Method for monitoring and predicting dynamic of carbon storage of ecosystem based on multi-source remote sensing data

Abstract

This invention discloses a method for dynamic monitoring and prediction of ecosystem carbon storage based on multi-source remote sensing data. The method includes: acquiring and preprocessing optical remote sensing, radar, LiDAR, UAV multispectral, and ground-based measured data; constructing a multi-scale spatiotemporal fusion network; utilizing a cross-modal attention mechanism and an improved long short-term memory network to achieve multi-source feature fusion; employing an adaptive feature selection network to screen core feature factors; constructing a deep forest model to collaboratively invert vegetation and soil carbon storage; constructing a spatiotemporal graph convolutional network coupling land use and climate scenarios to achieve multi-scenario carbon storage prediction for the next 1-5 years; and establishing an online incremental learning mechanism based on elastic weight consolidation to achieve dynamic model updates. This invention solves the problems of spatiotemporal resolution mismatch, insufficient fusion depth, and insufficient carbon pool coordination in multi-source remote sensing data, and can be used for ecosystem carbon sink monitoring.

Description

Technical Field

[0001] This invention relates to the fields of ecological environment monitoring, remote sensing information processing and carbon cycle simulation technology, and in particular to a method for dynamic monitoring and prediction of ecosystem carbon storage based on multi-source remote sensing data. Background Technology

[0002] Accurate estimation and dynamic prediction of carbon storage in terrestrial ecosystems are key supports for achieving the "dual carbon" goals and addressing climate change.

[0003] The existing technology has the following main shortcomings:

[0004] 1. The fusion of multi-source remote sensing data is shallow, and data from optical, radar, LiDAR and other sources have not been deeply fused, resulting in a prominent problem of mismatch between spatiotemporal resolution.

[0005] 2. The carbon storage prediction function is weak and cannot support future multi-scenario trend assessment;

[0006] 3. The model lacks an online update mechanism, resulting in poor monitoring timeliness;

[0007] 4. Insufficient synergistic inversion between vegetation carbon pool and soil carbon pool leads to systematic bias in the estimation of total carbon storage.

[0008] Therefore, there is a need for a dynamic monitoring and prediction method for ecosystem carbon storage based on multi-source remote sensing data that can deeply integrate multi-source data, achieve integrated monitoring and prediction, and has self-updating capabilities. Summary of the Invention

[0009] The present invention aims to at least partially solve one of the technical problems in the above-mentioned technologies.

[0010] To achieve the above objectives, the first aspect of this invention proposes a method for dynamic monitoring and prediction of ecosystem carbon storage based on multi-source remote sensing data, comprising the following steps:

[0011] S1: Acquire multi-source remote sensing data of the study area and perform spatiotemporal registration preprocessing; the multi-source remote sensing data includes optical remote sensing image data, radar remote sensing data, LiDAR point cloud data, UAV multispectral data and ground measurement data, and the spatiotemporal registration preprocessing includes generating high spatiotemporal resolution optical remote sensing data using a spatiotemporal adaptive reflectivity fusion model.

[0012] S2: Construct a multi-scale spatiotemporal fusion network to perform spatiotemporal resolution alignment and feature-level fusion on the preprocessed multi-source remote sensing data to generate a multi-source remote sensing fusion feature set. The multi-scale spatiotemporal fusion network includes a spatial attention module and a temporal feature extraction module. The spatial attention module uses a cross-modal attention mechanism to achieve spatial feature alignment between optical images and radar data. The temporal feature extraction module uses an improved long short-term memory network to extract the temporal evolution features of multi-temporal remote sensing data.

[0013] S3: Based on the multi-source remote sensing fusion feature set generated by S2, an adaptive feature selection network is used to screen the core feature factors related to ecosystem carbon storage, and a carbon storage estimation model based on the deep forest algorithm is constructed to output the carbon storage estimation value of each grid point in the study area. The carbon storage estimation value includes vegetation carbon storage and soil carbon storage, wherein the soil carbon storage is retrieved by Bayesian neural network collaborative inversion.

[0014] S4: Using the carbon storage estimate output by S3 as the baseline data, a carbon storage prediction model coupled with the land use dynamics module and the climate scenario module is constructed to predict the ecosystem carbon storage at future time points. The carbon storage prediction model adopts a spatiotemporal graph convolutional neural network, and its graph convolution operation adopts Chebyshev polynomial approximation.

[0015] S5: Establish an online incremental learning mechanism to dynamically update the parameters of the carbon storage estimation model and the prediction model when new multi-source remote sensing data are acquired, so as to realize continuous dynamic monitoring of carbon storage. The online incremental learning mechanism adopts the elastic weight consolidation regularization method.

[0016] In addition, the method for dynamic monitoring and prediction of ecosystem carbon storage based on multi-source remote sensing data proposed above according to the present invention may also have the following additional technical features:

[0017] Furthermore, the spatiotemporal adaptive reflectivity fusion model in S1 adopts the following formula: ;in, For the predicted time High-resolution reflectivity, For low-resolution reflectivity, For high-resolution reflectivity, The weighting coefficients are calculated based on time difference and spectral similarity, satisfying... , where n is the number of time phases involved in the fusion.

[0018] Furthermore, the cross-modal attention mechanism in S2 adopts the following formula: Where Q is the optical feature flattening matrix, and K and V are the radar feature flattening matrices. For feature dimensions.

[0019] Furthermore, the improved Long Short-Term Memory network in S2 employs the following formula for calculating temporal attention weights: ;in, Let v be the LSTM hidden state at time step t. , Here, T is a learnable parameter, and T is the timing length.

[0020] Furthermore, the adaptive feature selection network in S3 adopts the following objective function: ;in, For mutual information, For the selected feature subset, For candidate feature set, For carbon storage labeling; the deep forest algorithm adopts the following cascaded forest layer-by-layer update rule: ;in, For the first The input features of the layer For the first The class probability vector output by the layer; the variational lower bound loss function of the Bayesian neural network is: ;in, For training data, For network weights, For variational parameters, The divergence is Kullback-Leibler.

[0021] Furthermore, the normalized Laplacian matrix of the spatiotemporal graph convolutional neural network in S4 is defined as: ;in, It is an adjacency matrix. It is a diagonal matrix of degree. The identity matrix is ​​used; the graph convolution operation is approximated using Chebyshev polynomials. ;in, To scale the Laplacian matrix, For Chebyshev polynomials, For learnable parameters, Let be the order of the polynomial.

[0022] Furthermore, the loss function of the elastic weight consolidation regularization method in S5 is: ;in, For the loss function on the new data, These are the optimal parameters obtained from historical training. The diagonal elements of the Fisher information matrix are calculated using the following formula: Where λ is the balance coefficient, and its value ranges from 1 to 100.

[0023] A second aspect of this invention proposes a dynamic monitoring and prediction system for ecosystem carbon storage based on multi-source remote sensing data, comprising:

[0024] The system comprises the following modules: a multi-source remote sensing data acquisition module for acquiring optical remote sensing imagery, radar remote sensing data, LiDAR point cloud data, UAV multispectral data, and ground-based measured data for the study area; a data preprocessing module for spatiotemporal registration preprocessing of the acquired multi-source remote sensing data; a multi-scale spatiotemporal fusion module with a built-in multi-scale spatiotemporal fusion network for spatiotemporal resolution alignment and feature-level fusion of the preprocessed multi-source remote sensing data; a carbon storage estimation module with a built-in carbon storage estimation model based on a deep forest algorithm for outputting carbon storage estimates for each grid point in the study area; a carbon storage prediction module with a built-in carbon storage prediction model coupled with a land use dynamics module and a climate scenario module for scenario-based prediction of future carbon storage; a dynamic update module for establishing an online incremental learning mechanism to achieve real-time updates of model parameters; and a results output module for visually displaying the monitoring and prediction results of carbon storage.

[0025] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect above.

[0026] The fourth aspect of this invention proposes the application of the method described in the first aspect above in monitoring carbon storage in forest ecosystems, wetland ecosystems, or grassland ecosystems.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. Deep fusion of multi-source data: By constructing a multi-scale spatiotemporal fusion network and adopting a cross-modal attention mechanism and an improved long short-term memory network, feature-level fusion of optical, radar and LiDAR data is achieved, effectively solving the problem of spatiotemporal resolution mismatch and improving the accuracy of carbon storage estimation by more than 25%.

[0029] 2. Integrated monitoring and forecasting: By using a spatiotemporal graph convolutional neural network coupled with land use change and climate scenario data, multi-scenario carbon storage prediction for the next 1 to 5 years can be achieved, filling the gap in the prediction capabilities of existing methods;

[0030] 3. Vegetation-soil carbon pool co-inversion: By using deep forest algorithm and Bayesian neural network to co-estimate vegetation carbon storage and stratified soil carbon storage, the systematic bias of ignoring soil carbon pool is overcome.

[0031] 4. Online incremental self-updating: An online incremental learning mechanism based on elastic weight consolidation is established, which allows the model to be dynamically updated with new data, avoiding retraining and enabling continuous real-time monitoring of carbon storage, shortening the monitoring time to within 72 hours. Attached Figure Description

[0032] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0033] Figure 1 This is an overall flowchart of a method for dynamic monitoring and prediction of ecosystem carbon storage based on multi-source remote sensing data according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic block diagram of the structure of a multi-scale spatiotemporal fusion network for a dynamic monitoring and prediction method of ecosystem carbon storage based on multi-source remote sensing data according to an embodiment of the present invention.

[0035] Figure 3 This is a schematic block diagram of the deep forest algorithm for dynamic monitoring and prediction of ecosystem carbon storage based on multi-source remote sensing data according to an embodiment of the present invention.

[0036] Figure 4 This is a schematic block diagram of the spatiotemporal graph convolutional neural network structure of an ecosystem carbon storage dynamic monitoring and prediction method based on multi-source remote sensing data according to an embodiment of the present invention.

[0037] Figure 5 This is a flowchart illustrating the online incremental learning mechanism of an ecosystem carbon storage dynamic monitoring and prediction method based on multi-source remote sensing data according to an embodiment of the present invention:

[0038] As shown in the figure:

[0039] 1. Acquisition and preprocessing of multi-source remote sensing data; 2. Multi-scale spatiotemporal fusion; 3. Carbon storage estimation; 4. Carbon storage prediction; 5. Online incremental learning. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments.

[0041] Example 1

[0042] like Figures 1 to 5 As shown, this embodiment takes a forest ecosystem in a certain province of China as the study area (approximately 500,000 hectares, geographical range of 118°~122° east longitude and 30°~34° north latitude) to specifically implement and verify the method of the present invention.

[0043] I. S1: Acquisition and Preprocessing of Multi-Source Remote Sensing Data

[0044] 1.1 Acquisition of Multi-Source Remote Sensing Data

[0045] (1) Optical remote sensing data: Sentinel-2 L2A multispectral images (spatial resolution 10 meters, temporal resolution 5 days) of the study area from January 2020 to December 2024 were acquired and downloaded from the Copernicus Open Access Centre of the European Space Agency (https: / / scihub.copernicus.eu). Landsat 8 / 9 OLI / TIRS images (30-meter resolution, 16-day revisit period) were also acquired as supplementary data.

[0046] (2) Radar remote sensing data: acquire Sentinel-1 GRD-level synthetic aperture radar data (C-band, VV+VH dual polarization, spatial resolution 10 meters, temporal resolution 12 days), and ALOS-2 PALSAR-2 L-band radar data (25-meter resolution).

[0047] (3) LiDAR data: Airborne LiDAR point cloud data of the study area were obtained, with a point cloud density of ≥4 points / m², provided by the provincial surveying and mapping department (acquisition time: autumn 2022).

[0048] (4) UAV multispectral data: Fifteen typical sample areas (1km×1km each) were set up in the study area. Aerial photography was carried out using a DJI Matrice 300 RTK UAV equipped with a MicaSense RedEdge-MX multispectral camera. The flight altitude was 120 meters, and multispectral images of five bands, namely blue (475nm), green (560nm), red (668nm), red edge (717nm), and near-infrared (842nm), were acquired with a spatial resolution of about 0.05 meters.

[0049] (5) Ground measurement data: 120 fixed plots (40m×40m) were set up using a stratified random sampling method. Considering factors such as vegetation type (coniferous forest, broad-leaved forest, mixed coniferous and broad-leaved forest), altitude (<500m, 500~1000m, >1000m), slope (<15°, 15°~30°, >30°), tree species, diameter at breast height (DBH≥5cm), tree height, crown width and other parameters were recorded in each plot. The biomass of individual trees was calculated using the allometric growth equation (referring to the industry standard "Standing Tree Biomass Model and Carbon Measurement Parameters"). The vegetation carbon storage at the plot scale was obtained by summing the data. At the same time, 0~30cm topsoil samples (mixed samples) were collected at the center and four corners of each plot. The soil organic carbon content was determined by the potassium dichromate oxidation method.

[0050] (6) Auxiliary data: acquire ERA5 reanalysis climate data (monthly average temperature, monthly precipitation, solar radiation), SRTM digital elevation model (30-meter resolution), and 1:100,000 soil type map.

[0051] 1.2 Preprocessing Operations

[0052] Optical imagery: Atmospheric correction was performed using the Sen2Cor v2.8 plugin to eliminate atmospheric scattering and absorption. Topographic correction (C model) was performed using a 30-meter resolution DEM of the study area. Geometric fine correction was performed based on a 1:100,000 scale topographic map, with the correction error controlled within 0.5 pixels. Cloud and snow cover pixels were removed using a QA60 band mask.

[0053] Radar data: speckle noise filtering was performed using the Refined Lee filtering algorithm (7×7 pixels window size), terrain correction was performed using SRTM DEM (Gamma-MAP algorithm) to eliminate radiation distortion caused by terrain, and the backscattering coefficient was converted to decibels (dB).

[0054] LiDAR data: Point cloud filtering was performed using LAStools software to separate ground points from non-ground points. A 1m resolution DEM was generated from the ground points, and a DSM was generated from the first echo. Vegetation height (CHM = DSM - DEM) was calculated, and structural parameters such as tree height, canopy coverage, and canopy undulation index were extracted.

[0055] Spatiotemporal fusion: A spatiotemporal adaptive reflectivity fusion model is adopted. Using Landsat 8 (30 meters, 16 days) as the high spatial resolution benchmark and MODIS MOD09GA (500 meters, 1 day) as the high temporal resolution data, high spatiotemporal resolution optical data with a 30-meter resolution and a 16-day revisit period were generated through fusion. Weighting coefficients... Based on time difference and spectral similarity calculations, satisfying ∑ =1, the number of time phases participating in the fusion is n=3 (one time phase before and after and the current time phase).

[0056] UAV data registration: A 3D point cloud was generated using the Structure for Motion Restoration (SfM) algorithm. Thirty ground control points (GPS measured coordinates, accuracy 0.1 meters) were selected evenly distributed within the study area. The UAV point cloud coordinates were transformed to the WGS-84 / UTM projection coordinate system using the Iterative Closest Point (ICP) algorithm. The registration error was controlled within 0.1 meters.

[0057] Ground data validation: Using measured carbon storage data from 120 sample plots, the residuals between the preliminary remote sensing inversion values ​​(estimated using only optical data) and the measured values ​​at each pixel location were calculated. Outlier samples with residual absolute values ​​exceeding twice the standard deviation were removed (a total of 5 outlier sample plots were removed), and 115 valid sample plots were retained, of which 80 were used for model training and 35 were used for independent validation.

[0058] II. S2: Multi-scale Spatiotemporal Fusion 2

[0059] 2.1 Spatial Attention Module

[0060] Preprocessed Sentinel-2 optical images (10-meter resolution, bands: B2, B3, B4, B8, B11, B12, a total of 6 bands) and Sentinel-1 radar data (VV+VH dual polarization, 10-meter resolution) were input into two parallel residual convolutional neural networks (ResNet-18 skeleton, with the first convolutional layer modified to adapt to multispectral input). Each network contained 4 convolutional blocks, and each convolutional block contained two 3×3 convolutional layers (stride 1, padding 1). Each convolutional layer was followed by a batch normalization layer and a ReLU activation function, outputting a 64-dimensional depth feature map (spatial size is 1 / 4 of the original size, downsampled by convolution with a stride of 2).

[0061] Employ a cross-modal attention mechanism: Where Q is the optical feature flattening matrix, and K and V are the radar feature flattening matrices. =64 is the feature dimension, outputting fused features. Later reshaped into .

[0062] For LiDAR features (CHM, canopy coverage, etc.), a hierarchical feature projection method is adopted: the LiDAR features are downsampled to the same spatial size as the optical features, then mapped to 64 dimensions through a 1×1 convolution, and then combined with... By performing element-by-element addition and fusion, spatial fusion characteristics are obtained.

[0063] 2.2 Temporal Feature Extraction Module

[0064] The spatially fused features are then input into an improved Long Short-Term Memory (LSTM) network in chronological order (each phase lasts 16 days, totaling approximately 114 phases over 5 years). This LSTM network contains two hidden layers, each with 128 hidden units. A temporal attention layer is added to the standard LSTM structure, and the temporal attention weights are calculated using the following formula: ,in Let v be the LSTM hidden state at time step t. , The parameters are learnable, the time series length T = 114, and the final output is the temporal fusion feature. .

[0065] This embodiment uses the PyTorch 1.9 framework to implement the above network, employs the Adam optimizer, has an initial learning rate of 0.001, 100 training epochs, and a batch size of 32.

[0066] III. S3: Carbon Storage Estimation 3 Model Construction and Inversion

[0067] 3.1 Construction of Candidate Feature Set

[0068] Based on the multi-source remote sensing fusion features (128-dimensional temporal fusion features) output by S2 and the original auxiliary data, a total of 150 candidate feature factors were constructed, including:

[0069] Spectral characteristics: reflectance in each band, normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), soil-regulated vegetation index (SAVI), normalized difference moisture index (NDWI), red edge position index (REP), etc., totaling 38.

[0070] Radar features include: backscattering coefficients (VV, VH), polarization decomposition parameters (Entropy, Alpha, Anisotropy), and texture features (contrast, correlation, homogeneity), totaling 26 features.

[0071] LiDAR structural features include: average tree height, maximum tree height, canopy coverage, canopy undulation index, leaf area index (LAI), etc., totaling 15 features.

[0072] Topographic features: elevation, slope, aspect, topographic humidity index (TWI), mountain shadow, etc., a total of 12.

[0073] Climate characteristics: annual average temperature, annual precipitation, growing season accumulated temperature, drought index, etc., a total of 18.

[0074] Soil characteristics: soil type code, soil texture (content of sand, silt, and clay), soil pH value, etc., totaling 41 items.

[0075] 3.2 Adaptive Feature Selection

[0076] Embedded feature selection is performed using the maximum correlation minimum redundancy (mRMR) criterion based on mutual information, with the objective function being: ,in, For mutual information, For the selected feature subset, For candidate feature set, For carbon storage labeling, a greedy forward search was used to select 28 core feature factors from 150 candidate features, including NDVI, EVI, radar backscattering coefficient, average tree height, TWI, and annual precipitation.

[0077] 3.3 Deep Forest Model Construction

[0078] A carbon storage estimation model is constructed using the Deep Forest algorithm. The model structure is as follows:

[0079] Multi-granularity scanning layer: Three sliding windows of different sizes are used (window sizes are 100×100 pixels, 50×50 pixels and 25×25 pixels, with strides of 50 pixels, 25 pixels and 12 pixels respectively). Each window is processed by a fully random forest (100 decision trees) and a regular random forest (100 decision trees), and the output class probability vector (discretizing carbon reserves into 10 levels) is output. The outputs of multi-granularity scanning are concatenated to form an enhanced feature vector.

[0080] Cascaded Forest Layers: This layer comprises 5 layers, each consisting of 4 fully random forests and 4 ordinary random forests. Each forest contains 300 decision trees, and the update rules are implemented layer by layer according to the cascaded forest structure. ,in, For the first The input features of the layer For the first The class probability vectors output by the layer are controlled for overfitting using 3-fold cross-validation.

[0081] Output layer: The output of the last layer of the cascaded forest is averaged to obtain the estimated vegetation carbon storage value (unit: tC / ha) for each pixel.

[0082] 3.4 Coordinated Inversion of Soil Carbon Storage

[0083] Soil organic carbon density was retrieved hierarchically using a Bayesian neural network (BNN). The input variables were: estimated vegetation carbon storage, radar penetration signal index (calculated from VV / VH), topographic humidity index (TWI), mean annual temperature, annual precipitation, and soil type encoding (one-hot encoding). The BNN network contained three hidden layers (128 neurons per layer), and the output was the soil organic carbon density at two depths: 0–20 cm and 20–100 cm. Variational inference was used for training, and the loss function was variational lower bound (ELBO). ,in, For training data, For network weights, For variational parameters, The divergence is Kullback-Leibler, with 200 training epochs and a learning rate of 0.001. Monte Carlo dropout is used to approximate variational inference (sampling times 10).

[0084] The soil carbon storage (tC / ha) is obtained by multiplying the layered soil organic carbon density by the corresponding area and soil layer thickness (0~20cm: 0.2m, 20~100cm: 0.8m). The total carbon storage is the sum of vegetation carbon storage and soil carbon storage.

[0085] IV. S4: Carbon Storage Prediction Model Construction

[0086] 4.1 Construction of the Spatiotemporal Graph Structure

[0087] Using a 1km × 1km spatial grid as nodes (approximately 500,000 nodes in total; to reduce computational load, this embodiment divides the study area into 5,000 1km grids), an undirected graph G = (V, E) is constructed using the spatial adjacency relationships (Rook adjacency, i.e., shared edges) of adjacent grids as edges. The node characteristics are the monthly carbon storage sequence estimated by S3 (2020-2024, a total of 60 months).

[0088] 4.2 ST-GCN Model Construction

[0089] Spatiotemporal Graph Convolutional Neural Network (ST-GCN) is employed.

[0090] Normalized Laplace matrix: ,in, It is an adjacency matrix. It is a diagonal matrix of degree. It is an identity matrix.

[0091] Graph convolution operation (Chebyshev polynomial approximation): ,in, To scale the Laplacian matrix, For Chebyshev polynomials, For learnable parameters, the order of the polynomial =3, this embodiment sets 2 graph convolutional layers, each layer outputs 128-dimensional features.

[0092] Temporal convolutional layers: One-dimensional causal convolution is used, with a kernel size of 3 and an increasing dilation factor (1, 2, 4, 8) layer by layer, for a total of 4 stacked temporal convolutional modules. Each layer is followed by a gated linear unit (GLU) activation function.

[0093] Covariate input: Land use change scenario data (urban expansion, farmland protection, and ecological priority) and climate scenario data (SSP1-2.6, SSP2-4.5, and SSP5-8.5 in CMIP6) are used as additional node features, mapped to 128 dimensions through a fully connected layer, and then added element-wise with the graph convolution output.

[0094] Output: The model outputs monthly carbon storage forecasts (spatial distribution map) for the next 5 years (2025-2029) and generates 95% confidence intervals. The model is implemented using the PyTorch Geometric library, with 100 training epochs, an initial learning rate of 0.001, and a batch size of 32.

[0095] V. S5: Online Incremental Learning Mechanism

[0096] Online incremental learning is achieved using the Elastic Weight Consolidation (EWC) regularization method, with the loss function being... ,in, For the loss function on the new data, These are the optimal parameters obtained from historical training. The diagonal elements of the Fisher information matrix are calculated using the following formula: The balance coefficient λ = 10. When new multi-source remote sensing data is acquired each month, the model is fine-tuned with the current parameters as the initial values: 20 rounds of fine-tuning, batch size 32, learning rate is 1 / 10 of the initial learning rate (0.001) (i.e. 0.0001), and the optimizer is Adam.

[0097] VI. Accuracy Verification and Results

[0098] Accuracy was validated using 35 independent validation plots (not involved in model training), with the coefficient of determination as the evaluation metric. ), Root Mean Square Error (RMSE, unit: tC / ha), Mean Absolute Percentage Error (MAPE).

[0099] Verification results:

[0100] Vegetation carbon storage inversion = 0.91, RMSE = 7.2tC / ha, MAPE = 12.3%.

[0101] Soil carbon storage inversion (0~20cm): = 0.86, RMSE = 4.1tC / ha, MAPE = 15.8%.

[0102] Total carbon storage inversion: = 0.90, RMSE = 8.9tC / ha, MAPE = 13.5%.

[0103] Carbon storage forecast for the next 5 years (2025): Comparison with the reserved measured data for 2025. =0.83, RMSE=10.2tC / ha, prediction accuracy (the proportion of grid points with relative error <20%) is 84.

[0104] Comparative Experiment: Compared with the comparative example (using only optical remote sensing data + random forest), the total carbon storage estimation of this invention... The RMSE increased by 0.23 (from 0.67 to 0.90) and decreased by 51% (from 18.2 to 8.9 tC / ha), which was statistically significant (p<0.01, paired t-test).

[0105] Results visualization: Outputs a spatial distribution map of carbon storage in the study area in December 2024 (30-meter resolution), as well as time-series curves and spatial variation probability maps of carbon storage prediction under three climate scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5) from 2025 to 2029. The results are published through a WebGIS platform and support interactive querying by users.

[0106] Example 2

[0107] like Figures 1 to 5 This embodiment is basically the same as Embodiment 1, except that the study area is a wetland ecosystem in a certain province (approximately 100,000 hectares). The multi-source remote sensing data has been adjusted accordingly to be suitable for wetlands: Sentinel-1 water system mapping products have been added, and PALSAR-2 L-band radar data has been added to enhance the penetration capability of wetland water. The carbon storage prediction model has added a hydrological scenario module (three scenarios: high-water year, normal-water year, and low-water year). The verification results show that the wetland vegetation carbon storage inversion... =0.85, RMSE=5.3tC / ha, prediction accuracy 80%.

[0108] Comparative Example

[0109] To verify the technical effects of the present invention, the following prior art was selected for comparison:

[0110] Comparative Example 1: Sentinel-2 optical remote sensing data only + random forest regression (no multi-source fusion, no prediction function).

[0111] Comparative Example 2: Multi-source data fusion based on GEE platform + BP neural network (refer to patent CN119598392A).

[0112] Comparative Example 3: Traditional ground-based inventory method (every 5 years).

[0113] A comparative evaluation was conducted within the same study area (the forest ecosystem of Example 1) and the same time period (2020–2024), and the results are shown in the table below:

[0114] method Total carbon storage RMSE (tC / ha) Monitoring frequency Predictive ability Ground-measured dependence Example 1 0.90 8.9 Monthly Continuous (Multiple scenarios, 5 years) Reduced by 75% Comparative Example 1 0.67 18.2 Follow-up visit period (16 days) none high Comparative Example 2 0.81 13.5 During the revisit period none middle Comparative Example 3 0.95 (sample plot only) 5.2 (Sample plot only) Once every 5 years none Extremely high

[0115] As can be seen from the table above, the method of the present invention is close to the ground survey method in terms of estimation accuracy. (vs0.95), but the monitoring frequency has been increased from once every 5 years to monthly continuous monitoring, and it has multi-scenario prediction capabilities that existing methods do not have at all, compared to Comparative Example 2 (GEE+BP neural network). Improved by 11.1%, and reduced RMSE by 34.1%, compared to Comparative Example 1 (single optical data). The results show a 34.3% increase and a 51.1% decrease in RMSE, which verifies the significant technological advancement and practicality of the method of this invention.

[0116] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0117] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0118] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for dynamic monitoring and prediction of ecosystem carbon storage based on multi-source remote sensing data, characterized in that, Includes the following steps: S1: Acquisition and preprocessing of multi-source remote sensing data (1) - Acquire multi-source remote sensing data of the study area and perform spatiotemporal registration preprocessing. The multi-source remote sensing data includes optical remote sensing image data, radar remote sensing data, LiDAR point cloud data, UAV multispectral data and ground measured data. The spatiotemporal registration preprocessing includes generating high spatiotemporal resolution optical remote sensing data using a spatiotemporal adaptive reflectivity fusion model. S2: Multi-scale spatiotemporal fusion (2) — Construct a multi-scale spatiotemporal fusion network, perform spatiotemporal resolution alignment and feature-level fusion on the preprocessed multi-source remote sensing data, and generate a multi-source remote sensing fusion feature set. The multi-scale spatiotemporal fusion network includes a spatial attention module and a temporal feature extraction module. The spatial attention module uses a cross-modal attention mechanism to realize the spatial feature alignment between optical images and radar data. The temporal feature extraction module uses an improved long short-term memory network to extract the temporal evolution features of multi-temporal remote sensing data. S3: Carbon storage estimation (3) - Based on the multi-source remote sensing fusion feature set generated in S2, an adaptive feature selection network is used to screen the core feature factors related to ecosystem carbon storage, and a carbon storage estimation model based on deep forest algorithm is constructed to output the carbon storage estimation value of each grid point in the study area. The carbon storage estimation value includes vegetation carbon storage and soil carbon storage, wherein the soil carbon storage is retrieved by Bayesian neural network collaborative inversion. S4: Carbon storage prediction (4) — Using the carbon storage estimate output by S3 as the baseline data, a carbon storage prediction model coupled with the land use dynamics module and the climate scenario module is constructed to predict the ecosystem carbon storage at future time nodes. The carbon storage prediction model adopts a spatiotemporal graph convolutional neural network, and its graph convolution operation adopts Chebyshev polynomial approximation. S5: Online Incremental Learning (5) — Establish an online incremental learning mechanism. When new multi-source remote sensing data is acquired, the carbon storage estimation model and prediction model are dynamically updated to realize continuous dynamic monitoring of carbon storage. The online incremental learning mechanism adopts the elastic weight consolidation regularization method.

2. The method according to claim 1, characterized in that, The spatiotemporal adaptive reflectivity fusion model in S1 adopts the following formula: ； in, For the predicted time High-resolution reflectivity, For low-resolution reflectivity, For high-resolution reflectivity, The weighting coefficients are calculated based on time difference and spectral similarity, satisfying... , where n is the number of time phases involved in the fusion.

3. The method according to claim 1, characterized in that, The cross-modal attention mechanism in S2 uses the following formula: ； Where Q is the optical feature flattening matrix, and K and V are the radar feature flattening matrices. For feature dimensions.

4. The method according to claim 1, characterized in that, The improved Long Short-Term Memory network in S2 uses the following formula for calculating temporal attention weights: ； in, Let v be the LSTM hidden state at time step t. , Here, T is a learnable parameter, and T is the timing length.

5. The method according to claim 1, characterized in that, The adaptive feature selection network in S3 adopts the following objective function: ； in, For mutual information, For the selected feature subset, For candidate feature set, For carbon storage labels; The deep forest algorithm employs the following cascaded forest layer-by-layer update rule: ； in, For the first The input features of the layer For the first The class probability vector output by the layer; The variational lower bound loss function of the Bayesian neural network is: ； in, For training data, For network weights, For variational parameters, The divergence is Kullback-Leibler.

6. The method according to claim 1, characterized in that, The normalized Laplacian matrix of the spatiotemporal graph convolutional neural network in S4 is defined as follows: ； in, It is an adjacency matrix. It is a diagonal matrix of degree. It is the identity matrix; The graph convolution operation is approximated using Chebyshev polynomials: ； in, To scale the Laplacian matrix, For Chebyshev polynomials, For learnable parameters, Let be the order of the polynomial.

7. The method according to claim 1, characterized in that, The loss function of the elastic weight consolidation regularization method in S5 is: ； in, For the loss function on the new data, These are the optimal parameters obtained from historical training. The diagonal elements of the Fisher information matrix are calculated using the following formula: ； Wherein, λ is the balance coefficient, and its value ranges from 1 to 100.

8. A dynamic monitoring and prediction system for ecosystem carbon storage based on multi-source remote sensing data, characterized in that, include: The multi-source remote sensing data acquisition module is used to acquire optical remote sensing image data, radar remote sensing data, LiDAR point cloud data, UAV multispectral data and ground measurement data of the study area; The data preprocessing module is used to perform spatiotemporal registration preprocessing on the acquired multi-source remote sensing data; The multi-scale spatiotemporal fusion module has a built-in multi-scale spatiotemporal fusion network, which is used to perform spatiotemporal resolution alignment and feature-level fusion on preprocessed multi-source remote sensing data; The carbon storage estimation module has a built-in carbon storage estimation model based on the deep forest algorithm, which is used to output the carbon storage estimation value of each grid point in the study area. The carbon storage prediction module has a built-in carbon storage prediction model that couples the land use dynamics module and the climate scenario module, which is used to predict future carbon storage under different scenarios. The dynamic update module is used to establish an online incremental learning mechanism to achieve real-time updates of model parameters; The results output module is used to visualize the monitoring and prediction results of carbon storage.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method of any one of claims 1 to 7.

10. The application of the method according to any one of claims 1 to 7 in monitoring carbon storage in forest ecosystems, wetland ecosystems, or grassland ecosystems.

Images