Multi-source remote sensing and artificial intelligence fused wetland carbon reserve estimation method

Abstract

The invention relates to a multi-source remote sensing and artificial intelligence fused wetland system plant carbon reserve fine estimation method, which belongs to the technical field of wetland ecological monitoring and remote sensing inversion, and comprises the following steps: obtaining and preprocessing multi-source remote sensing image data; a two-stage progressive classification system is constructed in combination with land utilization types and vegetation community spatial distribution data acquired in the field, and vegetation fine class identification is realized based on a random forest and a graph convolutional neural network; the method comprises the following steps: fusing spectrum and texture features of multi-source remote sensing, combining actual measurement data of a biomass quadrat, and constructing a biomass regression estimation model through feature variable optimization; and based on the biomass estimation result and the carbon conversion coefficient corresponding to the vegetation subclass, calculating the carbon reserve and drawing a carbon density spatial distribution diagram. According to the method, the problems of low wetland vegetation classification precision, insufficient multi-source data collaborative utilization and poor biomass estimation model stability are effectively solved, and reliable technical support is provided for wetland ecosystem protection and dual-carbon strategy implementation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wetland ecological monitoring, in particular to a wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence. BACKGROUND

[0002] As a global important carbon sink, the accurate estimation of wetland carbon storage is a key technology for addressing climate change and ecological protection. The current mainstream technology mainly faces the following bottlenecks: (1) Existing technical bottlenecks and core challenges: single remote sensing classification system cannot distinguish the fine types of wetland vegetation (such as short herbaceous and tall herbaceous), resulting in serious biomass inversion error and transmission to the carbon storage estimation link; optical remote sensing is easily disturbed by clouds, while radar remote sensing data lacks spectral information, and existing research shows that simple superposition of the two has limited precision improvement; traditional estimation models or single machine learning algorithms have unstable regression prediction performance, making it difficult to control the accuracy of biomass estimation results.

[0003] (2) Technical development history and evolution trend: based on traditional spatial / spectral resolution remote sensing images, the carbon storage distribution map estimated by relying on statistical models has poor spatial continuity; the introduction of traditional machine learning and artificial feature engineering design has insufficient utilization of geospatial data; deep learning has been applied to wetland classification, but there is still no standardized technical framework for modeling the spatial relationship of vegetation. SUMMARY

[0004] To solve the problems of insufficient classification accuracy, limited data fusion effect and poor model stability in existing technologies, and to provide technical support for wetland protection and "double carbon" goals, the present application provides a wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence, which includes: (1) a two-stage progressive classification system: through the cascade architecture of supervised learning and deep learning, the vegetation classification accuracy is improved; (2) the "morphology-spectrum" features of Sentinel series satellite SAR and MSI data are fused to break through the data quality bottleneck in cloudy areas; (3) combining the advantages of graph convolutional neural network in spatial relationship modeling and the regression stability of ensemble learning algorithm, a carbon storage estimation model with higher accuracy and stronger robustness is established.

[0005] The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence provided by the present application has the following steps: S1, multi-source remote sensing image preprocessing and field data collection: obtain Sentinel-1 dual-polarization SAR data, Sentinel-2 multispectral data and high-resolution image data of Gaofen-2, and perform radiation correction, atmospheric correction, terrain correction and resampling preprocessing; at the same time, field investigation is carried out, and land use type spatial distribution data, stratified vegetation type verification data, biomass measurement data and carbon conversion coefficients of various vegetation are collected; S2, fine classification of land use and vegetation: extract spectral features, vegetation indices, texture features and SAR polarization features of multi-source data, use a two-level progressive classification framework, the first level uses Random Forest (RF) algorithm to divide the land into four categories of vegetation, water, buildings and unused land; the second level aims at the vegetation area, combining spatial adjacency relationship and multi-dimensional features, using Graph Convolution Neural Networks (GCN) to realize the fine classification of low herb, tall herb, seasonal beach plant and perennial tree; S3, biomass regression model construction and raster estimation: based on Pearson-Spearman joint correlation analysis and recursive feature elimination (RFE) algorithm to screen the feature variables significantly related to biomass; use multiple regression algorithms in parallel training, and through Voting and Stacking for integrated modeling, build and optimize high-precision biomass estimation model; multiple regression algorithms include linear model Partial Least Squares, Ridge, Lasso, Elastic Net, tree model Random Forest, Light GBM, XGBoost, CatBoost and other models Support Vector Machine, AdaBoost; on this basis, the spatialization of plant biomass in the whole study area is realized by using raster calculation; S4, carbon storage calculation and spatialization: according to the carbon conversion coefficient of different vegetation types, convert the biomass raster data into carbon storage raster data, summarize the carbon storage, realize the quantitative estimation and spatial expression of wetland system carbon storage; S5, carbon density mapping and management strategy generation: generate a 10m x 10m resolution carbon density spatial distribution map in the GIS platform, and propose wetland zoning differentiated protection and carbon sink enhancement strategies according to the carbon density pattern.

[0006] Preferably, the preprocessing of Sentinel-1 SAR data in S1 includes orbit correction, radiation correction, coherent spot noise elimination and terrain distortion correction; the preprocessing of Sentinel-2 MSI data includes atmospheric correction, band fusion and image cropping; the preprocessing of high-resolution two optical images includes radiation calibration, atmospheric correction and orthorectification.

[0007] Orbit correction is used to calibrate satellite orbital position information to ensure that the imaging data corresponds accurately to the actual position; radiometric correction is used to convert pixel values ​​into backscattering coefficients (decibels); speckle denoising is performed by using a Refined Lee filter to reduce noise; terrain distortion correction is performed by combining DEM data to correct geometric distortion; atmospheric correction is performed by using the Sen2Cor tool to generate an image of the atmospheric bottom reflectance; multispectral and panchromatic band fusion is performed and resampled to a 1m×1m resolution; image cropping is performed by stitching adjacent images together using ENVI software and cropping according to the boundaries of the study area.

[0008] Preferably, the field data collection in S1 includes: Land use type data collection: The spatial location and extent of vegetation, bare land, water bodies, roads and residential areas are obtained in time series, and spatial distribution data of different types of vegetation are accurately collected; Biomass measurement data collection: 1m×1m standard quadrats were set up, and vegetation parameters were collected in layers; for herbaceous plants, the whole-plant harvesting method was used to determine the dry weight; for woody plants, the diameter at breast height and tree height were measured, and after obtaining the relevant parameters, the Schumacher-Hall equation was used to estimate the dry weight; the obtained dry weight data were converted into quadrats with the same spatial resolution as the predicted data. Carbon conversion coefficient collection: Roots, stems and leaves of plants were collected by category, processed and then the carbon conversion coefficients were measured in the laboratory and the average value was calculated.

[0009] Preferably, in S2, the spectral characteristics include the reflectance of Sentinel-2 in the blue band, green band, red band, red edge band, near-infrared band, and short-wave infrared band. There are 11 vegetation indices, including the Normalized Difference Vegetation Index, Ratio Vegetation Index, Enhanced Vegetation Index, Difference Vegetation Index, Soil-Regulated Vegetation Index, Red-edged Normalized Difference Vegetation Index, Chlorophyll Monitoring Index, Chlorophyll Monitoring Index, Brightness Index, Redness Index, and Color Rendering Index. The texture features are extracted using the Gray-Level Co-occurrence Matrix (GLCM) method. The extracted feature parameters include mean, variance, uniformity, contrast, dissimilarity, entropy, second moment of angle, and correlation. The SAR features include the VV / VH backscattering coefficients and derived textures of Sentinel-1.

[0010] Preferably, in S2, the random forest algorithm determines the final classification result by constructing multiple decision trees and using a voting mechanism. The mathematical expression is: , In the formula, This represents the probability of belonging to a certain category. k For the number of decision trees, Indicates the first i A decision tree.

[0011] Preferably, in S2, the graph convolutional neural network adopts a convolutional neural network structure, including an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, and the precision evaluation adopts a confusion matrix to calculate overall precision, user precision, mapping precision and a Kappa coefficient.

[0012] Preferably, in S3, the calculation formula of the biomass estimation model is: , In the formula, represents the total biomass, represents the predicted grid biomass, represents the number of grids, represents the number of plant categories.

[0013] After integration, the preferred biomass estimation model (Stacking Regressor) has R² of 0.86, 0.91, 0.94 and 0.87 for dwarf herbs, tall herbs, seasonal beach plants and perennial trees, respectively.

[0014] Preferably, in S4, the calculation formula of the grid calculation model is: , In the formula, represents the biomass of a single plant, represents the carbon conversion coefficient of a single plant, represents the number of plant categories.

[0015] Preferably, in S5, the resolution of the drawn carbon density distribution map is 10m x 10m.

[0016] According to the specific embodiments provided by the present application, the following technical effects are disclosed: 1. Classification system innovation: a two-level progressive classification framework is adopted, the first layer adopts traditional supervised learning to implement land use classification and extract vegetation coverage areas; the second layer introduces deep learning to establish a fine recognition model to distinguish different types of plants (dwarf herbs / tall herbs / seasonal beach plants / perennial trees). By constructing a plant-specific classification model, the interference of the "same spectrum of different objects" phenomenon is effectively eliminated, aiming to realize the differential modeling of the biomass of different vegetation types.

[0017] 2. Multi-source data synergy: Integrate Sentinel-1 SAR and Sentinel-2 MSI multi-source remote sensing data to form a collaborative information extraction system: ① Use Sentinel-1 dual polarization (VV+VH) radar data to obtain vegetation canopy structure parameters all-weather; ② Use Sentinel-2 multispectral data (10 / 20 / 60m multi-scale resolution) to obtain rich vegetation spectral characteristics. Through the "morphological-spectral" feature fusion technology, a multi-dimensional representation model of vegetation growth status is established.

[0018] 3. Feature engineering optimization: Construct a multi-dimensional feature space containing spectral reflectance (visible / near-infrared, etc.), vegetation index (NDVI / EVI / NDWI, etc.), and texture features (GLCM / Gabor, etc.). Use Pearson-Spearman joint correlation analysis combined with recursive feature elimination algorithm to realize the optimal subset selection of feature variables, ensure the effectiveness of the input feature set of biological-physical representation, and reduce the influence of feature dimension disaster.

[0019] 4. Intelligent algorithm fusion: Establish a hybrid machine learning architecture: ① Random forest classification algorithm processes structured feature data to realize land use classification; ② Graph convolutional neural network inputs include adjacency matrix (based on spatial autocorrelation) and node features (spectrum+texture), to build a fine classification model of vegetation; ③ Use multiple regression algorithms such as linear (Partial Least Squares, Ridge, Lasso, Elastic Net), tree (Random Forest, Light GBM, XGBoost, CatBoost) and other (Support Vector Machine, AdaBoost) to respectively establish the nonlinear mapping relationship between "biomass-feature variables", and carry out single model parallel prediction; ④ Place the single regression model in the base model pool, and based on Voting and Stacking, carry out multi-model integration and optimization to improve the regression prediction performance and ensure the spatio-temporal stability of biomass estimation.

[0020] 5. Stereoscopic plot investigation: Establish a three-level ground measurement system. ① Land use plot: Conduct high-precision land use measurement, use high spatial resolution orthographic image to assist in delineating land class boundaries, and accurately map land use types and their spatial positions. ② Vegetation type verification sample: For different types of plant land, collect vegetation community spatial distribution atlas, and obtain the spatial position and spatial form of typical vegetation. ③ Biomass measurement sample: Based on the stratified sampling method, for herbaceous plants, use the whole plant harvesting method to measure the dry weight of 1m2 sample; for trees, use the Diameter-Based Height (DBH) and specific tree species correlation parameters to estimate the dry weight of 1m2 sample using the Schumacher-Hall equation.

[0021] 6. Carbon storage calculation system: Build a grid calculation model. Refer to the characteristics of the characteristic variable data, establish a 10m×10m standard grid calculation system, and unify the scale consistency of sample data and prediction data; stratified sampling according to vegetation type (≥10 samples per type), use an elemental analyzer to measure organic carbon content (i.e. carbon conversion coefficient; based on the GIS grid calculation system, use the formula to aggregate spatially, estimate the carbon storage of the wetland system, and output the carbon density distribution map. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 FIG. 1 is a flowchart of a wetland carbon storage estimation method according to an embodiment of the present application. Figure 2 FIG. 2 is a carbon density distribution map of four types of plants in the South Dongting Lake wetland according to an embodiment of the present application, wherein (a) represents the carbon density distribution of short herbaceous plants, (b) represents the carbon density distribution of tall herbaceous plants, (c) represents the carbon density distribution of seasonal beach plants, and (d) represents the carbon density distribution of perennial trees. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.​

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Example like Figure 1 As shown, a wetland carbon storage estimation method integrating multi-source remote sensing and artificial intelligence comprises the following steps: S1. Multi-source remote sensing image preprocessing and field data acquisition.

[0027] S1.1 Acquire Sentinel-1 SAR data, Sentinel-2 MSI data, and Gaofen-2 high-resolution image data to provide a data foundation for subsequent land use classification and plant type identification.

[0028] S1.2 Data Preprocessing. Preprocessing of Sentinel-1 SAR data includes orbit correction, radiometric correction, speckle noise reduction, and terrain distortion correction; preprocessing of Sentinel-2 MSI data includes atmospheric correction, band fusion, and image cropping; preprocessing of high-resolution optical images includes radiometric calibration, atmospheric correction, and orthorectification.

[0029] The specific steps for S1.3 field data collection include: Field sampling was conducted in standard quadrats (1m × 1m) in the South Dongting Lake wetland research area, completed in December 2022, April 2023, and June-July 2023. A systematic survey was carried out for different plant types, and the results are as follows: For Category A plants (low-growing herbaceous plants such as Artemisia capillaris and Leonurus japonicus), plant cover, average height, and total vegetation cover were recorded within the quadrats; for Category B plants (tall herbaceous plants such as Reed), average height, density, and community structure were measured; for Category C plants (seasonal herbaceous plants such as Artemisia annua), cover, average height, and total vegetation cover were recorded; and for Category D plants (trees such as Populus tomentosa and Willow), individual tree height, diameter at breast height (DBH), community density, and stratification were measured. After field sampling, biomass was determined in the laboratory. Herbaceous plant samples were sterilized at 105℃ for 30 minutes and then dried at 65℃ for 48 hours to constant weight, with dry weight biomass measured using an electronic balance. Tree biomass was calculated using the allometric growth equation, as shown in the following formula: , In the formula: AGB represents the total biomass of the tree (including roots), D represents the diameter at breast height (DBH), and H represents the tree height. A total of 828 representative sampling points were set up in the study, including 240 in category A, 250 in category B, 138 in category C, and 200 in category D. The locations of the sampling points were determined through the Ovi interactive map and field investigation.

[0030] S2. Feature Extraction and Land Use Classification S2.1 Extract spectral features, vegetation index, texture features and SAR features from multi-source data.

[0031] Spectral features were extracted from Sentinel-2 multispectral data, including the blue band (B2, 496.9 nm), green band (B3, 560.0 nm), red band (B4, 664.5 nm), red-edge band (B5-B7, 703.9-782.5 nm), near-infrared band (B8, 835.1 nm), and short-wave infrared band (B11-B12, 1613.7-2202.4 nm). These bands were mainly used for vegetation classification and biomass estimation.

[0032] Regarding vegetation indices, the NDVI (Normalized Difference Vegetation Index) is calculated using the formula NDVI=(B8-B4) / (B8+B4), with a value range of -1 to 1. For vegetation areas, the value is typically 0.2-0.8, used to reflect vegetation growth status and coverage. Simultaneously, 11 vegetation indices are calculated, including RVI (Ratio Vegetation Index), EVI (Enhanced Vegetation Index), DVI (Differential Vegetation Index), SAVI (Soil-Adjusted Vegetation Index), NDVIre (Red Edge Normalized Difference Vegetation Index), MTCI (Chlorophyll Monitoring Index), Brightness Index (BI), Redness Index (RI), and Color Rendering Index (CI).

[0033] Texture features were extracted using the Gray-Level Co-occurrence Matrix (GLCM) method, including eight feature parameters: mean, variance, homogeneity, contrast, dissimilarity, entropy, angular second moment, and correlation. A 3×3 pixel window was used for extraction, with an X / Y offset step size of 1, and 64 levels of grayscale compression were applied. Data sources included Sentinel-1 VV / VH polarization data and Sentinel-2 NDVI calculation results.

[0034] SAR features include the VV / VH backscattering coefficients and derived textures of Sentinel-1.

[0035] S2.2 adopts a two-level progressive classification architecture for land use classification. The first-level classification architecture uses the random forest algorithm to classify the land into four categories: vegetation, water bodies, buildings, and unused land. The second-level classification architecture uses a graph convolutional neural network to classify the vegetation coverage area into four categories: Category A: dwarf herbaceous plants (mainly Artemisia argyi and Leonurus japonicus), Category B: tall herbaceous plants (mainly Reed), Category C: seasonal tidal flat plants (mainly Artemisia annua), and Category D: perennial trees (mainly Populus tomentosa and Willow).

[0036] The first-level classification architecture, Random Forest (RF), is based on the ensemble learning framework proposed by Breiman. It determines the final classification result by constructing multiple decision trees and using a voting mechanism. The mathematical expression is: , In the formula, This represents the probability of belonging to a certain category. k For the number of decision trees, Indicates the first i A decision tree, where 1 represents being voted for and 0 represents not being voted for.

[0037] The second-level classification architecture uses a graph convolutional neural network (CNN) structure, including an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. Accuracy is evaluated using a confusion matrix to calculate overall accuracy, user accuracy, producer accuracy, and the Kappa coefficient.

[0038] In practical applications, Random Forest achieved an overall accuracy of 90.15% (Kappa=0.90) in primary classification, but lagged behind Graph Convolutional Neural Networks (90.55% overall accuracy, Kappa=0.87) in secondary plant classification. Specifically, the accuracy of building mapping was low in primary classification (73.26%), mainly due to imbalanced samples; the accuracy of user classes B and D was high in secondary classification (86.08% and 95.17%, respectively), while class C had the lowest accuracy due to seasonality. All classification results were validated using a 7:3 training / validation set split, with a total of 32,164 samples.

[0039] S3. Construction and estimation of biomass regression model S3.1 Feature Variable Screening: The Pearson correlation coefficient is used to screen feature variables. The calculation formula is as follows: , In the above formula, X and Y represent variables, Represents the i-th variable X. Represents the i-th variable, , The value of r represents the mean of the variables. The absolute value of r indicates the degree of correlation, ranging from -1 to 1. The larger the absolute value of r, the stronger the correlation between the independent and dependent variables. A positive value of r indicates a positive correlation between the independent and dependent variables, while a negative value of r indicates a negative correlation.

[0040] The analysis results showed that for Class A plants, NDVI was significantly positively correlated with biomass (r=0.355, p<0.01), and the B4 band was significantly negatively correlated (r=-0.193, p<0.01), with the MTCI index showing the highest correlation (r=0.379, p<0.01). For Class B plants, the B4 band was significantly positively correlated (r=0.287, p<0.01), and NDVI was significantly negatively correlated (r=-0.313, p<0.01), with texture feature variance showing the strongest correlation (r=0.722, p<0.01). For Class C plants, the B2 band showed a significant positive correlation (r=0.156, p<0.05), the NDVI showed a significant negative correlation (r=-0.116, p<0.05), and the VV polarization backscattering coefficient showed a significant negative correlation (r=-0.142, p<0.05). For Class D plants, the NDVI showed a significant positive correlation (r=0.197, p<0.01), the VV backscattering coefficient showed the strongest correlation (r=0.547, p<0.01), and the texture feature entropy showed a significant positive correlation (r=0.347, p<0.01). The screening criteria included significance level (p<0.05 or p<0.01), ranking by absolute value of correlation coefficient, and low autocorrelation between variables. The final candidate variables included: 11 in category A (B4, SAVI, NDVI, etc.), 10 in category B (B4, NDVI, VH, etc.), 10 in category C (B2, B4, VV, etc.), and 8 in category D (B2, B3, VV, etc.). All analyses were based on 80% of the training sample (n=662) and used a two-tailed test.

[0041] S3.2. Multiple regression algorithms were trained in parallel and integrated using Voting and Stacking to optimize the final prediction model. This embodiment simultaneously trained ten regression models (Partial Least Squares, Ridge, Lasso, Elastic Net, Random Forest, Light GBM, XGBoost, CatBoost, Support Vector Machine, and AdaBoost), then placed them in a model pool and integrated them using Voting and Stacking respectively. Comparative training showed that the Stacking ensemble learning model using XGBoost as the meta-model had the best regression prediction performance. Model training used 80% of the sample data (n=662), with input variables being features selected through Pearson correlation analysis, including B4 band, NDVI index, and SAR texture features. The coefficients of determination (R²) for the regression prediction results of the optimized model were: Class A 0.86 (RMSE=13.87), Class B 0.91 (RMSE=78.28), Class C 0.94 (RMSE=18.60), and Class D 0.87 (RMSE=35.21). Model evaluation employed four metrics: coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). Experimental results show that the Stacking ensemble learning model exhibits the best prediction accuracy and stability in biomass estimation for all four plant types. The model's superior performance is attributed to the effectiveness of its ensemble learning framework: firstly, Stacking generates meta-features through cross-validation, reducing the risk of overfitting; secondly, XGBoost, as a meta-model, further enhances the model's generalization ability through its built-in regularization term and gradient boosting framework.

[0042] S3.3 Construct a biomass estimation model and use grid calculation to obtain the biomass of all plants in the study area.

[0043] This embodiment uses XGBoost as the meta-model in a Stacking ensemble learning model for biomass estimation. Specifically, for a 10m × 10m Sentinel-2 pixel (area 100m²), the model prediction is directly multiplied by the pixel area to obtain the biomass of that pixel. Taking Reed B as an example, the root mean square error (RMSE) predicted by the Stacking ensemble model is 78.28 kg / 100m², which is calculated using the formula: , The calculation shows that, among which These are measured values. The values ​​are predicted, and n is the number of samples. The prediction accuracies for other categories are as follows: Category A herbaceous plants (RMSE = 13.87 kg / 100 m²), Category C seasonal plants (RMSE = 18.60 kg / 100 m²), and Category D trees (RMSE = 35.21 kg / 100 m²). All biomass estimates are based on selected characteristic variables, including Sentinel-2's B4 band, NDVI vegetation index, and Sentinel-1's VV / VH backscattering coefficients, and the model's reliability is verified by the coefficient of determination (R² = 0.86–0.94) and mean absolute percentage error (MAPE = 0.003–0.023). The final pixel-level biomass calculation results will be used for carbon storage estimation.

[0044] S4. Carbon storage calculation and spatialization.

[0045] S4.1 Determination of Carbon Conversion Coefficients: Root, stem, and leaf samples were collected from representative plants of four wetland plant types. After drying, grinding, and mixing, the organic carbon content was determined using an elemental analyzer. The average value of the results from multiple samples of the same plant type was calculated to obtain the carbon conversion coefficient for each plant category: Category A (dwarf herbaceous plants) 42.99%, Category B (tall herbaceous plants) 43.59%, Category C (seasonal tidal flat plants) 38.58%, and Category D (trees) 45.12%. This determination method is scientifically reliable, and the obtained coefficients accurately reflect the actual carbon sequestration characteristics of different wetland plant types, providing a solid data foundation for subsequent carbon storage calculations.

[0046] S4.2 Carbon Storage Calculation: The carbon storage of wetland plants is estimated using the carbon storage calculation formula. The calculation formula for the raster computing model is: , In the formula, Indicates the biomass of a single plant species. Indicates the carbon transformation coefficient of a single plant species. This indicates the number of major plant categories.

[0047] The calculation results show that the total carbon storage of plants in the South Dongting Lake wetland is 7,079,301.69 tons, specifically distributed as follows: Category B reeds have the highest carbon storage (3,940,368.74 tons, accounting for 55.66%), followed by Category D trees (3,063,323.10 tons), while Category A dwarf herbs (73,873.68 tons) and Category C seasonal plants (1,736.17 tons) have relatively small proportions. This result is consistent with the plant distribution characteristics observed in the field survey. Category B reeds, due to their widest distribution area (45,493.49 hm², accounting for 71.65% of the total plant area) and stable growth environment, became the main carbon sink contributor. Category D trees, although with a smaller distribution area (5,174.60 hm²), had a higher carbon storage per plant (average carbon density 5,919.92 kg / 100 m²), resulting in a larger total carbon storage. Category C plants, growing only in dry season mudflats and with a limited distribution range (3,376.39 hm²), had the lowest carbon storage share (0.024%). All calculations were based on random forest classification (overall accuracy 90.55%) and Stacking biomass estimation (R² = 0.86–0.94), ensuring the scientific validity and reliability of the carbon storage estimation.

[0048] S5. Carbon density spatial distribution analysis Based on biomass data and carbon conversion coefficients predicted by the XGBoost model, carbon density distribution maps of four plant species in the South Dongting Lake wetland were generated, such as... Figure 2As shown, (a) represents the carbon density distribution of dwarf herbaceous plants, (b) represents the carbon density distribution of tall herbaceous plants, (c) represents the carbon density distribution of seasonal tidal flat plants, and (d) represents the carbon density distribution of perennial trees. A detailed analysis of their spatial characteristics revealed the following results: Category D trees (poplar and willow) had the highest carbon density (average 5,919.92 kg / 0.01 hm²), but were scattered and mainly concentrated in the central wetland area; Category B reeds were the most widely distributed (accounting for 71.65% of the total plant area) and had a relatively uniform carbon density (average 866.13 kg / 0.01 hm²), making them the main contributor to the carbon storage in the study area (3,940,368.74 tons, accounting for 55.66% of the total); Category A dwarf herbs (Artemisia and Leonurus) had a moderate carbon density (average 78.2 kg / 0.01 hm²) and were mainly distributed in the northeastern region; Category C seasonal plants (Artemisia annua) had the lowest carbon density (average 5.14 kg / 0.01 hm²) and were evenly distributed in the lake area mudflats. This spatial distribution pattern is closely related to the hydrological characteristics of the study area: Category B reeds are widely distributed due to their adaptation to a stable wetland environment; Category D trees require relatively dry habitats and are therefore distributed in patches; Category C plants are affected by seasonal water level changes and only grow on exposed mudflats during the dry season. Carbon density was calculated using the formula: Carbon density = Biomass × Carbon conversion coefficient. All spatial analyses were based on random forest classification results (Kappa = 0.87) and 10m resolution pixels from Sentinel-2 imagery.

[0049] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art to meet different specific practical needs according to actual circumstances. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims. In the above description, numerous specific details have been set forth to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to implement the present invention. In other instances, to avoid obscuring the present invention, well-known techniques, such as specific construction details, operating conditions, and other technical conditions, have not been specifically described.

[0050] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for estimating wetland carbon storage by integrating multi-source remote sensing and artificial intelligence, characterized in that, The steps are as follows: S1. Multi-source remote sensing image preprocessing and field data acquisition: Acquire Sentinel-1 dual-polarization SAR data, Sentinel-2 multispectral data, and Gaofen-2 high-resolution image data, and perform radiometric correction, atmospheric correction, topographic correction, and resampling preprocessing respectively; at the same time, conduct field surveys to collect spatial distribution data of land use types, verification data of layered vegetation types, measured biomass data, and carbon conversion coefficients of various vegetation types. S2. Fine Classification of Land Use and Vegetation: Spectral features, vegetation indices, texture features, and SAR polarization features of multi-source data are extracted. A two-level progressive classification architecture is adopted. The first level uses a random forest algorithm to divide land into four categories: vegetation, water bodies, buildings, and unused land. The second level focuses on vegetation areas and combines spatial adjacency relationships and multi-dimensional features to achieve fine classification of dwarf herbaceous plants, tall herbaceous plants, seasonal tidal flat plants, and perennial trees using graph convolutional neural networks. S3. Biomass regression model construction and raster estimation: Screening of feature variables that are significantly related to biomass based on Pearson-Spearman joint correlation analysis and recursive feature elimination algorithm; Multiple regression algorithms were trained in parallel, and integrated modeling was performed using Voting and Stacking to construct and select a high-precision biomass estimation model. The regression algorithms included linear models such as Partial Least Squares, Ridge, Lasso, and Elastic Net, tree models such as Random Forest, Light GBM, XGBoost, and CatBoost, and other models such as Support Vector Machine and AdaBoost. Based on this, grid computing was used to achieve spatial estimation of plant biomass in the entire study area. S4. Carbon storage calculation and spatialization: Based on the carbon conversion coefficient of different vegetation types, biomass raster data is converted into carbon storage raster data, carbon storage is summarized, and quantitative estimation and spatialized expression of carbon storage in wetland systems are realized. S5. Carbon density mapping and management strategy generation: Generate a 10m×10m resolution spatial distribution map of carbon density in the GIS platform, and propose wetland zoning differentiated protection and carbon sink enhancement strategies based on the carbon density pattern.

2. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 1, characterized in that: Preprocessing of Sentinel-1 SAR data in S1 includes orbit correction, radiometric correction, speckle noise reduction, and terrain distortion correction; preprocessing of Sentinel-2 MSI data includes atmospheric correction, band fusion, and image cropping; and preprocessing of Gaofen-2 images includes radiometric calibration, atmospheric correction, and orthorectification.

3. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 2, characterized in that: Field data collection in S1 includes: Land use type data collection: The spatial location and extent of vegetation, bare land, water bodies, roads and residential areas are obtained in time series, and spatial distribution data of different types of vegetation are accurately collected; Biomass data collection: 1m×1m standard quadrats were set up and vegetation parameters were collected in layers; for herbaceous plants, the whole-plant harvesting method was used to determine dry weight; for woody plants, diameter at breast height and tree height were measured to obtain relevant parameters, and the Schumacher-Hall equation was applied to estimate dry weight; the obtained dry weight data were converted into quadrats with the same spatial resolution as the predicted data. Carbon conversion coefficient collection: Roots, stems and leaves of plants were collected by category, processed and then the carbon conversion coefficients were measured in the laboratory and the average value was calculated.

4. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 3, characterized in that: In S2, the spectral characteristics include the reflectance of Sentinel-2 in the blue band, green band, red band, red edge band, near-infrared band, and short-wave infrared band. There are 11 vegetation indices, including: Normalized Difference Vegetation Index, Ratio Vegetation Index, Enhanced Vegetation Index, Difference Vegetation Index, Soil-Regulated Vegetation Index, Red-edged Normalized Difference Vegetation Index, Chlorophyll Monitoring Index, Chlorophyll Monitoring Index, Brightness Index, Redness Index, and Color Rendering Index. The texture features are extracted using the gray-level co-occurrence matrix method. The extracted feature parameters include mean, variance, uniformity, contrast, dissimilarity, entropy, second moment of angle, and correlation. The SAR features include the VV / VH backscattering coefficients and derived textures of Sentinel-1.

5. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 4, characterized in that: In S2, the random forest algorithm determines the final classification result by constructing multiple decision trees and using a voting mechanism. The mathematical expression is: , In the formula, This represents the probability of belonging to a certain category. k For the number of decision trees, Indicates the first i A decision tree.

6. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 5, characterized in that: In S2, the graph convolutional neural network adopts a convolutional neural network structure, including an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The accuracy evaluation uses a confusion matrix to calculate the overall accuracy, user accuracy, graph accuracy, and Kappa coefficient.

7. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 6, characterized in that: In S3, the calculation formula for the biomass estimation model is: , In the formula, Indicates the total biomass. This represents the predicted grid biomass. Indicates the number of grid cells. This indicates the number of major plant categories.

8. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 7, characterized in that: In S4, the calculation formula for the raster calculation model is: , In the formula, Indicates the biomass of a single plant species. Indicates the carbon transformation coefficient of a single plant species. This indicates the number of major plant categories.

9. The wetland carbon storage estimation method based on multi-source remote sensing and artificial intelligence fusion as described in claim 8, characterized in that: In S5, the carbon density distribution map is plotted with a resolution of 10m×10m.

Images