliminated noise effects such as tidal interference, and the extracted initial vegetation contours accurately distinguished between vegetated and non-vegetated areas, ensuring the reliability of the classification starting point. The construction of the phenological decision tree model in this embodiment is a core innovation. Based on the NDVI thresholding method, and incorporating a machine learning-enhanced node partitioning mechanism, it can autonomously mine key classification features and optimize node priorities, improving the accuracy of vegetation type differentiation. Furthermore, the introduction of phenological slope analysis further strengthens the ability to identify differences in vegetation attenuation during the withering period, significantly reducing the confusion rate of similar vegetation types and making the classification results for Suaeda salsa, Phragmites communis, Spartina alterniflora, and Ligusticum striatum more accurate. In the carbon sink estimation stage, precise vegetation cover range is combined with corresponding carbon density parameters through spatial overlay calculations. Simultaneously, a dual-channel assessment framework constructed by combining sediment deposition inversion and soil organic carbon calibration is employed, encompassing both vegetation carbon sinks and soil carbon burial, comprehensively covering key processes in the ecosystem's carbon cycle. Overall, the embodiments of this application achieve end-to-end optimization from fine vegetation classification to accurate estimation of total carbon content, effectively compensating for the shortcomings of traditional methods in dynamic feature capture, classification accuracy, and assessment completeness, providing reliable technical support for the scientific assessment of carbon sink functions in coastal salt marsh ecosystems.

[0067] Please see Figure 2 , Figure 2 A system for estimating carbon sinks in coastal salt marsh vegetation based on multi-temporal phenological characteristics is provided for embodiments of this application. The system includes: an acquisition module for acquiring remote sensing images of a target coastal salt marsh area during key phenological periods; The classification module is used to process the remote sensing image using an unsupervised classification method to extract the initial vegetation outline of the vegetation cover area within the target region. The module is used to construct a phenological decision tree model based on the NDVI threshold method. The model incorporates a machine learning-enhanced node partitioning mechanism to optimize the node hierarchy and judgment priority of the decision tree. The phenological slope analysis method is used to strengthen the ability of the phenological decision tree model to identify the phenological differences of various types of vegetation during the withering period. The decision module is used to input the initial vegetation outline into the phenological decision tree model, and the phenological decision tree model performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation coverage range of Suaeda salsa, Reed, Spartina alterniflora and Ligusticum striatum in the target coastal salt marsh area. The estimation module is used to obtain the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. Combined with the vegetation cover range corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, the total carbon content of the target coastal salt marsh area is estimated by spatial overlay calculation.

[0068] In some implementations, the coastal salt marsh vegetation carbon sink estimation system based on multi-temporal phenological characteristics can be applied to terminal devices. It should be noted that, for the sake of convenience and brevity, the specific working process of the coastal salt marsh vegetation carbon sink estimation system based on multi-temporal phenological characteristics described above can be referred to the corresponding process in the aforementioned embodiments of the coastal salt marsh vegetation carbon sink estimation method based on multi-temporal phenological characteristics, and will not be repeated here.

[0069] Please see Figure 3 , Figure 3 This is a schematic block diagram of the structure of a terminal device provided in an embodiment of this application.

[0070] like Figure 3 As shown, the terminal device 300 includes a processor 301 and a memory 302, which are connected via a bus 303, such as an I2C bus. Specifically, the processor 301 provides computing and control capabilities to support the operation of the entire terminal device. The processor 301 can be a central processing unit, or it can be other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the embodiments of this application and does not constitute a limitation on the terminal device to which the embodiments of this application are applied. A specific server may include more or fewer components than shown in the diagram, or combine certain components, or have different component arrangements. The processor is used to run a computer program stored in the memory, and when executing the computer program, implements any of the methods provided in the embodiments of this application for estimating carbon sinks in coastal salt marshes based on multi-temporal phenological characteristics. It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the terminal device described above can be referred to the aforementioned embodiments of the method for estimating carbon sinks in coastal salt marshes based on multi-temporal phenological characteristics, and will not be repeated here.

[0071] This application also provides a storage medium for computer-readable storage, wherein the storage medium stores one or more programs that can be executed by one or more processors to implement the steps of any of the methods for estimating carbon sinks of coastal salt marsh vegetation based on multi-temporal phenological characteristics as provided in the specification of this application.

Claims

1. A method for estimating carbon sequestration in coastal salt marsh vegetation based on multi-temporal phenological characteristics, characterized in that, The method includes: Acquire remote sensing images of the target coastal salt marsh area during key phenological periods; The remote sensing image is processed using an unsupervised classification method to extract the initial vegetation outline of the vegetation cover area within the target region; A phenological decision tree model was constructed based on the NDVI threshold method. A machine learning-enhanced node partitioning mechanism was incorporated into the phenological decision tree model to optimize the node hierarchy and judgment priority of the decision tree. The phenological slope analysis method was used to enhance the ability of the phenological decision tree model to identify the phenological differences of various types of vegetation during the withering period. The initial vegetation outline is input into the phenological decision tree model. The phenological decision tree model performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation coverage of Suaeda salsa, Reed, Spartina alterniflora, and Ligusticum striatum in the target coastal salt marsh area. Obtain the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. Combine the vegetation cover ranges corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, and complete the estimation of the total carbon content of the target coastal salt marsh area through spatial overlay calculation.

2. The method according to claim 1, characterized in that, The key phenological periods include at least one of the following: June, August, and November; acquiring remote sensing images of the target coastal salt marsh area during the key phenological periods includes: Remote sensing image data of the target coastal salt marsh area during at least one key phenological period in June, August, and November were collected using remote sensing satellites or airborne remote sensing equipment. The remote sensing image data included spectral information for extracting the NDVI vegetation index.

3. The method according to claim 1, characterized in that, The process of processing the remote sensing image using an unsupervised classification method to extract the initial vegetation contour of the vegetation cover area within the target region includes: To address the image noise characteristics of the target coastal salt marsh area caused by tidal interference and water reflection, the remote sensing image is preprocessed in a targeted manner. An adaptive filtering algorithm is used to remove salt-and-pepper noise and strip noise from the remote sensing image. In addition, radiometric bias correction and geometric fine correction are completed by combining regional topographic data and tidal correction model to ensure the accuracy of image spectral information and spatial location. For the preprocessed remote sensing images, an improved K-means unsupervised classification algorithm is used to perform hierarchical automatic clustering analysis on the land cover types in the preprocessed remote sensing images by dynamically adjusting the cluster centers and iterative thresholds. Based on the spectral response differences of the NDVI vegetation index, vegetation-covered areas and non-vegetation-covered areas are distinguished. By setting a threshold range for vegetation spectral features from the clustering results, the set of pixels corresponding to the vegetation coverage area is selected. Combined with morphological filtering to remove discrete noise pixels, the spatial distribution range of the pixel set is defined as the initial vegetation outline of the vegetation coverage range in the target coastal salt marsh area.

4. The method according to claim 3, characterized in that, The phenological decision tree model is constructed using the NDVI threshold method as its core. Machine learning is incorporated into this model to enhance node partitioning, optimizing the node hierarchy and decision priority. Furthermore, phenological slope analysis is used to strengthen the model's ability to identify phenological differences in various vegetation types during the withering period. This includes: Using the NDVI vegetation index extracted from the preprocessed remote sensing image as the core parameter, and combining the growth characteristics of coastal salt marsh vegetation to set a dynamic initial NDVI threshold range, a model framework for a phenological decision tree model adapted to the salt marsh environment is constructed. The random forest algorithm is embedded in the model framework as a machine learning enhancement node partitioning mechanism. The vegetation index combination data of each key phenological period after radiation bias correction and geometric fine correction are input. The vegetation index combination data includes NDVI and EVI2. The weight relationship between different key phenological periods and index combinations on vegetation classification is autonomously mined through feature importance ranking algorithm. The core features with the highest classification contribution among Suaeda salsa, Reed, Spartina alterniflora and Ligusticum striatum are identified. The node hierarchy order and judgment priority of the decision tree are dynamically adjusted according to the weight ranking of the core features. Based on NDVI data from key phenological periods, a linear regression algorithm was used to calculate the NDVI change slopes from June to August and August to November. The slope value from August to November was used as the core dynamic indicator of the phenological period slope analysis method. An adaptive threshold was set by comparing the difference in NDVI decay rates between Spartina alterniflora and Phragmites australis during the withering period. The adaptive threshold was embedded in the decision tree as a key verification node to form a secondary verification of the feature optimization results, thereby enhancing the sensitivity and recognition accuracy of the phenological decision tree model to vegetation phenological differences during the withering period, thus completing the construction of the phenological decision tree model.

5. The method according to claim 4, characterized in that, The method of constructing a phenological decision tree model based on the NDVI threshold method, incorporating a machine learning-enhanced node partitioning mechanism to optimize the node hierarchy and judgment priority of the decision tree, and employing phenological slope analysis to strengthen the model's ability to identify phenological differences in various vegetation types during the withering period, further includes: Preprocess the tide table data of the target coastal salt marsh area, the real-time tide level data monitored by the hydrological station, and the salinity monitoring data of the target coastal salt marsh area to remove outliers and fill in missing values; Based on the preprocessed tide table data and tide level data, the correlation between tide level and vegetation type distribution is analyzed. The tide level classification threshold is determined by combining the growth characteristics of Suaeda salsa and Spartina alterniflora. A tide level threshold node is added to the constructed phenological decision tree model. The tide level threshold node is set to prioritize the classification of Suaeda salsa in the low tide bare area and prioritize the classification of Spartina alterniflora in the high tide area. Based on the relationship between tidal height and salinity on the estimated net exchange of the ecosystem, and combined with the difference analysis between measured vegetation distribution data and preliminary model classification results under different salinity gradients, vegetation salinity correction factors and applicable salinity ranges for Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum were determined, forming a ternary correspondence table containing salinity range, vegetation type, and correction factor. The salinity correction factor was integrated into the classification result output layer of the phenological decision tree model to optimize the classification accuracy of the phenological decision tree model.

6. The method according to claim 5, characterized in that, After integrating the salinity correction factor into the classification result output layer of the phenological decision tree model, the method further includes: Based on the added tide level threshold node, the determined salinity correction factor, and the original vegetation classification features in the phenological decision tree model, a three-dimensional decision rule is constructed that includes three dimensions: tide level, salinity, and vegetation type. The original vegetation classification features include one of the following: NDVI threshold and phenological slope. The three-dimensional decision rule is used to represent the correspondence between the tide level interval, salinity range, and the classification results of Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. Typical vegetation distribution areas of the target coastal salt marsh region were selected as validation datasets. Three-dimensional decision rules were applied to the model classification process. The classification accuracy was calculated by comparing the measured vegetation types in the validation datasets with the output results of the phenological decision tree model. Based on the accuracy calculation results, the division threshold of the tide level threshold node and the values ​​of the salinity correction factor and the applicable salinity range are iteratively optimized. If the vegetation classification error in a certain tide level range exceeds the preset value, the tide level threshold of the target coastal salt marsh area is adjusted. If the classification deviation is large in a certain salinity range, the salinity correction factor of the corresponding vegetation type is corrected until the model classification accuracy meets the preset requirements, so as to optimize the classification accuracy of the phenological decision tree model.

7. The method according to claim 4, characterized in that, The NDVI data based on key phenological periods are used to calculate the NDVI change slopes from June to August and August to November using a linear regression algorithm. The slope value from August to November is used as the core dynamic indicator of the phenological period slope analysis method. An adaptive threshold is set by comparing the difference in NDVI decay rates between Spartina alterniflora and Phragmites australis during the withering period. This adaptive threshold is embedded in a decision tree as a key verification node to form a secondary verification of the feature optimization results, including: The NDVI data after preprocessing the key phenological periods were smoothed over time to eliminate short-term fluctuations. The NDVI change slope during the peak growth period from June to August and the withering period from August to November was calculated using a weighted linear regression algorithm. The NDVI data in November was given the highest weight in the calculation of the slope during the withering period to highlight the characteristics of the final decline. Based on the distribution characteristics of the withering slope of Spartina alterniflora and Phragmites communis in the historical vegetation classification sample database, the slope interval is divided by dynamic clustering algorithm, and differentiated adaptive thresholds are set for different vegetation types. For example, the withering slope threshold of Spartina alterniflora is set to -0.02 / month to -0.05 / month, and that of Phragmites communis is set to -0.01 / month to -0.03 / month, thus forming a threshold system adapted to each vegetation type. Thresholds are embedded in the decision tree model as independent validation nodes to cross-validate the preliminary classification results after feature optimization. When the slope value of a sample exceeds the threshold range of the corresponding vegetation type, a secondary discrimination mechanism is triggered. Based on the validated withering slope value, vegetation carbon density parameters are dynamically allocated to construct a correlation model between slope and biomass decline. For Spartina alterniflora and Phragmites australis, the calculation method is the product between the carbon density as the base value and the 1.2 times standardized slope value. The larger the absolute value of the slope, the higher the decline coefficient and the lower the carbon density value. At the same time, an abnormal slope warning mechanism is established. When the withering slope of a certain area deviates from the mean of the same type of vegetation by more than 2 standard deviations, the area is marked as a suspicious area to be verified in the field.

8. The method according to claim 1, characterized in that, The initial vegetation outline is input into the phenological decision tree model, which performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation coverage areas of Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum within the target coastal salt marsh area, including: The initial vegetation outline is converted into a model input format at the pixel scale, and combined with the preprocessed key phenological period NDVI data and vegetation index combination data, and then input into the phenological decision tree model. The multi-level classification process of the model is initiated. The first-level classification is based on the dynamic initial NDVI threshold range for coarse classification. Interference pixels that do not conform to the basic vegetation spectral characteristics are removed, and the set of vegetation pixels corresponding to Suaeda salsa, Reed, Spartina alterniflora and Ligusticum striatum are selected as the first-level classification result. The second-level classification utilizes a machine learning-enhanced node partitioning mechanism. Based on the weight ranking of core features, a random forest optimization algorithm with dynamic weight allocation is used to select key phenological periods and index combinations. Combining the growth characteristics of coastal salt marsh vegetation in different phenological periods, dynamically changing weights are assigned to the vegetation indices of key phenological periods. The selected key phenological periods and index combinations are used to dynamically classify the first-level classification results, prioritizing the distinction between vegetation types whose NDVI feature differences reach the set conditions, such as Spartina alterniflora in the high NDVI value range and Suaeda salsa in the medium-low NDVI value range, to obtain the second-level classification results. The third-level classification uses a phenological slope verification mechanism. For vegetation types that are easily confused in the second-level classification results, the slope of NDVI change during the withering period from August to November is compared with the adaptive threshold for discrimination. Pixels with slope values ​​that meet the threshold range of Spartina alterniflora are marked as Spartina alterniflora, and those that meet the threshold range of Reed are marked as Reed, thus obtaining the third-level classification results. A three-dimensional decision rule integrating tide level threshold nodes and salinity correction factors is used to spatially verify the classification results at each level. Vegetation pixels in the low tide bare area are combined with salinity correction factors to optimize the classification boundary of Suaeda salsa, while pixels in the high tide area are used to strengthen the classification weight of Spartina alterniflora. Isolated noise pixels in the classification results at each level are removed by morphological post-processing, vegetation cover boundaries are smoothed, and the vector data of vegetation cover for each of the distinguished species, such as Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, are output.

9. The method according to claim 1, characterized in that, The process of obtaining the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Scutellaria baicalensis, and combining these with the corresponding vegetation cover ranges, to estimate the total carbon content of the target coastal salt marsh area through spatial overlay calculations, includes: Sampling points were set up in the target coastal salt marsh area according to the first preset size of the unit grid. Field sampling was carried out on three types of vegetation: Suaeda salsa, Reed and Spartina alterniflora. The number of samples for each type of vegetation was no less than the set number. By collecting vegetation samples and measuring the carbon content of each type of vegetation, a basic carbon density raw dataset containing the coordinates of the sampling points and the carbon content values ​​was generated. The basic carbon density raw dataset was spatially matched with regional-scale vegetation carbon density literature data. Kriging interpolation was used to spatially interpolate and calibrate the field sampling data. Cross-validation was used to adjust the interpolation parameters to expand the coverage of the sampling point data, so that the spatial distribution of the calibrated basic carbon density parameters was adapted to the vegetation growth environment characteristics of the target area. Based on the interpolated basic carbon density parameters, a multi-maintenance positive matrix was established using the slope of NDVI change, salinity monitoring data, tide data, and correction coefficients during the wilt period from August to November. A differentiated parameter calling model is constructed based on a multi-maintenance positive matrix. The differentiated parameter calling model includes: Spartina alterniflora model, Phragmites australis model, and Suaeda salsa model. The Spartina alterniflora model takes the base carbon density value as input, and calculates the output by weighting the standardized slope value taken from the pixel slope label and the tide correction coefficient taken from the tide attribute label. The Phragmites australis model multiplies the correction coefficient corresponding to the salinity interval label with the base value. The Suaeda salsa model determines the baseline value fluctuation multiple based on the low tide exposure duration label, forming a three-dimensional carbon density parameter library containing data association rules. The vector data of vegetation cover range of each of the following species are converted into a second preset size of cell raster. The phenological slope label, salinity interval label and tidal attribute label of the cell are matched with the three-dimensional carbon density parameter library through spatial coordinate association. The corresponding model is called to complete the cell-level carbon density assignment and generate a carbon density raster dataset. Based on the aforementioned carbon density raster dataset, a hierarchical weighted summation algorithm was used to calculate the carbon sink, where the cell weights were determined by weighted fusion of classification confidence and August NDVI values; A sampling area representing a predetermined proportion of the total area is selected. The measured values ​​of soil profile carbon storage in the sampling area are compared with the model estimates of the corresponding area. The regional error correction coefficient is calculated and fed back to the carbon density raster dataset for correction. The output includes the estimated results of the total carbon sink in the target area, the carbon sink contribution rate of each vegetation type, and the spatial distribution heat map of carbon density.

10. A carbon sequestration estimation system for coastal salt marsh vegetation based on multi-temporal phenological characteristics, characterized in that, The system includes: The acquisition module is used to acquire remote sensing images of the target coastal salt marsh area during key phenological periods; The classification module is used to process the remote sensing image using an unsupervised classification method to extract the initial vegetation outline of the vegetation cover area within the target region. The module is used to construct a phenological decision tree model based on the NDVI threshold method. The model incorporates a machine learning-enhanced node partitioning mechanism to optimize the node hierarchy and judgment priority of the decision tree. The phenological slope analysis method is used to strengthen the ability of the phenological decision tree model to identify the phenological differences of various types of vegetation during the withering period. The decision module is used to input the initial vegetation outline into the phenological decision tree model, and the phenological decision tree model performs multi-level classification of vegetation types based on the initial vegetation outline to distinguish the vegetation coverage range of Suaeda salsa, Reed, Spartina alterniflora and Ligusticum striatum in the target coastal salt marsh area. The estimation module is used to obtain the vegetation carbon density parameters corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum. Combined with the vegetation cover range corresponding to Suaeda salsa, Phragmites australis, Spartina alterniflora, and Ligusticum striatum, the total carbon content of the target coastal salt marsh area is estimated by spatial overlay calculation.

Images