Ecological carbon sink potential prediction method and system based on fusion attention mechanism

Abstract

The invention discloses an ecological carbon sink potential prediction method and system based on a fusion attention mechanism, and relates to the technical field of ecological carbon sink monitoring and image processing, and the method comprises the steps: obtaining a to-be-evaluated scene reflectivity image and meteorological variable data of a to-be-evaluated time period; image local features are extracted from the reflectivity image through CNN, and meteorological variable features are extracted from meteorological variable data through BiLSTM; calculating a meteorological time sequence vector based on an attention mechanism; dynamic weights of the image local features and the meteorological time sequence vector features are calculated, and weight fusion features are obtained through the dynamic weights corresponding to the image local features and the meteorological time sequence vector features; inputting the weight fusion features into a full connection layer for processing, and mapping feature variables to a prediction space to obtain predicted ecological carbon sink data of the scene to be evaluated in the time period to be evaluated; the technical difficulty that the carbon sink change cannot be accurately predicted under the large-range and high-precision practical demand of the current land ecosystem carbon sink estimation is solved.

Description

Technical Field

[0001] This invention relates to the field of ecological carbon sink monitoring and image processing technology, and more specifically, to a method and system for predicting ecological carbon sink potential based on a fusion attention mechanism. Background Technology

[0002] Traditional methods are mainly divided into two categories based on data sources and model approaches: "top-down" and "bottom-up." "Bottom-up" methods extend ground observations and simulations at the sample point or grid scale to the regional scale, primarily relying on ground sampling. Commonly used methods include inventory methods, eddy covariance methods, and ecological process simulation methods. "Top-down" methods mainly refer to carbon sink retrieval techniques, which calculate terrestrial ecosystem carbon sinks using radiative transfer models based on observation data from different remote sensing platforms. This model-based simulation method, due to its clear technical solutions and efficient algorithm execution, can achieve large-scale ecosystem carbon sink assessments, making it an effective technique favored by many ecologists. Meanwhile, with the rapid development of remote sensing and computer vision technologies, estimating terrestrial ecosystem carbon sinks using high-resolution satellite remote sensing data has become a research hotspot. However, due to the complex environment of terrestrial ecosystems, accurately describing the quantitative relationship between past and future terrestrial ecosystem carbon sinks and ecological environmental factors remains a significant challenge. Existing "top-down" ecological carbon sink estimation techniques mainly include: Patent CN117575359B discloses an ecosystem carbon sink assessment method for unified supervision. The method uses land patches as unified spatial units and determines the amount of carbon dioxide that the ecosystem can absorb and fix by inputting ecosystem carbon sink assessment data into the ecosystem carbon sink assessment model.

[0003] Patent CN115310757B discloses a high-precision, dynamically variable resolution ecological carbon sequestration assessment method and system. This scheme is based on the YIBs vegetation model and adapts to data with different formats and spatial resolutions by modifying the spatial resolution and format of the data at multiple spatial scales. The converted ecological carbon sequestration-related variables are then input into the YIBs vegetation model to calculate the carbon sequestration amount. This method effectively solves the problem of data format and resolution mismatch caused by different data sources through spatial interpolation and scaling methods.

[0004] Patent CN119003679B discloses a remote sensing-based method for monitoring ecosystem carbon sinks. This method involves collecting remote sensing data of wetland protected areas in real time via remote sensing satellites, parsing and calculating wetland carbon sink data at ground stations, and then constructing a carbon sink monitoring model to calculate and output the regional carbon sink amount and extend it to the overall carbon sink amount.

[0005] The paper, "Estimation of Carbon Sinks in Terrestrial Ecosystems and Identification of Protected Areas in Zhejiang Province under Future Climate Backgrounds," proposes using an Integrated Biosphere Model (ISIS) to simulate terrestrial ecosystem carbon sinks at a kilometer-grid accuracy and identifies and delineates the spatial extent of carbon sink protected areas based on different climate backgrounds. Through quantitative modeling, the paper explores the level and spatial distribution characteristics of terrestrial ecosystem carbon sinks at the regional scale and quantitatively assesses the effectiveness of optimization strategies in maintaining terrestrial ecosystem carbon sinks.

[0006] The existing technical solutions for ecological carbon sequestration based on the "top-down" approach have the following main problems: The patent with announcement number CN117575359B assessed the total ecological carbon by constructing a unified evaluation unit, but the proposal lacked a technical description of the carbon sink estimation and prediction model, and did not establish or clarify the quantitative relationship between the total ecological carbon and climate environmental factors.

[0007] The patent with announcement number CN115310757B constructs a YIBs vegetation model and uses a multi-scale spatial resolution and format dynamic correction method to realize a high-precision dynamic variable resolution ecological carbon sink assessment system based on physical processes. However, this system relies on the YIBs model to achieve high-precision ecological carbon sink simulation. It requires multiple parameters for input data, including control information, calculation parameters, driving data, initial field data, and environmental data. Each data item is further subdivided into multiple data sources, which has significant limitations when applied to long-term, large-scale ecological carbon sink assessment.

[0008] The patent with announcement number CN119003679B mentions using real-time data from remote sensing satellites to construct a carbon sink estimation model for monitoring ecosystem carbon sinks. However, this scheme only estimates carbon sinks based on real-time data for a region and cannot achieve quantitative assessment and prediction of long-term, large-scale ecological carbon sinks for both historical and future periods.

[0009] The paper "Estimation of Carbon Sinks in Terrestrial Ecosystems and Identification of Protected Areas in Zhejiang Province under Future Climate Background" uses the Integrated Biosphere Model (IBIS) to simulate carbon sinks in terrestrial ecosystems at the kilometer grid scale and explores the level of carbon sinks in terrestrial ecosystems and their spatial distribution characteristics at the regional scale. However, under the application requirements of fine ecological carbon sink assessment, the kilometer-scale results are difficult to meet the actual operational needs.

[0010] Therefore, this application is hereby submitted. Summary of the Invention

[0011] The purpose of this invention is to provide a method and system for predicting ecological carbon sink potential based on a fusion attention mechanism. Targeting high-resolution ecological carbon sink assessment scenarios, this invention uses ecological carbon sink estimation and ecological carbon sink potential prediction models, combined with remote sensing reflectance imagery and future climate scenario data, to achieve long-term, high-spatial-resolution estimation and prediction of terrestrial ecosystem carbon sinks at regional and even global scales. This solves the technical challenge of accurately predicting carbon sink changes under the current real-world demand for large-scale and high-precision terrestrial ecosystem carbon sink estimation.

[0012] The above-mentioned technical objective of the present invention is achieved through the following technical solution: Firstly, this application provides a method for predicting the potential of ecological carbon sinks based on a fusion attention mechanism, including the following specific steps: S1, acquire reflectance images of the scene to be evaluated, as well as meteorological variable data of the scene to be evaluated in the future evaluation period; S2, local image features are extracted from reflectance images using CNN, and meteorological variable features are extracted from meteorological variable data using BiLSTM; S3, based on the attention mechanism, the meteorological variable features are weighted and the meteorological time series vector corresponding to the meteorological variable features with fused attention weights is calculated; S4, based on the dynamic weight allocation mechanism, calculates the dynamic weights of local image features and meteorological time-series vector features respectively, and obtains the weighted fusion features through the corresponding dynamic weights; S5 inputs the weighted fusion features into the fully connected layer for processing, and maps the feature variables to the prediction space to obtain the predicted ecological carbon sink data of the scenario to be evaluated during the evaluation period.

[0013] Based on the above technical solution, the present invention can be further improved as follows.

[0014] Furthermore, the above calculation yields a meteorological time-series vector corresponding to the fused attention weights for the meteorological variable features, specifically: ,in: , ; In the formula, Representing the characteristics of meteorological time series vectors, For calculating attention weights, Let T represent the attention score at time step t, and T represent the total number of time steps. , , They represent the learnable parameters, This represents the hidden state of the BiLSTM at time step t. This represents the context vector.

[0015] Furthermore, the aforementioned dynamic weight allocation mechanism is specifically as follows: ; ; In the formula, This represents the dynamic weights corresponding to local features of the image. Dynamic weights representing the characteristics of meteorological time-series vectors. , They represent the learnable parameters, The quality of remote sensing images is represented by the proportion of cloud cover, and the reflectance image is the reflectance image. ; To assess the reliability of meteorological variable data, model uncertainty in the CMIP6 dataset is used as a representation. ; For temporal integrity, it is characterized by the data missing rate, and .

[0016] Furthermore, the aforementioned weighted fusion features are specifically as follows: ; In the formula, For weighted fusion features, Indicates local features of the image. Representing the characteristics of meteorological time series vectors, This represents the dynamic weights corresponding to local features of the image. The dynamic weights represent the characteristics of meteorological time series vectors.

[0017] Furthermore, steps S2-S5 above are executed using the constructed CNN-BiLSTM-Attention hybrid model, which is obtained through the following method: Construct training samples, which include spatiotemporally aligned historical meteorological data, historical reflectance images, and historical carbon sink data; The historical time-series reflectance images and the historical meteorological data for a specified period are input into a preset hybrid network model for model training to obtain the predicted carbon sink data for the specified period. The loss function of the hybrid network model is calculated based on the predicted carbon sink data and the historical carbon sink data for the corresponding period until the training termination condition is met. The hybrid network model that meets the training termination condition is determined as the CNN-BiLSTM-Attention hybrid model.

[0018] Furthermore, the loss function described above is as follows: ; In the formula, Represents historical carbon sequestration data. To predict carbon sink data, This represents the number of samples.

[0019] Furthermore, the aforementioned historical carbon sequestration data was obtained through the following methods: Optical reflectance images were acquired using Sentinel-2A, leaf area index was retrieved using a radiative transfer model, and historical carbon sink data were estimated using a two-leaf light energy utilization model.

[0020] Secondly, this application provides an ecological carbon sink potential prediction system based on a fusion attention mechanism, applicable to any one of the ecological carbon sink potential prediction methods based on a fusion attention mechanism in the first aspect, comprising: The sample data acquisition module is used to acquire reflectance images of the scene to be evaluated, as well as meteorological variable data of the scene to be evaluated in the future evaluation period. The data feature extraction module is used to extract local image features from reflectance images using CNN and meteorological variable features from meteorological variable data using BiLSTM. The meteorological feature weighting module is used to weight meteorological variable features based on the attention mechanism and calculate the meteorological time series vector corresponding to the meteorological variable features with fused attention weights. The fusion feature acquisition module is used to calculate the dynamic weights of local image features and meteorological time-series vector features based on a dynamic weight allocation mechanism, and obtain weighted fusion features through the corresponding dynamic weights. The ecological carbon sink calculation module is used to input the weighted fusion features into the fully connected layer for processing, and to map the feature variables to the prediction space to obtain the predicted ecological carbon sink data of the scenario to be evaluated during the evaluation period.

[0021] Thirdly, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method of any one of the first aspects.

[0022] Fourthly, this application provides a non-transitory computer-readable storage medium that stores computer instructions that cause a computer to perform any of the methods in the first aspect.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects: 1. Compared with the current technology of using the Integrated Biosphere Model (IBIS) to simulate terrestrial ecosystem carbon sinks at the kilometer grid scale, this application proposes to combine the advantages of CNN to extract local features, BiLSTM to capture long-term dependencies in long-term data, and attention mechanism to dynamically adjust the attention to the input sequence. This can achieve terrestrial ecological carbon sink estimation at the ten-meter level and predict the potential level and changing trend of terrestrial ecological carbon sinks of a considerable number of orders of magnitude up to 2300.

[0024] 2. Breaking away from the dependence of traditional models on specific data formats and resolutions, this model adapts to multi-source remote sensing and climate data through a dynamic weighting mechanism and a multimodal data preprocessing module, making it suitable for carbon sink prediction at different regional and time scales. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the prediction method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the connection of the prediction system in an embodiment of the present invention; Figure 3 This is a schematic diagram of the connection of an electronic device in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0029] In the description of the embodiments of the present invention, "multiple" means at least two.

[0030] Example 1: To address the technical challenge of accurately predicting carbon sink changes in terrestrial ecosystem carbon sink estimation under the current practical needs of large-scale and high-precision estimation, this example provides a method for predicting ecological carbon sink potential based on a fusion attention mechanism, such as... Figure 1 As shown, the specific steps include the following: S1, acquire reflectance images of the scene to be evaluated, as well as meteorological variable data for the scene to be evaluated in the future evaluation period.

[0031] Among them, the Sentinel-2A can be used to acquire optical reflectance images. For meteorological variable data for future assessment periods, the data is based on the integration of historical and future climate scenario data from the Coupled Model Intercomparison Project Phase 6 (CMIP6), which mainly includes four climate scenarios: SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5. The data can be obtained up to 2300 years in advance, including meteorological variable data such as temperature, precipitation, and radiation.

[0032] S2 extracts local image features from reflectance images using CNN and meteorological variable features from meteorological variable data using BiLSTM.

[0033] Among them, local image features can be implemented using one-dimensional convolutional layers (CNN-1d), which can support multiple input features and adapt to the complex scenarios of multi-feature prediction in terrestrial ecosystem carbon sinks; its main calculation process can be abstracted into the following formula: ; In the formula, It is an index in the batch. It is the index of the output channel. It is the position index in the convolution kernel. It is the expansion coefficient. It is the size of the fill on the left. This represents the position index of the input sequence. c is Input channel index, in channels This represents the length of the input sequence; the length of the output sequence (output channels) can be calculated based on parameters such as the input sequence length, kernel size, stride, and padding, as shown in the following formula: In the formula, Indicates the number of input channels. This indicates the number of output channels of the convolutional layer. Indicates the kernel size. This represents the stride of the convolution. Indicates the fill amount.

[0034] Furthermore, meteorological variable characteristics are processed using BiLSTM. BiLSTM (Bidirectional Long Short-Term Memory) is a special type of recurrent neural network (RNN) that captures the contextual information at each time step of the sequence by simultaneously processing forward and backward information of the sequence data, effectively solving the problems of vanishing and exploding gradients in RNNs. BiLSTM is often used to capture long-term dependencies in long-term time-series data, controlling information transmission through gating mechanisms (input gate, forget gate, output gate). It mainly consists of two LSTM networks: a forward LSTM and a backward LSTM, which can be mathematically abstracted as follows: For the input sequence X={x1,x2,x3,…x t The forward LSTM processes the input sequentially from the first element to the last element of the sequence and generates the forward hidden state. .

[0035] Forgotten Gate: ; Input Gate: ; Candidate update unit status: ; Cell status update: ; Output gate: ; Hidden status update: .

[0036] Furthermore, the backward LSTM processes the output from the last element to the first element of the sequence and generates the backward hidden state. .

[0037] Forgotten Gate: ; Input Gate: ; Candidate update unit status: ; Cell status update:

[0038] Output gate: ; Hidden status update: ; Hidden state fusion: At each time step t, the forward and backward hidden states are concatenated to form the final bidirectional hidden state. ; In the above formula, , , These represent the forget gate, input gate, and output gate, respectively. Neuron state, It is in a hidden state. It is a weight matrix. It is a bias term. It is the sigmoid function. The superscripts f and b correspond to the forward and backward LSTM respectively. This indicates pixel-by-pixel multiplication. This represents the input sequence.

[0039] Specifically, BiLSTM utilizes the contextual information before and after each time step in the time series to provide a more comprehensive feature representation when processing sequence data, thereby obtaining the meteorological variable features mentioned above.

[0040] S3, based on the attention mechanism, performs weighted processing on the meteorological variable features and calculates the meteorological time series vector corresponding to the meteorological variable features with fused attention weights.

[0041] Optionally, the above calculation yields a meteorological time-series vector corresponding to the fused attention weights for the meteorological variable features, specifically: ,in: ,

[0042] In the formula, Representing the characteristics of meteorological time series vectors, For calculating attention weights, Let T represent the attention score at time step t, and T represent the total number of time steps. , , They represent the learnable parameters, This represents the hidden state of the BiLSTM at time step t. This represents the context vector.

[0043] S4, based on the dynamic weight allocation mechanism, calculates the dynamic weights of local image features and meteorological time-series vector features respectively, and obtains the weighted fusion features through the corresponding dynamic weights.

[0044] The dynamic weight allocation mechanism dynamically adjusts the contribution of each network (CNN / BiLSTM) based on real-time data quality (such as cloud cover in remote sensing images and climate data uncertainty). This mechanism aims to maximize prediction accuracy by updating weight parameters using a policy gradient algorithm, thereby improving the model's robustness in complex climatic environments. The specific steps are as follows: 1) Data quality indicator quantification: Real-time extraction of quality characteristics of input data, mainly including: Remote sensing image quality: characterized by cloud cover percentage ( The smaller the value, the higher the image quality.

[0045] Climate data reliability: Characterized by model uncertainty in the CMIP6 dataset ( The smaller the value, the higher the reliability.

[0046] Temporal integrity: characterized by data missing rate ( The smaller the value, the better the data integrity.

[0047] 2) Dynamic Weight Function: A dynamic weight function is designed based on data quality metrics to output the CNN network weights. and BiLSTM weights When the quality of remote sensing data is high ( (smaller values), increase the weights of the CNN network. The importance of enhancing local features. When the reliability of climate time-series data is high ( , (If the value is small), increase the BiLSTM weights. This enhances the importance of time-series features; the dynamic weighting function is calculated as follows: , ;in, , Set initial values ​​for learnable parameters. = =0.5.

[0048] 3) Weight fusion and prediction output: The local features extracted by the CNN are fused together... ) and BiLSTM extraction of temporal features ( Based on dynamic weight fusion: The fused features are processed through an attention mechanism and a fully connected layer to output a predicted value of carbon sequestration potential.

[0049] 4) Weight parameter optimization: Using the mean squared error (MSE) between the predicted and actual values ​​as the loss function, the parameters of the weight function are optimized synchronously during model training. , ).

[0050] S5 inputs the weighted fusion features into the fully connected layer for processing, and maps the feature variables to the prediction space to obtain the predicted ecological carbon sink data of the scenario to be evaluated during the evaluation period.

[0051] Among them, the fully connected layer (FC) will be obtained through weight fusion. Mapped to the prediction space, future carbon sequestration data is output through a fully connected layer; the output layer uses a linear activation function, and the loss function can be the mean squared error (MSE); its mathematical expression can be: In the formula, For the predicted results, This is the weight matrix. Bias term.

[0052] Optionally, steps S2-S5 above are executed using the constructed CNN-BiLSTM-Attention hybrid model, which is obtained in the following way: S51. Construct training samples, which include spatiotemporally aligned historical meteorological data, historical reflectance images, and historical carbon sink data.

[0053] Specifically, the net primary productivity (NPP) of ecological carbon sinks and climate variable data such as temperature, precipitation, and radiation from different historical (2017-2025) scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5) in the Sixth Coupled Model Intercomparison Project (CMIP6) will be obtained. The data will be sampled to the same resolution using spatiotemporal alignment methods (such as bilinear interpolation and spatiotemporal kriging interpolation), and time-series alignment will be performed using time-series labels (DOY). The data will be placed in the same folder as a training dataset in a sample-label manner.

[0054] S52 inputs historical time-series reflectance images and historical meteorological data for a specified period into a preset hybrid network model for processing to obtain predicted carbon sink data for the specified period.

[0055] S53 calculates the loss function of the hybrid network model based on the predicted carbon sink data and the historical carbon sink data of the corresponding period until the training termination condition is met. The hybrid network model that meets the training termination condition is determined as the CNN-BiLSTM-Attention hybrid model.

[0056] This study utilizes a CNN-BiLSTM-Attention hybrid model built on the PyTorch deep learning framework. The datasets are divided into training, validation, and test datasets in an 8:1:1 ratio using time-series segmentation. During training, the mean squared error (MSE) is chosen as the loss function to measure the difference between the model's predictions and the actual values. The loss function calculation formula is as follows: ; In the formula, Represents historical carbon sequestration data. To predict carbon sink data, The sample size is 1; the optimizer can be Adam optimizer, the initial learning rate is 0.001, the learning rate strategy uses cosine annealing to dynamically adjust the learning rate, and the batch size can be set to 8; after training, the model performance can be evaluated using the test dataset, and evaluation metrics such as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) can be calculated.

[0057] Specifically, by integrating the PROSAIL+TL-LUE carbon storage estimation technology route and CMIP6 historical and future multi-scenario climate data, interpolation methods are used to align climate variables with ecological carbon sink estimation results in time and space, forming a multi-source, multi-modal dataset, and realizing full-chain modeling of historical carbon sink-future climate-potential prediction. At the same time, a deep fusion model of a one-dimensional convolutional neural network (CNN-1D) with an attention mechanism and a bidirectional long short-term memory network (BiLSTM) is applied to achieve the triple function synergy of local feature extraction, long-term dependence capture, and key period focus of carbon sink data, adapting to the characteristics of ecological carbon sink "local climate response + long-term cumulative effect", thereby improving the accuracy of prediction.

[0058] The historical carbon sequestration data mentioned above was obtained through the following methods: Optical reflectance images were acquired using Sentinel-2A. The leaf area index (LAI) was retrieved using the PROSAIL radiative transfer model, and then combined with the two-leaf light use efficiency model (TL-LUE) to estimate historical terrestrial ecosystem carbon sinks, calculating net primary production (NPP). Subsequently, sliding window statistics and seasonal decomposition trend components were used, combined with climate factors (such as temperature, precipitation, and radiation) to form a multi-climate variable input. These climate variables were then mapped to carbon sink data for corresponding dates using time-series labels, integrating them to form a carbon sink dataset. This dataset includes historical meteorological data, historical reflectance images, and historical carbon sink data. 1. Regarding the PROSAIL radiative transfer model, it is derived by coupling the PROSPECT leaf model and the SAIL canopy structure model. This model explains vegetation reflectance from biophysical characteristics, thus expressing the optical properties of vegetation. The PROSPECT model (leaf scale) simulates the reflectance and transmittance of vegetation leaves in the 400-2500 nm range. The model has two types of input parameters: leaf reflectance and transmittance vary with structural and biochemical parameters such as leaf structure parameters, leaf chlorophyll content, water thickness, dry matter content, brown pigment, and leaf carotenoid content. The SAIL model (canopy scale) calculates the bidirectional reflectance of the canopy top as a function of leaf reflectance and transmittance derived from PROSPECT and other input parameters (LAI, average leaf tilt angle, hotspot size, solar zenith angle, field of view zenith angle, relative azimuth angle, brightness coefficient, and diffuse / direct light ratio). The PROSAIL model can be formally expressed as: (Equation 8-1) In the formula, Where N is the canopy reflectivity; N is the blade structural parameter; Chlorophyll content; Moisture content; Dry matter content; LAI is leaf area index; ALA is mean leaf tilt angle; For hotspot parameters; For soil moisture, For soil brightness; To observe the zenith angle; The solar zenith angle; This is the azimuth angle between the sun and the observation.

[0059] The PROSAIL model can be used to simulate the bidirectional reflectivity of the top of the canopy in the range of 400-2500 nm. Then, the LAI is inverted using a lookup table method. The main steps are as follows: 1) Generate a dataset combining LAI with other parameters, and calculate the corresponding... ; 2) Calculate the measured reflectance of the image and the LUT. The mean square error; 3) Select the LAI corresponding to the LUT record with the smallest RMSE as the inversion result, as shown in the following formula: (Equation 8-2) 2. Calculation of net primary productivity of carbon sinks in historical terrestrial ecosystems based on the TL-LUE model: 1) Inversion of Mao's primary productive forces The Two-Leaf Light Use Efficiency (TL-LUE) model, derived from the MOD17 algorithm, is based on the BEPS model. It divides the canopy into sun-facing and shade-facing leaf groups, improving the calculation of photosynthetically active radiation (APAR) and net primary productivity (GPP) absorbed by the canopy. The formula for calculating vegetation canopy net primary productivity (GPP) is as follows: (Equation 8-3) In the formula, and These are the maximum light energy utilization efficiencies of the sun-facing blade and the shaded blade, respectively. and It is the amount of incident photosynthetically active radiation (PAR) absorbed by leaves that are exposed to sunlight or shaded, and its calculation formula is as follows: (Equation 8-4) (Equation 8-5) in, Canopy reflectance related to vegetation type; The average leaf tilt angle is set to 60 degrees for canopies where the leaf angles are spherically distributed. The solar zenith angle; , and These represent the diffuse component, direct component, and diffuse PAR under the canopy, respectively. The formula for calculating multiple scattering of direct radiation is as follows: (Equation 8-6) and It refers to the leaf LAI that is exposed to and shaded by the sun, based on the canopy LAI and the clustering index ( Divide it by the solar zenith angle; The calculation formula is as follows: (Equation 8-7) In equation (8-3), and The scalar value for the lowest temperature of VPD is calculated as follows: (Equation 8-8) (Equation 8-9) In the above formula, VPD is the daily average VPD. The lowest temperature of the day can be calculated from half-hourly data. and , and These are the minimum and maximum parameter values ​​of saturated vapor pressure difference and temperature, determined based on vegetation type.

[0060] 2) NPP calculation Net primary productivity is gross primary productivity minus plant autotrophic respiration, and its calculation formula is as follows: (Equation 8-10) In the formula, NPP is the net primary productivity of vegetation, GPP is the gross primary productivity, and Ra is the autotrophic respiration of plants.

[0061] Furthermore, the spatiotemporal interpolation method described above is implemented through the following steps: 1) Data preprocessing and spatiotemporal benchmark unification Spatial baseline: The resolution of the net primary productivity data of vegetation was resampled to 1km resolution to facilitate adaptation with CMIP6 climate scenario data, and bilinear interpolation was used for processing.

[0062] Time base: Aggregating climate variables from CMIP 6-day scale climate scenario data to monthly scale means can be abstracted as follows: (Equation 8-11) in, Let be the mean of the climate variables in month t. Let t be the number of days in month t. This represents the meteorological data for day d of month t. Furthermore, carbon sink data is also synthesized to a monthly scale using a mean-based synthesis method to ensure complete alignment with the time-series labels of the meteorological data.

[0063] 2) Spatial interpolation: Spatial interpolation employs the spatiotemporal kriging method to interpolate low-resolution CMIP6 grid data to a 1km*1km grid size. Spatiotemporal kriging extends traditional kriging interpolation to the time dimension, and its predicted values... It is a known point The weighted sum. Its calculation formula is as follows: (Equation 8-12) Among them, weight By minimizing the prediction variance Solve for the solution that satisfies the unbiasedness constraint. .

[0064] The most crucial component of spacetime kriging interpolation is the spacetime mutability function. Describing spatial distance and time interval The following data differences: (Equation 8-13) 3) Spatiotemporal co-kriging interpolation method Calculate the spatial variability function and time variation function The spatiotemporal autocorrelation of the data is characterized by the following formula: (Equation 8-14) (Equation 8-15) in, For spatial lag distance, Due to time lag, and Z represents the lagged sample size, and Z represents the climate variable.

[0065] The spatiotemporal covariance function is: (Equation 8-16) in, , Where a represents the spatial range and b represents the temporal range. The base value.

[0066] Interpolation prediction: for target point Its climate variable prediction values for: (Equation 8-17) in, For interpolation weights, Let the Lagrange multiplier be a given number, satisfying the following constraints: , (Equation 8-18) in, The spatial distance between the target point and the sample point k. For time difference.

[0067] Example 2: This application provides an ecological carbon sink potential prediction system based on a fusion attention mechanism, applied to the ecological carbon sink potential prediction method based on a fusion attention mechanism in Example 1, such as... Figure 2 As shown, it includes: The sample data acquisition module is used to acquire reflectance images of the scene to be evaluated, as well as meteorological variable data of the scene to be evaluated in the future evaluation period. The data feature extraction module is used to extract local image features from reflectance images using CNN and meteorological variable features from meteorological variable data using BiLSTM. The meteorological feature weighting module is used to weight meteorological variable features based on the attention mechanism and calculate the meteorological time series vector corresponding to the meteorological variable features with fused attention weights. The fusion feature acquisition module is used to calculate the dynamic weights of local image features and meteorological time-series vector features based on a dynamic weight allocation mechanism, and obtain weighted fusion features through the corresponding dynamic weights. The ecological carbon sink calculation module is used to input the weighted fusion features into the fully connected layer for processing, and to map the feature variables to the prediction space to obtain the predicted ecological carbon sink data of the scenario to be evaluated during the evaluation period.

[0068] Example 3: This application provides an electronic device, such as... Figure 3 As shown, it includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method of Embodiment 1.

[0069] Example 4: This application provides a non-transitory computer-readable storage medium that stores computer instructions that cause a computer to execute the method of Example 1.

[0070] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0074] Those skilled in the art will understand that all or part of the steps in the above facts and methods can be implemented by a program instructing related hardware. The program or the program described therein can be stored in a computer-readable storage medium. When the program is executed, it includes the following steps: at this time, the corresponding method steps are introduced. The storage medium can be ROM / RAM, magnetic disk, optical disk, etc.

[0075] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting ecological carbon sink potential based on a fusion attention mechanism, characterized in that, The specific steps include the following: S1, acquire reflectance images of the scene to be evaluated, as well as meteorological variable data of the scene to be evaluated in the future evaluation period; S2, local image features are extracted from the reflectance image using CNN, and meteorological variable features are extracted from the meteorological variable data using BiLSTM; S3, the meteorological variable features are weighted based on the attention mechanism, and the meteorological time series vector corresponding to the meteorological variable features with fused attention weights is calculated; S4. Based on the dynamic weight allocation mechanism, the dynamic weights of the local image features and the meteorological time-series vector features are calculated respectively, and the weighted fusion features are obtained through the corresponding dynamic weights. S5, the weighted fusion features are input into the fully connected layer for processing, and the feature variables are mapped to the prediction space to obtain the predicted ecological carbon sink data of the scenario to be evaluated during the evaluation period.

2. The method for predicting ecological carbon sink potential based on a fusion attention mechanism according to claim 1, characterized in that, The calculation yields a meteorological time-series vector with fused attention weights corresponding to the meteorological variable features, specifically: ,in: ， ； In the formula, Representing the characteristics of meteorological time series vectors, For calculating the attention weights, Let T represent the attention score at time step t, and T represent the total number of time steps. , , They represent the learnable parameters, This represents the hidden state of the BiLSTM at time step t. This represents the context vector.

3. The method for predicting ecological carbon sink potential based on a fusion attention mechanism according to claim 1, characterized in that, The dynamic weight allocation mechanism is as follows: ； ； In the formula, This represents the dynamic weights corresponding to local features of the image. Dynamic weights representing the characteristics of meteorological time-series vectors. , They represent the learnable parameters, The quality of remote sensing images is represented by the proportion of cloud cover, and the reflectance image is the reflectance image. ; To assess the reliability of meteorological variable data, model uncertainty in the CMIP6 dataset is used as a representation. ; For temporal integrity, it is characterized by the data missing rate, and .

4. The method for predicting ecological carbon sink potential based on a fusion attention mechanism according to claim 3, characterized in that, The weighted fusion feature is specifically as follows: ； In the formula, For weighted fusion features, Indicates local features of the image. Representing the characteristics of meteorological time series vectors, This represents the dynamic weights corresponding to local features of the image. The dynamic weights represent the characteristics of meteorological time series vectors.

5. A method for predicting ecological carbon sink potential based on a fusion attention mechanism according to any one of claims 1-4, characterized in that, Steps S2-S5 are executed using the constructed CNN-BiLSTM-Attention hybrid model, which is obtained through the following method: Construct training samples, which include spatiotemporally aligned historical meteorological data, historical reflectance images, and historical carbon sink data; The historical time-series reflectance images and the historical meteorological data for a specified period are input into a preset hybrid network model for model training to obtain the predicted carbon sink data for the specified period. The loss function of the hybrid network model is calculated based on the predicted carbon sink data and the historical carbon sink data for the corresponding period until the training termination condition is met. The hybrid network model that meets the training termination condition is determined as the CNN-BiLSTM-Attention hybrid model.

6. The method for predicting ecological carbon sink potential based on a fusion attention mechanism according to claim 5, characterized in that, The loss function is specifically as follows: ； In the formula, Represents historical carbon sequestration data. To predict carbon sink data, This represents the number of samples.

7. The method for predicting ecological carbon sink potential based on a fusion attention mechanism according to claim 5, characterized in that, The historical carbon sequestration data was obtained through the following methods: Optical reflectance images were acquired using Sentinel-2A, leaf area index was inverted using a radiative transfer model, and historical carbon sink data were estimated using a two-leaf light energy utilization model.

8. A system for predicting ecological carbon sink potential based on a fusion attention mechanism, characterized in that, include: The sample data acquisition module is used to acquire reflectance images of the scene to be evaluated, as well as meteorological variable data of the scene to be evaluated in the future evaluation period. The data feature extraction module is used to extract local image features from the reflectance image using CNN and to extract meteorological variable features from the meteorological variable data using BiLSTM. The meteorological feature weighting module is used to perform weighted processing on the meteorological variable features based on the attention mechanism, and to calculate the meteorological time series vector corresponding to the meteorological variable features with fused attention weights; The fusion feature acquisition module is used to calculate the dynamic weights of the local image features and the meteorological time-series vector features based on a dynamic weight allocation mechanism, and obtain the weighted fusion features through the corresponding dynamic weights. The ecological carbon sink calculation module is used to input the weighted fusion features into the fully connected layer for processing, and to map the feature variables to the prediction space to obtain the predicted ecological carbon sink data of the scenario to be evaluated during the evaluation period.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the ecological carbon sink potential prediction method based on the fusion attention mechanism as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the ecological carbon sink potential prediction method based on the fusion attention mechanism according to any one of claims 1-7.

Images