A method for dynamic monitoring of forest carbon sequestration

Abstract

This invention provides a method for dynamic monitoring of forest carbon sinks, comprising the following steps: Step 1, acquiring multi-source remote sensing image data, meteorological data, and basic ecological data; Step 2, preprocessing the remote sensing data to construct carbon state characteristics; Step 3, using a lightweight Swing Transformer for carbon state perception modeling; Step 4, constructing a carbon cycle mechanism model and calculating theoretical carbon changes; Step 5, training and inference of a GAN error compensation model. This method is primarily driven by a carbon cycle mechanism model, ensuring that the results satisfy ecological and physical laws of the carbon cycle; it introduces an intelligent model for carbon state perception and error compensation, improving the accuracy of carbon storage estimation; and it achieves collaborative work between the mechanism model and the intelligent model, balancing stability and flexibility.

Description

Technical Field

[0001] This invention belongs to the field of ecological environment monitoring and remote sensing information processing technology, and specifically relates to a method for dynamic monitoring of forest carbon sequestration. Background Technology

[0002] Against the backdrop of global climate change, forest carbon sinks, as one of the most important carbon absorption and storage units in terrestrial ecosystems, are of great significance for mitigating greenhouse gas emissions and achieving carbon neutrality. How to conduct long-term, continuous, and precise monitoring of forest carbon storage and its dynamic changes is a key issue in current ecological environment management and carbon emission accounting.

[0003] Current technologies for monitoring forest carbon storage still face numerous limitations. First, while traditional ground-based sample plot surveys offer high accuracy, they are costly, time-consuming, and have limited spatial coverage, hindering large-scale, high-frequency dynamic monitoring. Second, optical remote sensing methods are susceptible to cloud cover, atmospheric conditions, and canopy saturation effects, exhibiting significantly reduced sensitivity to carbon storage in high-biomass areas, thus limiting estimation accuracy. Furthermore, existing remote sensing inversion models largely rely on empirical statistical relationships, lacking a sufficient representation of ecosystem processes, exhibiting weak generalization ability, and poor stability when applied across regions. Simultaneously, multi-source data fusion and scale conversion still face technical bottlenecks, hindering the effective construction of an integrated, continuous, and refined carbon sink monitoring system encompassing air, space, and ground. These shortcomings severely restrict the improvement of accurate forest carbon sink accounting and dynamic assessment capabilities, necessitating the development of more robust, physically well-defined, and operationally viable monitoring technologies. Summary of the Invention

[0004] Purpose of the invention: The technical problem to be solved by the present invention is to provide a dynamic monitoring method for forest carbon sinks, which is applicable to regional and even large-scale forest carbon sink monitoring, carbon storage accounting and ecological environment supervision.

[0005] The method includes the following steps:

[0006] Step 1: Acquire multi-source remote sensing image data, meteorological data, and basic ecological data;

[0007] Step 2: Preprocess the remote sensing data to construct carbon state characteristics;

[0008] Step 3, Lightweight Swing Transformer carbon state-aware modeling;

[0009] Step 4: Construct a carbon cycle mechanism model and calculate the theoretical carbon change;

[0010] Step 5: Training and inference of the GAN error compensation model.

[0011] Step 1 includes the following steps:

[0012] Step 101: Construct forest region definition and spatial grid;

[0013] First, the target forest monitoring area is determined based on the administrative divisions or ecological zoning boundaries of the study area. The target area is divided into regular raster units with a uniform spatial resolution:

[0014] ,

[0015] in, This represents the i-th grid cell; N is the total number of grid cells; each grid cell is used as an independent object for carbon storage calculation and updating.

[0016] Step 102: Collect multi-source remote sensing data;

[0017] At each time scale t, acquire multi-source remote sensing image data for the corresponding region. The remote sensing image data includes red band R data. Near-infrared (NIR) data SWIR data in the shortwave infrared band At the same time, radar imagery or lidar data are introduced as auxiliary input data;

[0018] Multi-source remote sensing image data Represented as:

[0019] ,

[0020] in, This represents auxiliary input data;

[0021] Step 103: Obtain data on ecological and environmental driving factors;

[0022] Simultaneous collection of ecological and environmental driving factor datasets related to forest carbon cycle processes Including temperature data Precipitation data Photosynthetically active radiation Soil moisture and soil organic carbon content The unified expression is:

[0023] ;

[0024] Step 104: Construct carbon storage data for ground samples;

[0025] Obtain measured forest plot carbon storage data , is represented as:

[0026] ,

[0027] in, Let be the biomass of the j-th vegetation type at time t; The biomass-to-carbon conversion coefficient; M is the number of vegetation types;

[0028] Step 105: Construct an initial carbon storage baseline;

[0029] Construct a baseline map of forest carbon storage at the initial stage of the system:

[0030] ,

[0031] Where i represents the grid number; It is the initial carbon storage value of the i-th grid cell; This represents the true initial carbon storage value of the i-th grid cell, obtained based on actual observations or inversion results.

[0032] Step 106: Data standardization and unification processing;

[0033] Standardization processing was performed on all multi-source remote sensing image data, ecological environment driving factor data, and carbon storage data:

[0034] ,

[0035] in, and Let X be the mean and standard deviation of the data X, respectively. This represents the standardized data;

[0036] Step 107: Construct the overall system data flow;

[0037] Input data at time t Represented as:

[0038] ,

[0039] in, This represents the current state of forest carbon storage.

[0040] Step 108: Define the overall calculation process;

[0041] In each time period The operation process is defined as follows:

[0042] ,

[0043] in, For carbon cycle mechanism model functions; This is a carbon state sensing function based on a lightweight Swin Transformer; Generate adversarial network functions for error compensation; Carbon change calculated using an ecological model; This is based on the lightweight Swing Transformer model to perceive the carbon state; This is the amount of error compensation corrected by a Generative Adversarial Network (GAN). This represents the carbon storage at the next moment.

[0044] Step 2 specifically includes the following steps:

[0045] Step 201: Geometric correction of remote sensing images;

[0046] raw remote sensing images Geometric correction is performed to unify the image coordinates to the standard geographic reference coordinate system. Mathematically, this is expressed as:

[0047] ,

[0048] in, Represents geometric transformation functions; To complete the geometrically corrected remote sensing image data; Indicates image coordinates;

[0049] Step 202: Radiation calibration and atmospheric correction;

[0050] Convert the digital quantization value DN of remote sensing imagery into the actual surface reflectance value R:

[0051] ,

[0052] Where L is the radiance; d is the Earth-Sun distance; It is the solar constant; The solar zenith angle is used. Atmospheric scattering and absorption effects are removed using an atmospheric correction model to obtain a true surface reflectance image. ;

[0053] Step 203: Unify the time scale of multi-source remote sensing data;

[0054] For data with inconsistent time periods, a linear interpolation method is used:

[0055] ,

[0056] in, , Obtain time points for two adjacent images; Standardized remote sensing image data for the target time point t; The earlier of two adjacent actual observation time points The original remote sensing image data; The later of two adjacent actual observation time points The original remote sensing image data;

[0057] Step 204: Removal of clouds and abnormal pixels;

[0058] Construct the cloud mask function M(x,y):

[0059] ,

[0060] Invalid pixels are removed using the following formula:

[0061] ;

[0062] in, The effective remote sensing image data at time t after removing clouds, fog, and abnormal pixels; The image data of the true surface reflectance at time t after radiometric calibration and atmospheric correction;

[0063] Step 205: Construct spectral indices;

[0064] Within each grid cell, the following vegetation indices are calculated to construct a set of spectral features reflecting the vegetation growth status:

[0065] Normalized Difference Vegetation Index (NDVI):

[0066] ,

[0067] Enhanced Vegetation Index (EVI):

[0068] ,

[0069] Soil-modifying vegetation index (SAVI):

[0070] ,

[0071] Where R, NIR, and B represent the red light band, near-infrared band, and blue light band, respectively; L is the soil brightness adjustment parameter;

[0072] Step 206: Construct texture features;

[0073] Extracting image texture features using the gray-level co-occurrence matrix:

[0074] ,

[0075] ,

[0076] ,

[0077] Where P(i,j) is the element in the i-th row and j-th column of the gray-level co-occurrence matrix; This refers to the degree of difference in grayscale values ​​between adjacent pixels in a remotely sensed image. This refers to the uniformity of grayscale distribution and the fineness of texture in remote sensing images; It refers to the similarity of gray values ​​between adjacent pixels in a remote sensing image;

[0078] Step 207: Construct a multi-feature fusion vector;

[0079] By fusing spectral indices, texture features, and original band data, a unified feature vector is formed:

[0080] ,

[0081] in, Let be the comprehensive feature vector of the i-th grid cell at time t;

[0082] Step 208: Feature normalization processing;

[0083] Normalize the eigenvectors:

[0084] ,

[0085] in, The maximum value of the comprehensive feature vector set at time t. Let be the minimum value of the set of integrated feature vectors at time t;

[0086] Map all feature components to the interval [0,1]; This represents the normalized composite feature vector of the i-th raster cell at time t;

[0087] Step 209: Constructing an initial characterization of the carbon state;

[0088] By using linear mapping or shallow neural networks, the comprehensive feature vector is mapped to the initial carbon state estimate. :

[0089] ,

[0090] Where W and b are trainable parameters;

[0091] Step 210: Construct a remote sensing carbon state feature dataset :

[0092] .

[0093] Step 3 specifically includes the following steps:

[0094] Step 301: Construct the input feature tensor;

[0095] eigenvectors Reconstructed into a multi-channel image tensor form based on spatial location. Where H and W represent the image height and width, respectively; C represents the number of feature channels; and U represents the real number space.

[0096] Step 302: Patch Embedding and Linear Mapping;

[0097] The input image is divided into several image patches of size P×P. Each image patch is flattened into a vector and an initial feature representation is obtained through linear mapping.

[0098] ,

[0099] Where P is the side length of the image patch. and These are trainable parameters; It represents the feature embedding representation at the patch level.

[0100] Step 303: Construct a window partitioning and window self-attention mechanism;

[0101] The feature map is divided into local windows of size M×M, and multi-head self-attention computation is performed within each window, mathematically expressed as:

[0102] ,

[0103] Where M is the window side length parameter, used to divide the input feature map into local windows of size M×M, limiting the local scope of the self-attention calculation; Attention(Q,K,V) is the multi-head self-attention calculation function, which mines the spatial correlation information between features through the interaction of query vector Q, key vector K, and value vector V. Let T denote the normalized exponential function, T denote the transpose, and Q, K, and V be obtained from the input features through linear transformations.

[0104] , , ,

[0105] in , , The weight matrix is ​​trainable.

[0106] Step 304: Window panning mechanism;

[0107] In adjacent Transformer layers, the window is translated by half the window size, enabling information exchange between different windows. Mathematically, this can be represented as:

[0108] ,

[0109] in, This is the feature map after window translation, which is the output of the input feature map after processing by the Shift window translation operation function; Shift is the window translation operation function, used to translate a local window on the feature map by a set distance along the x-axis and y-axis, with intermediate parameters... This enables information interaction between different windows in adjacent Transformer layers, balancing the accuracy of local feature modeling with the ability to fuse global information.

[0110] Step 305: Construct a multi-scale feature pyramid by downsampling layer by layer and channel expansion;

[0111] The model constructs a multi-scale feature representation system through layer-by-layer downsampling and feature channel expansion: the lower layer focuses on depicting image details and local textures; the middle layer focuses on depicting forest stand structure and vegetation distribution; and the upper layer focuses on depicting the spatial pattern of large-scale carbon states. Through this hierarchical structure, the model can simultaneously possess a comprehensive perception capability of both details and the global picture.

[0112] Step 306: Design lightweight structural constraints, including:

[0113] Channel trimming:

[0114] The importance of intermediate feature channels used for carbon state representation in the Transformer network is evaluated, and only feature channels with high contribution to carbon state perception are retained. "High contribution" in this embodiment can be determined in any of the following ways: based on the correlation coefficient between the channel feature and the target carbon state estimate; based on the average response intensity of the absolute value of the channel feature gradient; based on the weight update magnitude of the channel feature during training. In a specific embodiment, after sorting by channel contribution, only the top 50% to 70% of feature channels are retained, and the remaining channels are pruned to reduce redundant computation while maintaining the main carbon state information.

[0115] Attention head compression constraint:

[0116] Attention head compression: The number of attention heads in a multi-head self-attention structure is constrained to reduce the computational overhead of multi-head attention. In this embodiment, the compression of the number of attention heads follows the following principles: when multiple attention heads produce similar responses to changes in the carbon state in the same region, they are considered redundant. Redundant attention heads are merged or pruned, and only attention heads with strong ability to characterize spatial carbon heterogeneity are retained.

[0117] In one exemplary implementation, the original 8-12 attention heads can be compressed to 4-6 attention heads, significantly reducing the computational load of the model while maintaining spatial perception capabilities.

[0118] Parameter sharing constraints:

[0119] Parameter sharing: A parameter sharing mechanism is introduced between different layers of the Transformer network, allowing multiple layers to share some weight parameters to reduce the overall parameter size of the model. In specific embodiments, adjacent Transformer layers may share at least one of the following: linear mapping parameters in the multi-head attention module; or a portion of the weight matrix in the feedforward network.

[0120] By sharing parameters, the storage and training costs of the model can be reduced without significantly affecting its expressive power.

[0121] Quantization calculation constraints:

[0122] Quantization calculation: Low-bit-width quantization is performed on the model weight parameters and intermediate features to further reduce storage and computational complexity.

[0123] In this embodiment, the model weights and intermediate features can be quantized from floating-point representation to 8-bit or 16-bit fixed-point representation, ensuring that the carbon state estimation error is within an acceptable range and achieving lightweight model deployment.

[0124] Step 307: Construct the apparent output layer of carbon state;

[0125] At the end of the Transformer network after completing the lightweight structural constraints, a carbon state appearance output layer is constructed to map the high-dimensional features output by the network to the carbon state appearance estimates of the corresponding spatiotemporal units. In this embodiment, to reduce the computational complexity and parameter size of the output layer, a fully connected structure with low-dimensional mapping is used to compress the high-dimensional features. That is, the feature vector output by the Transformer network is first compressed in dimension before the carbon state appearance estimation is completed. For example, in an exemplary implementation, the feature vector output by the Transformer network can be 256-dimensional or 512-dimensional. By setting the output layer to map it to a 1-dimensional or low-dimensional carbon state appearance estimate, the number of parameters in the output layer is significantly reduced. Specifically, the carbon state appearance estimate is calculated as follows:

[0126] ,

[0127] in, Let be the final feature vector of the i-th grid cell; W and b are the trainable parameters of the output layer.

[0128] By using the aforementioned low-dimensional mapping method, the output layer parameter size and computational load can be effectively reduced while ensuring that the apparent estimation error of carbon state is within an acceptable range, thus further meeting the requirements for lightweight model deployment.

[0129] Step 308: Training rules for the carbon state-aware model;

[0130] In this embodiment, the carbon state perception model is a carbon state appearance estimation model built on a lightweight Transformer structure, which is used to perceive and characterize the spatiotemporal variation characteristics of forest carbon state from multi-source remote sensing features. The training of the carbon state perception model adopts a combination of supervised learning and weakly supervised learning, wherein: the supervision signal comes from carbon storage or biomass observation data obtained from sample plot surveys; the weak supervision signal comes from carbon change trend constraints output by the mechanistic model and historical carbon storage statistics.

[0131] By introducing a training method that combines supervised and weakly supervised training, under limited experimental sample conditions, the model is guided to learn carbon state representation results that conform to the carbon cycle mechanism, thereby improving the stability and generalization ability of the model training.

[0132] Step 309: Spatial continuity and physical rationality constraints;

[0133] Spatial continuity and physical plausibility constraints are imposed on the output of the carbon state perception model. The constraint rules include:

[0134] Spatial continuity constraint: The carbon state change amplitude of adjacent grid cells at the same time should not change abruptly. In an exemplary embodiment, when the difference in carbon state change of adjacent grid cells exceeds a preset threshold, it is determined that the spatial continuity constraint is not met.

[0135] Reasonableness constraint of change direction: The direction of carbon state change should be consistent with the vegetation growth, degradation or disturbance process. When the direction of carbon state change output by the model is opposite to the trend of the corresponding vegetation index change, it is judged as not meeting the physical reasonableness constraint.

[0136] Ecological rationality constraint on output amplitude: The carbon state change amplitude output by the model should be within the preset ecological rationality range. In an exemplary implementation, when the carbon state change exceeds the historical statistical range or the reasonable range given by the mechanism model within a unit time, it is judged as an abnormal output.

[0137] If the model output violates any constraint rule, the model parameters are backtracked or a retraining process is triggered to ensure the physical rationality and spatial consistency of the model output.

[0138] Step 310: The model ultimately outputs the following two types of results:

[0139] apparent carbon state estimate ;

[0140] High-dimensional carbon state eigenvectors ;

[0141] in, Used to construct mechanistic error terms; high-dimensional carbon state feature vectors ) is used as input for GAN error compensation models.

[0142] In step 4, the carbon cycle mechanism model uses net ecosystem productivity (NEP) as the core variable, and its basic expression is:

[0143] ,

[0144] in, For total primary productivity; For the plant's autotrophic respiration; For soil heterotrophic respiration; NEP(t) characterizes the net carbon uptake capacity of the forest at time t.

[0145] Step 4 specifically includes the following steps:

[0146] Step 401: Constructing the set of input variables required for a carbon cycle mechanism model :

[0147] ,

[0148] Where PAR represents photosynthetically active radiation, Temp represents temperature, Prec represents precipitation, SM represents soil moisture, B represents vegetation type and initial biomass, and SOC represents soil organic carbon content.

[0149] Step 402: Construct the rules for calculating Total Primary Productivity (GPP);

[0150] The following methods are used to estimate the total primary productivity (GPP) of technology:

[0151] GPP = Light energy input × Vegetation absorption capacity × Environmental stress regulation coefficient;

[0152] Step 403: Autotrophic Respiration Construction rules: Completed through table lookup or model parameter library. Engineering calculations.

[0153] Step 404: Heterotrophic respiration Construction rules: The higher the soil organic carbon content, the greater the amount of carbon released during decomposition; the higher the temperature, the faster the decomposition rate; both excessively low and excessively high soil moisture will inhibit heterotrophic respiration intensity; in specific calculations, heterotrophic respiration intensity is positively correlated with soil organic carbon content; when the soil temperature is in the range of 10~30 ℃, the heterotrophic respiration rate increases approximately exponentially with increasing temperature; when the soil moisture content is lower than 40% or higher than 90% of field capacity, the heterotrophic respiration intensity decreases proportionally, which is used to characterize the inhibitory effect of drought stress or hypoxia.

[0154] Step 405: Regional parameter adaptive mechanism; To address the differences in ecological conditions across different regions, a regional correction coefficient is introduced to adaptively adjust the parameters of the mechanistic model. This regional correction coefficient includes at least three aspects: topography type, vegetation type, and water conditions. For mountainous, hilly, and plain areas, different sets of carbon cycle parameters are used to reflect the influence of topography on the carbon flux process. For coniferous, broadleaf, and mixed forest areas, different sets of physiological parameters are used to reflect differences in vegetation type. For arid and humid areas, different water response parameters are used to reflect the regulatory effect of water conditions on the carbon exchange process.

[0155] Step 406: Calculate the theoretical change in carbon;

[0156] The theoretical carbon change is calculated using the following formula. :

[0157] ;

[0158] in, For the net ecosystem productivity at time t, Indicates the corresponding time step;

[0159] Step 407: Spatial Raster Scale Expansion: Each cell or grid undergoes an independent mechanistic calculation; the spatial distribution form is obtained. ;

[0160] Step 408: Physical rationality constraint check;

[0161] Physical rationality constraints are imposed on the carbon change results output by the mechanistic model to avoid abnormal outputs that do not conform to the laws of the ecological carbon cycle. The physical rationality constraints include the following quantitative judgment rules:

[0162] Consistency constraint of change direction: the change in carbon per unit time The symbol is compared with the vegetation index change trend of the corresponding grid cell. When the carbon change is positive and the vegetation index shows a significant downward trend, or when the carbon change is negative and the vegetation index shows a significant upward trend, it is determined that the consistency constraint of change direction is not met.

[0163] Reasonableness constraint of change range: Compare the carbon change range with the reasonable change range given by historical statistical distribution or mechanistic model. If the range of changes exceeds the historical range of the corresponding region at the same time scale, it is determined that the range of reasonableness constraints are not met.

[0164] Temporal continuity constraint: When comparing carbon change results within adjacent time steps, if there is an abrupt change in the magnitude of carbon change within a continuous time scale that cannot be explained by known ecological disturbance events, it is determined that the temporal continuity constraint is not met.

[0165] When the model output violates any quantitative judgment rule, the mechanistic model parameters of the corresponding region are backtracked and corrected, or the region is marked as an abnormal region, in order to ensure the physical rationality and stability of the carbon change results.

[0166] Step 409: Mechanism Model Output Standardization: [This section appears to be incomplete and requires further context.] Perform scale normalization;

[0167] Step 410: Finally, generate the mechanistic model output dataset, including the following data: theoretical carbon change. Updated carbon storage It serves as the main input and physical benchmark for the GAN error compensation model.

[0168] In step 5, a Generative Adversarial Network (GAN) is introduced as the GAN error compensation model.

[0169] Step 5 specifically includes the following steps:

[0170] Step 501: Construct error samples;

[0171] During the model training phase, an error sample set is constructed using historical time-series data:

[0172] Calculating theoretical carbon reserves using a carbon cycle mechanism model ;

[0173] Obtaining the apparent carbon state using a lightweight Swin Transformer ;

[0174] To obtain the true carbon storage, combine sample site surveys or long-term monitoring data. ;

[0175] Construct training sample pairs for input: Output: ,in The high-dimensional carbon state feature vector output by the Swing Transformer; Let be the carbon change error term for the i-th spatial unit at time t; This represents the actual change in carbon storage in the i-th spatial unit during the period from time t to t+1.

[0176] Step 502: Generator input structure design;

[0177] The generator input consists of the following two parts:

[0178] The mechanism error characteristics section is used to reflect the systematic deviation trend of the mechanism model at different time scales;

[0179] The carbon state space characteristic part is used to reflect the structural differences in error in different spatial regions.

[0180] The two parts are concatenated and used as the unified input vector for the generator.

[0181] Step 503: Generator function definition: Given the current carbon state space characteristics and historical mechanism error characteristics, output a carbon storage error compensation amount to make the mechanism model prediction results closer to the actual observation results;

[0182] Step 504: Discriminator function definition;

[0183] The discriminator is used to determine whether the input error compensation result originates from the true error distribution. The input includes:

[0184] True error compensation amount Generator output error compensation amount ;

[0185] The discriminator makes a comprehensive judgment based on the error magnitude, error distribution characteristics, and degree of matching with spatial features, and outputs the probability value that the sample comes from the true error distribution.

[0186] Step 505: Construct an adversarial training mechanism;

[0187] During training, the generator and discriminator are optimized alternately through an adversarial process: the discriminator continuously improves its ability to distinguish between the real error and the generated error; the generator continuously improves its ability to approximate the real error distribution with its output.

[0188] Step 506: Mechanism Consistency Constraint Mechanism: A mechanism consistency constraint is applied to the GAN error compensation results, requiring that the error compensation direction be consistent with the actual carbon change direction, i.e., the error compensation result must not change the positive or negative trend of carbon storage change; at the same time, the error compensation magnitude must not exceed a certain proportion of the change predicted by the mechanism model, so as to avoid excessive interference of the data-driven model on the physical mechanism-dominated results; wherein, the "certain proportion" is set as the carbon change output by the mechanism model. 30% to 50% is used to limit the adjustment range of the error compensation results.

[0189] Step 507: Error compensation amplitude control rules;

[0190] Output results of GAN error compensation model An amplitude limit is applied, and when the error compensation amplitude exceeds the set proportional threshold of the change in the output of the mechanism model, automatic reduction or rollback processing is performed.

[0191] The error compensation range must meet the following constraints:

[0192] ,

[0193] When the constraints are not met, the error compensation result is scaled proportionally to the allowable range, or it is rolled back to the output result of the mechanism model, so as to ensure the stability and interpretability of the overall carbon change estimation result.

[0194] Step 508: GAN error compensation model inference and execution process;

[0195] For each time step t and spatial grid cell i, perform the following operation: Obtain the theoretical carbon change output by the mechanistic model. Obtain the high-dimensional carbon state feature vector output by the Swing Transformer network. and will and These features serve as input to the GAN error compensation model; the generator network outputs the error compensation amount. Subsequently, the mechanism consistency constraint check and error compensation magnitude control rules are executed sequentially, and the constrained compensation results are sent to the fusion inference module. The fusion inference module is a logical function module used to perform weighted fusion of the mechanism model output and the GAN error compensation results to generate the final carbon state change estimation result. The above inference process is executed independently at each time step to ensure the continuity and controllability of the error compensation results in the time series.

[0196] Step 509: Perform time smoothing on the output of the GAN error compensation model;

[0197] Step 510: Anomalous Disturbance Enhancement Mode;

[0198] When abnormal perturbation characteristics are detected in the carbon state estimation results, the system automatically enters the abnormal perturbation enhancement mode, appropriately increases the GAN error compensation weight, and marks the current region as abnormal. Specifically, when the error compensation amplitude exceeds twice the standard deviation of the historical average for three or more consecutive time steps, or when the carbon state change rate of adjacent spatial grids significantly exceeds the historical maximum change amplitude and is clearly inconsistent with historical evolution patterns, it is judged as an abnormal perturbation situation, used to characterize sudden impacts such as fires, logging, or extreme weather events.

[0199] Step 511: Output the error compensation result;

[0200] The final output error compensation result dataset includes the following data: mechanistic model error correction amount. Abnormal disturbance identification information and error confidence assessment indicators.

[0201] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0202] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0203] This invention addresses the core problems in traditional forest carbon sink monitoring, such as low resolution of remote sensing images, high computational complexity of models leading to poor real-time performance and insufficient cross-regional generalization. It proposes a collaborative optimization of a lightweight Swin Transformer model and GAN super-resolution reconstruction to solve the technical bottlenecks of high-precision forest classification and real-time dynamic carbon storage tracking. In particular, it addresses the deployment requirements of resource-constrained embedded devices (such as drones) and overcomes the key challenge of balancing lightweight models with performance.

[0204] The present invention has the following beneficial effects:

[0205] 1) The carbon cycle mechanism model is the primary model to ensure that the results meet the ecological and physical laws of the carbon cycle;

[0206] 2) Introduce intelligent models for carbon state perception and error compensation to improve the accuracy of carbon storage estimation;

[0207] 3) Achieve collaborative operation between the mechanistic model and the intelligent model, balancing stability and flexibility;

[0208] 4) The results are well interpretable, and the changes in carbon storage are composed of both mechanistic and compensatory terms;

[0209] 5) The system has a clear structure, strong engineering deployability, and is suitable for forestry carbon sink accounting and ecological supervision applications. Attached Figure Description

[0210] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0211] Figure 1 This is a training graph for a GAN model.

[0212] Figure 2This is a diagram of the method architecture of the present invention. Detailed Implementation

[0213] like Figure 2 As shown, this embodiment provides a method for dynamic monitoring of forest carbon sinks. Its overall process is constructed along the main lines of "data acquisition—mechanism modeling—intelligent sensing—error compensation—fusion reasoning—closed-loop correction," forming a dynamic collaborative system dominated by carbon cycle mechanisms and enhanced by deep learning models. The overall system operation flow is represented as follows:

[0214] ,

[0215] in, Let t be the forest carbon storage at time t; This represents the theoretical carbon change calculated based on a carbon cycle mechanism model. This refers to the error compensation term based on the output of the GAN network. This is the final prediction result of forest carbon storage at the next time step obtained through fusion reasoning.

[0216] This formula clarifies the core idea of ​​the method of this invention: the change in forest carbon storage is first given by the carbon cycle mechanism model under physical constraints, and then the complex nonlinear error part that is difficult to be characterized by the mechanism model is compensated and corrected by the GAN network, thereby improving the prediction accuracy while ensuring physical consistency.

[0217] In this embodiment, the method includes the following steps:

[0218] Step 1: Acquire multi-source remote sensing imagery, meteorological data, and basic ecological data; including the following steps:

[0219] Step 101: Construct forest region definition and spatial grid;

[0220] First, the target forest monitoring area is determined based on the administrative divisions or ecological zoning boundaries of the study area. The target area is divided into regular raster units with a uniform spatial resolution: , in, This represents the i-th grid cell; N is the total number of grid cells.

[0221] Each grid cell serves as an independent object for carbon storage calculation and updating, thereby enabling refined modeling of forest carbon storage in the spatial dimension.

[0222] Step 102: Collect multi-source remote sensing data;

[0223] At each time scale t, acquire multi-source remote sensing image data for the corresponding region. The remote sensing images include at least red band (R), near-infrared band (NIR), and short-wave infrared band (SWIR) data, and optionally radar images or lidar data as auxiliary input.

[0224] The remote sensing dataset is represented as follows:

[0225] ,

[0226] in, This indicates an optional auxiliary remote sensing data source.

[0227] Step 103: Obtain data on ecological and environmental driving factors;

[0228] Simultaneous collection of datasets of ecological and environmental driving factors closely related to forest carbon cycle processes. Including temperature data Precipitation data Photosynthetically active radiation Soil moisture and soil organic carbon content The unified expression is:

[0229] ,

[0230] The aforementioned environmental factors serve as important inputs to the carbon cycle mechanism model, used to characterize the physical processes of carbon absorption and release in forest ecosystems.

[0231] Step 104: Construct carbon storage data for ground samples;

[0232] Obtain measured carbon storage data from forest plots within a selected area. The data can be obtained through sample plot surveys, forest inventory data, or long-term ecological monitoring stations. Sample carbon storage is expressed as:

[0233] ,

[0234] in, Let be the biomass of the j-th vegetation type at time t; denoted as the corresponding biomass-to-carbon conversion coefficient; M represents the number of vegetation types. This sample data serves as an important basis for model training, calibration, and error feedback.

[0235] Step 105: Construct an initial carbon storage baseline;

[0236] Based on historical remote sensing data and sample plot survey results, a baseline map of forest carbon storage at the initial moment of the system was constructed:

[0237] ,

[0238] Where i represents the grid number; It is the initial carbon storage value of the i-th grid cell;

[0239] Step 106: Data standardization and unification processing

[0240] All remote sensing data, ecological environment data, and carbon storage sample data were standardized:

[0241] ,

[0242] in, and denoted as the mean and standard deviation of data X, respectively.

[0243] This step ensures consistency in numerical scale across different data sources, thereby improving the stability of model training.

[0244] Step 107: Construct the overall system data flow;

[0245] Therefore, the system's input data at time t can be uniformly expressed as:

[0246] ,

[0247] in, For remote sensing image data; As a driving factor of the ecological environment; This represents the current state of forest carbon storage.

[0248] This dataset serves as a unified input for subsequent carbon cycle mechanism models, Swin Transformer perception models, and GAN error compensation models.

[0249] Step 108: Define the overall system calculation process;

[0250] In this embodiment, the system in each time period The operation process is defined as follows:

[0251] ,

[0252] in, For carbon cycle mechanism model functions; This is a carbon state sensing function based on a lightweight Swin Transformer; Generate an adversarial network function for error compensation.

[0253] Step 2: Preprocess the remote sensing data to construct carbon state characteristics;

[0254] In this embodiment, to ensure that the remote sensing data accurately reflects the true physical state of the forest ecosystem, systematic preprocessing and feature construction of the multi-source remote sensing image data are required before entering the lightweight Swin Transformer network. The purpose of this step is to eliminate noise introduced by sensor differences, atmospheric interference, geometric distortion, and temporal scale differences in the remote sensing data, ensuring that the input data meets the dual requirements of data consistency and physical interpretability for subsequent deep learning models and carbon cycle mechanism models.

[0255] Step 2 specifically includes the following steps:

[0256] Step 201: Geometric correction of remote sensing images;

[0257] raw remote sensing images Geometric correction is performed to unify the image coordinate system to the standard geographic reference coordinate system, mathematically expressed as:

[0258] ,

[0259] in, Represents geometric transformation functions; To complete the geometric correction of the remote sensing image data.

[0260] Step 202: Radiation calibration and atmospheric correction;

[0261] Convert the digital quantization (DN) values ​​of remote sensing images into actual surface reflectance values:

[0262] ,

[0263] Where L is the radiance; d is the Earth-Sun distance; It is the solar constant; This is the solar zenith angle.

[0264] Atmospheric scattering and absorption effects are removed by using an atmospheric correction model to obtain a true surface reflectance image. .

[0265] Step 203: Unify the time scale of multi-source remote sensing data;

[0266] Because different remote sensing data sources have different acquisition periods, it is necessary to standardize the image timescale. For data with inconsistent timeframes, a linear interpolation method is used.

[0267] ,

[0268] in, , Time points are obtained for two adjacent images.

[0269] Step 204: Removal of clouds and abnormal pixels;

[0270] Construct the cloud mask function M(x,y):

[0271] ,

[0272] Invalid pixels are removed using the following formula:

[0273] ,

[0274] To ensure the reliability of subsequent calculation data.

[0275] Step 205: Construct spectral indices;

[0276] Multiple vegetation indices are calculated within each grid cell to construct a set of spectral features reflecting the vegetation growth status:

[0277] Normalized Difference Vegetation Index (NDVI):

[0278] ,

[0279] Enhanced Vegetation Index (EVI):

[0280] ,

[0281] Soil-modifying vegetation index (SAVI):

[0282] ,

[0283] Where R, NIR, and B represent the red light band, near-infrared band, and blue light band, respectively; L is the soil brightness adjustment parameter.

[0284] Step 206: Construct texture features;

[0285] Image texture features, such as contrast, correlation, energy, and homogeneity, are extracted using the Gray-Level Co-occurrence Matrix (GLCM).

[0286] ,

[0287] ,

[0288] ,

[0289] Where P(i,j) are elements of the gray-level co-occurrence matrix. Texture features are used to characterize the complexity of forest canopy structure and are closely related to forest biomass and carbon storage.

[0290] Step 207: Construct a multi-feature fusion vector;

[0291] By fusing spectral indices, texture features, and original band data, a unified feature vector is formed:

[0292] ,

[0293] in, Let be the comprehensive feature vector of the i-th grid cell at time t.

[0294] Step 208: Feature normalization processing;

[0295] Normalize the eigenvectors:

[0296] ,

[0297] Mapping all feature components to the [0,1] interval improves the stability of network training.

[0298] Step 209: Constructing an initial characterization of the carbon state;

[0299] The integrated feature vector is mapped to an initial carbon state estimate using a linear mapping or a shallow neural network:

[0300] ,

[0301] Where W and b are trainable parameters.

[0302] This initial carbon state is used to provide monitoring signals and structural constraints for the Swing Transformer.

[0303] Step 210: Construct a remote sensing carbon state feature dataset;

[0304] The final dataset used for training the carbon state-aware model is as follows:

[0305] ,

[0306] This dataset contains spectral features, spatial structure features, and prior information on carbon state, providing high-quality input for the subsequent lightweight Swin Transformer model.

[0307] Step 3, Lightweight Swing Transformer carbon state-aware modeling;

[0308] In this embodiment, a lightweight Swing Transformer is used to construct a forest carbon state perception model. Its core objective is to significantly reduce the model parameter scale and computational complexity while ensuring high-precision modeling of the spatial structural features and multi-scale semantic information of remote sensing images. This allows the model to adapt to the long-term, continuous, and stable operation requirements of large-scale forest areas. The model does not directly output the final carbon storage value, but rather outputs an estimated value of the apparent carbon state characterizing the spatial distribution of forest carbon state, along with its corresponding high-dimensional feature vector. This provides perceptual input for subsequent carbon cycle mechanism models and GAN error compensation models. Step 3 specifically includes the following steps:

[0309] Step 301: Construct the input feature tensor;

[0310] eigenvectors Reconstructed into a multi-channel image tensor form based on its spatial location: Where H and W are the image height and width, respectively; C is the number of feature channels, corresponding to the combination of various spectral indices, texture features, and original band features.

[0311] Step 302: Patch Embedding and Linear Mapping;

[0312] The input image is divided into several image patches of size P×P. Each image patch is flattened into a vector and an initial feature representation is obtained through linear mapping.

[0313] ,

[0314] in, and These are trainable parameters; It represents the feature embedding representation at the patch level.

[0315] Step 303: Construct a window partitioning and window self-attention mechanism;

[0316] The feature map is divided into local windows of size M×M, and multi-head self-attention computation is performed within each window. Its mathematical expression is as follows:

[0317] ,

[0318] Wherein, Q, K, and V are obtained from the input features through linear transformations:

[0319] , , ,

[0320] in , , This is a trainable weight matrix.

[0321] Step 304: Shifted Window Mechanism;

[0322] In adjacent Transformer layers, the window is translated by half the window size, enabling information exchange between different windows. Its mathematical representation is as follows:

[0323] ,

[0324] in, ;

[0325] This approach achieves an organic unity between local modeling and global perception capabilities.

[0326] Step 305: Construct a hierarchical feature pyramid;

[0327] By constructing a multi-scale feature pyramid through layer-by-layer downsampling and channel expansion, the model can simultaneously characterize fine-grained texture features and high-level semantic features, thereby improving its ability to express differences in forest canopy structure and biomass distribution.

[0328] Step 306: Design lightweight structural constraints;

[0329] To reduce the computational load and storage requirements of the model, the following engineering strategies are introduced in the structural design:

[0330] Channel pruning: Reduce the number of redundant feature channels and retain only those feature channels that contribute highly to carbon state perception;

[0331] Attention head compression: reducing the number of attention heads in a multi-head attention structure;

[0332] Parameter sharing: Sharing some weight parameters across different layers;

[0333] Quantization calculation: Low-bit-width quantization is performed on the model weights and intermediate features.

[0334] Through the above measures, the model maintains both sufficient expressive power and good engineering deployability.

[0335] Step 307: Construct the apparent output layer of carbon state;

[0336] The high-dimensional features output by the Transformer network are converted into carbon state appearance estimates through a fully connected mapping layer:

[0337] ,

[0338] in, is the final feature vector of the i-th grid cell; W and b are the trainable parameters of the output layer.

[0339] Step 308: Training Rules for the Carbon State-Aware Model

[0340] The model training adopts a combination of supervised and weakly supervised methods: the supervision signal comes from the carbon storage data of the sample site survey and the output results of the mechanism model; the model output is constrained to be consistent with the direction of change of the actual carbon storage on the numerical scale; and the model output results are prevented from fluctuating drastically in the time dimension.

[0341] The following objectives will be the focus of the training process:

[0342] Improve the ability to characterize the spatial distribution of carbon states; enhance the sensitivity to differences in forest stand structure; and ensure that the model output results do not conflict with ecological mechanisms.

[0343] Step 309: Spatial continuity and physical rationality constraints;

[0344] The following constraints are imposed on the model output: the carbon state changes of adjacent grid cells should have spatial continuity; the direction of output value changes should be consistent with the vegetation growth or degradation process; and the output amplitude should be within a reasonable ecological range. If the above constraints are violated, the model parameters should be rolled back or retrained.

[0345] Step 310: Output the carbon state feature vector;

[0346] The model ultimately outputs two types of results:

[0347] apparent carbon state estimate ;

[0348] High-dimensional carbon state eigenvectors ;

[0349] in, Used to construct mechanistic error terms; ) is used as input for GAN error compensation models.

[0350] Step 4: Construct a carbon cycle mechanism model and calculate the theoretical carbon change;

[0351] In this embodiment, to ensure that the dynamic monitoring results of forest carbon storage have clear ecological significance and physical interpretability, a forest carbon cycle mechanism model is used as the dominant physical module of the entire system. This mechanism model is used to characterize the intrinsic mechanisms of carbon absorption, carbon allocation, carbon respiration, and carbon release processes in the forest ecosystem, and provides a theoretical basis for subsequent GAN error compensation and fusion inference. The carbon cycle mechanism model uses net ecosystem productivity (NEP) as the core variable, and its basic expression is:

[0352] ,

[0353] in, For total primary productivity; For the plant's autotrophic respiration; For soil heterotrophic respiration; NEP(t) characterizes the net carbon uptake capacity of the forest at time t.

[0354] Step 4 specifically includes the following steps:

[0355] Step 401: Construct the input variables for the mechanistic model;

[0356] The set of input variables required to construct a carbon cycle mechanism model:

[0357] Photosynthetically active radiation (PAR), temperature (Temp), precipitation (Prec), soil moisture (SM), vegetation type and initial biomass (B), and soil organic carbon (SOC);

[0358] Variables are uniformly denoted as:

[0359] ,

[0360] These variables are all derived from remote sensing data inversion results, meteorological station data, and historical ecological survey data.

[0361] Step 402: Construct GPP calculation rules;

[0362] In this embodiment, the calculation of GPP no longer uses complex derivation formulas, but is implemented in an engineering manner according to the following physical logic: calculate vegetation indices (NDVI, EVI, etc.) based on remote sensing images to invert the vegetation's ability to absorb light energy; calculate the effective light energy that the vegetation can obtain based on PAR data; and construct a light energy utilization efficiency adjustment coefficient through temperature and moisture factors.

[0363] Multiplying the above three factors together, we obtain the engineering estimate of GPP:

[0364] GPP = Light energy input × Vegetation absorption capacity × Environmental stress regulation coefficient;

[0365] Step 403: Autotrophic Respiration Construction rules;

[0366] Autotrophic respiration is primarily determined by both vegetation biomass and ambient temperature. The following rules apply in implementation: autotrophic respiration is positively correlated with vegetation biomass; it exhibits an exponential or piecewise linear positive correlation with ambient temperature; and different respiration coefficient parameters can be configured for different vegetation types. The system completes this process through table lookups or model parameter databases. Engineering calculations.

[0367] Step 404: Heterotrophic respiration The system constructs a heterotrophic respiration calculation module based on the following rules: Heterotrophic respiration is primarily influenced by soil organic carbon content, temperature, and soil moisture. The higher the soil organic carbon content, the greater the amount of carbon released during decomposition; the higher the temperature, the faster the decomposition rate; and both excessively low and high soil moisture inhibit heterotrophic respiration intensity.

[0368] Step 405: Regional parameter adaptive mechanism; To address the differences in ecological conditions across different regions, a regional correction coefficient is introduced: different parameter sets are used for mountains, hills, and plains; different parameter sets are used for coniferous forests, broad-leaved forests, and mixed forests; and different water response parameters are used for arid and humid regions. This approach enables the mechanistic model to have regional generalization capabilities.

[0369] Step 406: Calculate the theoretical change in carbon;

[0370] After completing GPP, After calculation, proceed according to the core formula:

[0371] ,

[0372] This serves as the theoretical output of the mechanistic model for changes in forest carbon storage.

[0373] Step 407: Spatial grid scale expansion;

[0374] Extending the above calculation process to each grid cell: each cell or grid independently performs a mechanism calculation; resulting in the spatial distribution form. ;

[0375] Step 408: Physical rationality constraint check;

[0376] The following rules are applied to the mechanism output results: the direction of carbon storage change should conform to the vegetation growth cycle; the magnitude of carbon change should be within the reasonable range of historical statistics; if it exceeds the range, parameter correction or anomaly marking will be triggered.

[0377] Step 409: Standardize the output of the mechanistic model;

[0378] right Scale normalization is performed to adapt it to the input of subsequent GAN models.

[0379] Step S410: Output the mechanism model data;

[0380] The final mechanistic model output dataset is as follows:

[0381] Theoretical carbon change Updated carbon storage It serves as the main input and physical benchmark for the GAN error compensation model.

[0382] Step 5, GAN error compensation model training and inference;

[0383] In this embodiment, to address the systematic biases of the carbon cycle mechanism model under conditions such as complex terrain, mixed forests with multiple species, extreme climate disturbances, and remote sensing noise, a Generative Adversarial Network (GAN) is introduced as an error compensation module. It is important to emphasize that in this embodiment, the GAN does not directly predict forest carbon storage; instead, it only models and compensates for the "error term" between the output of the carbon cycle mechanism model and the actual carbon storage, ensuring that it always operates within the physical constraints of the mechanism model, thereby guaranteeing the ecological rationality and physical consistency of the prediction results. Step 5 specifically includes the following steps:

[0384] Step 501: Construct error samples;

[0385] During the model training phase, an error sample set is constructed using historical time-series data:

[0386] Theoretical carbon storage was calculated using a carbon cycle mechanism model: [Results] ;

[0387] The apparent carbon state was obtained using a lightweight Swin Transformer: ;

[0388] To obtain the true carbon storage by combining sample site surveys or long-term monitoring data: ;

[0389] The training sample pairs are constructed using the three types of data mentioned above as input: Output: ,in This is the high-dimensional carbon state feature vector output by the Swing Transformer.

[0390] Step 502: Generator input structure design;

[0391] The generator input consists of two parts:

[0392] The mechanism error characteristics section is used to reflect the systematic deviation trend of the mechanism model at different time scales;

[0393] The carbon state space feature is used to reflect the structural differences in error across different spatial regions.

[0394] The two are concatenated and used as a unified input vector for the generator, enabling GAN to have both time-aware and space-aware capabilities.

[0395] Step 503: Define generator functionality;

[0396] The generator's function can be described as follows: given the current carbon state space characteristics and historical mechanistic error characteristics, it outputs a reasonable carbon storage error compensation amount, making the mechanistic model prediction results closer to the actual observation results. The generator is not allowed to output arbitrary correction amounts; its output is limited to the local correction range of the mechanistic model prediction results.

[0397] Step 504: Discriminator function definition;

[0398] The discriminator is used to determine whether the input error compensation result originates from the true error distribution. Its input includes:

[0399] True error compensation amount Generator output error compensation amount ;

[0400] The discriminator comprehensively judges the error magnitude, error distribution characteristics, and the degree of matching with spatial features, and outputs the probability value that the sample comes from the true error distribution.

[0401] Step 505: Construction of the adversarial training mechanism;

[0402] During training, the generator and discriminator are optimized alternately through adversarial training: the discriminator continuously improves its ability to distinguish between "true error" and "generated error"; the generator continuously improves its ability to approximate the true error distribution with its output. Through adversarial training, the error compensation amount output by the generator gradually acquires the statistical characteristics of the true error.

[0403] Step 506: Mechanism consistency constraint mechanism;

[0404] To prevent GANs from undermining the physical dominance of the carbon cycle mechanism model, the following constraint rules are introduced into the system:

[0405] The direction of error compensation should be consistent with the actual direction of carbon change; the magnitude of error compensation should not exceed a certain proportion of the change in the mechanism model; and the result after error compensation should satisfy the basic physical laws of the carbon cycle mechanism. Through rule constraints, it is ensured that GANs always operate within a "physically reasonable range".

[0406] Step 507: Error compensation amplitude control rules;

[0407] The system imposes amplitude limits on the GAN output: when Exceed When setting a certain percentage threshold, the system automatically reduces or reverts to its previous state. This percentage threshold can be dynamically configured based on the regional ecological stability, thereby achieving an adaptive strategy of "small compensation in stable areas and strong compensation in disturbed areas".

[0408] Step 508: GAN inference process;

[0409] During the real-time operation phase of the system, the following is executed for each time step and spatial grid cell:

[0410] Obtain the output of the mechanism model ; Obtain the output features of the Swing Transformer Construct the input features for the GAN; output the error compensation amount from the generator. Perform physical constraint checks and amplitude constraints; send the results to the fusion inference module.

[0411] Step 509: Stabilize the GAN output results;

[0412] To prevent excessive fluctuations in error compensation over a short period of time, time smoothing is performed on the GAN output results, so that the error compensation changes continuously over time, which is consistent with the gradual nature of carbon change in the ecosystem.

[0413] Step 510: Anomalous Disturbance Enhancement Mode;

[0414] When the system detects the following:

[0415] The error compensation amplitude increased significantly over multiple consecutive periods; the spatial distribution of carbon states showed abrupt changes; and it clearly deviated from historical statistical patterns.

[0416] If the system determines that the current area may be affected by fire, logging, or extreme weather disturbances, it will automatically enter the "abnormal disturbance enhancement mode", appropriately increase the GAN compensation weight, and mark the area as abnormal.

[0417] Step 511: Output the error compensation result;

[0418] The final output error compensation result dataset includes: mechanistic model error correction amount. Abnormal disturbance identification information; error confidence assessment index. In this embodiment, the error compensation result is directly fused with the calculation result of the mechanistic model to complete the final update and output of forest carbon storage.

[0419] To verify the advantages of the error compensation model adopted in this invention in terms of computational efficiency and operational stability, this embodiment compares and analyzes the performance of different generative adversarial network structures in the forest carbon sink dynamic monitoring task, as shown in Table 1. Table 1 compares the differences in computational speed and stability of three models: ResBaGAN based on deep residual networks, traditional DCGAN, and TimeSwin-GAN adopted in this invention. Among them, ResBaGAN, due to the introduction of a deep ResNet structure, has a large number of model parameters and a heavy computational burden, and is prone to mode collapse in small-sample remote sensing data scenarios; DCGAN has a relatively simple structure and fast computation speed, but lacks structural optimization for the forest carbon sink dynamic monitoring task, and its stability is greatly affected by the scale of training data and the choice of network structure. In contrast, the TimeSwin-GAN proposed in this invention, by introducing a lightweight Swing Transformer structure, significantly reduces the number of parameters while ensuring the model's expressive power, reducing the number of parameters by about 40%–60%, and improving the overall computational efficiency by about 3 times. Meanwhile, through a dynamic update mechanism and closed-loop control strategy, the periodic updating of historical data and adaptive adjustment of model parameters are achieved, thereby ensuring the stability and robustness of the model during long-term operation.

[0420] like Figure 1The diagram shows the training structure of the Generative Adversarial Network (GAN) in this embodiment. In the diagram, "Generative Adversarial Network" represents the overall structure of the GAN; "G" represents the Generator module; "D" represents the Discriminator module; "Latent Space" represents the latent feature space; "Noise" represents a random noise vector; "Real Samples" represents real error samples; "Generated FakeSamples" represents error-compensated samples output by the generator; "Is D Correct?" indicates the discriminator's judgment of the authenticity of the input samples; "Fine Tune Training" represents the adversarial optimization training process between the generator and the discriminator; arrows indicate data flow, and the circular "⊕" symbol represents feature concatenation or vector fusion operations. This GAN includes two core modules: a generator and a discriminator. The generator generates error-compensated samples given a noise vector and carbon state features as input, while the discriminator determines whether the input error samples originate from the real error distribution. In this embodiment, the generator's input includes the theoretical carbon change output by the mechanistic model, the high-dimensional carbon state feature vector output by the lightweight Swin Transformer, and a random noise vector; the generator's output is the carbon storage error compensation. During training, real error samples and error compensation samples output by the generator are jointly input into the discriminator for discrimination. The discriminator outputs the probability value of a sample belonging to the real error distribution; this discrimination result is fed back to the generator through a backpropagation mechanism to guide the generator to continuously optimize its output. Through the above adversarial training mechanism, the generator gradually learns the statistical distribution characteristics of the mechanistic model's prediction error, thereby generating an error compensation term that better conforms to the actual carbon storage change pattern. Through this generative adversarial network training process, adaptive learning and dynamic correction of the carbon cycle mechanism model's prediction error are achieved. While ensuring the dominance of the physical constraints of the mechanistic model, the accuracy and stability of the carbon storage change estimation results are improved, providing reliable error compensation input for subsequent carbon storage fusion inference. Figure 2The diagram shows the overall architecture of the forest carbon sequestration dynamic monitoring method of this invention. The method, from top to bottom, includes a data acquisition and feature perception stage, a model construction and theoretical calculation stage, and a model training and error compensation stage. These stages form a closed-loop collaborative workflow through data flow and feedback mechanisms. In the data acquisition and feature perception stage, the system first completes the determination of forest areas and the construction of spatial grids, collects multi-source remote sensing image data, ecological environment driving factor data, and ground sample carbon storage data, and performs preprocessing operations such as geometric correction, radiometric calibration, atmospheric correction, and cloud and anomalous pixel removal on the above data. Further, spectral indices, texture features, and multi-feature fusion vectors are constructed to form initial carbon state characterization data. In the model construction and theoretical calculation stage, a carbon state perception model is constructed based on the lightweight Swin Transformer to extract the spatial distribution characteristics of forest carbon state; simultaneously, a carbon cycle mechanism model is constructed to calculate the theoretical change in forest carbon storage, and physical rationality constraints are applied to the model output results. In the model training and error compensation phases, an error compensation model is constructed based on a generative adversarial network (GAN). Through adversarial training, the systematic deviation between the predicted results of the mechanistic model and actual observations is learned. The error compensation results are then fused with the output of the mechanistic model for inference, ultimately yielding a dynamic update of forest carbon storage. This architectural design enables the collaborative work of the mechanistic model and the data-driven model, improving the accuracy and stability of dynamic monitoring results for forest carbon sequestration.

[0421] Table 1

[0422]

[0423] This invention provides a method for dynamic monitoring of forest carbon sequestration. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for dynamic monitoring of forest carbon sequestration, characterized in that, Includes the following steps: Step 1: Acquire multi-source remote sensing image data, meteorological data, and basic ecological data; Step 2: Preprocess the remote sensing data to construct carbon state characteristics; Step 3, Lightweight Swing Transformer carbon state-aware modeling; Step 4: Construct a carbon cycle mechanism model and calculate the theoretical carbon change; Step 5: Training and inference of the GAN error compensation model.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 101: Construct forest region definition and spatial grid; First, the target forest monitoring area is determined based on the administrative divisions or ecological zoning boundaries of the study area. The target area is divided into regular raster units with a uniform spatial resolution: ， in, This represents the i-th grid cell; N is the total number of grid cells; each grid cell is used as an independent object for carbon storage calculation and updating. Step 102: Collect multi-source remote sensing data; At each time scale t, acquire multi-source remote sensing image data for the corresponding region. The remote sensing image data includes red band R data. Near-infrared (NIR) data SWIR data in the shortwave infrared band At the same time, radar imagery or lidar data are introduced as auxiliary input data; Multi-source remote sensing image data Represented as: ， in, This represents auxiliary input data; Step 103: Obtain data on ecological and environmental driving factors; Simultaneous collection of ecological and environmental driving factor datasets related to forest carbon cycle processes Including temperature data Precipitation data Photosynthetically active radiation Soil moisture and soil organic carbon content The unified expression is: ； Step 104: Construct carbon storage data for ground samples; Obtain measured forest plot carbon storage data , is represented as: ， in, Let be the biomass of the j-th vegetation type at time t; The biomass-to-carbon conversion coefficient; M is the number of vegetation types; Step 105: Construct an initial carbon storage baseline; Construct a baseline map of forest carbon storage at the initial stage of the system: ， Where i represents the grid number; It is the initial carbon storage value of the i-th grid cell; This represents the true initial carbon storage value of the i-th grid cell, obtained based on actual observations or inversion results. Step 106: Data standardization and unification processing; Standardization processing was performed on all multi-source remote sensing image data, ecological environment driving factor data, and carbon storage data: ， in, and Let X be the mean and standard deviation of the data X, respectively. This represents the standardized data; Step 107: Construct the overall system data flow; Input data at time t Represented as: ， in, This represents the current state of forest carbon storage. Step 108: Define the overall calculation process; In each time period The operation process is defined as follows: ， in, For carbon cycle mechanism model functions; This is a carbon state sensing function based on a lightweight Swin Transformer; Generate adversarial network functions for error compensation; Carbon change calculated using an ecological model; This is based on the lightweight Swing Transformer model to perceive the carbon state; This is the amount of error compensation corrected by a Generative Adversarial Network (GAN). This represents the carbon storage at the next moment.

3. The method according to claim 2, characterized in that, Step 2 specifically includes the following steps: Step 201: Geometric correction of remote sensing images; raw remote sensing images Geometric correction is performed to unify the image coordinates to the standard geographic reference coordinate system. Mathematically, this is expressed as: ， in, Represents geometric transformation functions; To complete the geometrically corrected remote sensing image data; Indicates image coordinates; Step 202: Radiation calibration and atmospheric correction; Convert the digital quantization value DN of remote sensing imagery into the actual surface reflectance value R: ， Where L is the radiance; d is the Earth-Sun distance; It is the solar constant; The solar zenith angle is used; atmospheric scattering and absorption effects are removed using an atmospheric correction model to obtain a true surface reflectance image. ; Step 203: Unify the time scale of multi-source remote sensing data; For data with inconsistent time periods, a linear interpolation method is used: ， in, , Obtain time points for two adjacent images; Standardized remote sensing image data for the target time point t; The earlier of two adjacent actual observation time points The original remote sensing image data; The later of two adjacent actual observation time points The original remote sensing image data; Step 204: Removal of clouds and abnormal pixels; Construct the cloud mask function M(x,y): ， Invalid pixels are removed using the following formula: ； in, The effective remote sensing image data at time t after removing clouds, fog, and abnormal pixels; The image data of the true surface reflectance at time t after radiometric calibration and atmospheric correction; Step 205: Construct spectral indices; Within each grid cell, the following vegetation indices are calculated to construct a set of spectral features reflecting the vegetation growth status: Normalized Difference Vegetation Index (NDVI): ， Enhanced Vegetation Index (EVI): ， Soil-modifying vegetation index (SAVI): ， Where R, NIR, and B represent the red light band, near-infrared band, and blue light band, respectively; L is the soil brightness adjustment parameter; Step 206: Construct texture features; Extracting image texture features using the gray-level co-occurrence matrix: ， ， ， Where P(i,j) is the element in the i-th row and j-th column of the gray-level co-occurrence matrix; This refers to the degree of difference in grayscale values ​​between adjacent pixels in a remotely sensed image. This refers to the uniformity of grayscale distribution and the fineness of texture in remote sensing images; It refers to the similarity of gray values ​​between adjacent pixels in a remote sensing image; Step 207: Construct a multi-feature fusion vector; By fusing spectral indices, texture features, and original band data, a unified feature vector is formed: ， in, Let be the comprehensive feature vector of the i-th grid cell at time t; Step 208: Feature normalization processing; Normalize the eigenvectors: ， in, The maximum value of the comprehensive feature vector set at time t. Let be the minimum value of the set of integrated feature vectors at time t; Map all feature components to the interval [0,1]; This represents the normalized composite feature vector of the i-th raster cell at time t; Step 209: Constructing an initial characterization of the carbon state; By using linear mapping or shallow neural networks, the comprehensive feature vector is mapped to the initial carbon state estimate. : ， Where W and b are trainable parameters; Step 210: Construct a remote sensing carbon state feature dataset : 。 4. The method according to claim 3, characterized in that, Step 3 specifically includes the following steps: Step 301: Construct the input feature tensor; eigenvectors Reconstructed into a multi-channel image tensor form based on spatial location. Where H and W represent the image height and width, respectively; C represents the number of feature channels; and U represents the real number space. Step 302: Patch Embedding and Linear Mapping; The input image is divided into several image patches of size P×P. Each image patch is flattened into a vector and an initial feature representation is obtained through linear mapping. ， Where P is the side length of the image patch. and These are trainable parameters; Patch-level feature embedding representation; Step 303: Construct a window partitioning and window self-attention mechanism; The feature map is divided into local windows of size M×M, and multi-head self-attention computation is performed within each window, mathematically expressed as: ， Where M is the window side length parameter, used to divide the input feature map into local windows of size M×M; Attention(Q,K,V) is the multi-head self-attention calculation function, which mines spatial correlation information between features through the interaction of query vector Q, key vector K, and value vector V. Let T denote the normalized exponential function, T denote the transpose, and Q, K, and V be obtained from the input features through linear transformations. , , ， in , , The weight matrix is ​​trainable. Step 304: Window panning mechanism; In adjacent Transformer layers, the window is translated by half the window size, enabling information exchange between different windows. Mathematically, this can be represented as: ， in, This is the feature map after the window has been translated; Shift is the window translation operation function, used to translate a local window on the feature map by a set distance along the x-axis and y-axis, with intermediate parameters... ; Step 305: Construct a multi-scale feature pyramid by downsampling layer by layer and channel expansion; Step 306: Design lightweight structural constraints, including: Channel pruning: The importance of intermediate feature channels used for carbon state representation in the Transformer network is evaluated, and only feature channels that contribute more to carbon state perception are retained. Attention head compression constraint: Constrains the number of attention heads in a multi-head self-attention structure; Parameter sharing constraint: Introduce a parameter sharing mechanism between different layers of the Transformer network; Quantization constraints: Low-bit-width quantization is performed on the model weight parameters and intermediate features; Step 307: Construct the apparent output layer of carbon state; At the end of the Transformer network after completing the lightweight structural constraints, a carbon state appearance output layer is constructed to map the high-dimensional features of the network output to the carbon state appearance estimate of the corresponding spatiotemporal unit. The carbon state appearance estimate is calculated as follows: ， in, Let be the final feature vector of the i-th grid cell; W and b are the trainable parameters of the output layer. Step 308: Training rules for the carbon state-aware model; The carbon state perception model is a carbon state appearance estimation model built on a lightweight Transformer structure, used to perceive and characterize the spatiotemporal variation characteristics of forest carbon state from multi-source remote sensing features. The training of the carbon state perception model adopts a combination of supervised learning and weakly supervised learning, wherein: the supervision signal comes from carbon storage or biomass observation data obtained from sample plot surveys; the weak supervision signal comes from carbon change trend constraints output by the mechanistic model and historical carbon storage statistics. Step 309: Spatial continuity and physical rationality constraints; Spatial continuity and physical plausibility constraints are imposed on the output of the carbon state perception model. The constraint rules include: Spatial continuity constraint: When the difference in carbon state change between adjacent grid cells exceeds a preset threshold, it is determined that the spatial continuity constraint is not met; Reasonableness constraint of change direction: When the direction of carbon state change output by the model is opposite to the trend of the corresponding vegetation index change, it is determined that the physical reasonableness constraint is not met. Output amplitude ecological rationality constraint: When the change in carbon state per unit time exceeds the historical statistical range or the reasonable range given by the mechanism model, it is judged as abnormal output; If the model output violates any constraint rule, the model parameters will be backtracked or a retraining process will be triggered. Step 310: The model ultimately outputs the following two types of results: apparent carbon state estimate ; High-dimensional carbon state eigenvectors ; in, Used to construct mechanistic error terms; high-dimensional carbon state feature vectors ) is used as input for GAN error compensation models.

5. The method according to claim 4, characterized in that, In step 4, the carbon cycle mechanism model uses net ecosystem productivity (NEP) as the core variable, and its basic expression is: ， in, For total primary productivity; For the plant's autotrophic respiration; For soil heterotrophic respiration; NEP(t) characterizes the net carbon uptake capacity of the forest at time t.

6. The method according to claim 5, characterized in that, Step 4 specifically includes the following steps: Step 401: Constructing the set of input variables required for a carbon cycle mechanism model : ， Where PAR represents photosynthetically active radiation, Temp represents temperature, Prec represents precipitation, SM represents soil moisture, B represents vegetation type and initial biomass, and SOC represents soil organic carbon content. Step 402: Construct the rules for calculating Total Primary Productivity (GPP); The following methods are used to estimate the total primary productivity (GPP) of technology: GPP = Light energy input × Vegetation absorption capacity × Environmental stress regulation coefficient; Step 403: Autotrophic Respiration Construction rules: Completed through table lookup or model parameter library. Engineering calculations; Step 404: Heterotrophic respiration Construction rules: Heterotrophic respiration intensity is positively correlated with soil organic carbon content; Step 405: Regional parameter adaptive mechanism; To address the differences in ecological conditions across different regions, a regional correction coefficient is introduced to adaptively adjust the parameters of the mechanistic model. This regional correction coefficient includes three aspects: topography type, vegetation type, and water conditions. For mountainous, hilly, and plain areas, different sets of carbon cycle parameters are used to reflect the influence of topography on the carbon flux process. For coniferous, broadleaf, and mixed forest areas, different sets of physiological parameters are used to reflect differences in vegetation type. For arid and humid areas, different water response parameters are used to reflect the regulatory effect of water conditions on the carbon exchange process. Step 406: Calculate the theoretical change in carbon; The theoretical carbon change is calculated using the following formula. : ； in, For the net ecosystem productivity at time t, Indicates the corresponding time step; Step 407: Spatial Raster Scale Expansion: Each cell or grid is independently subjected to a mechanistic calculation to obtain the spatial distribution form. ; Step 408: Physical rationality constraint check; Physical rationality constraints are imposed on the carbon change results output by the mechanistic model, and these constraints include the following quantitative judgment rules: Consistency constraint of change direction: the change in carbon per unit time The symbol is compared with the vegetation index change trend of the corresponding grid cell. When the carbon change is positive and the vegetation index shows a significant downward trend, or when the carbon change is negative and the vegetation index shows a significant upward trend, it is determined that the consistency constraint of change direction is not met. Reasonableness constraint of change range: Compare the carbon change range with the reasonable change range given by historical statistical distribution or mechanistic model. If the range of changes exceeds the historical range of the corresponding region at the same time scale, it is determined that the range of reasonableness constraints are not met. Temporal continuity constraint: When comparing carbon change results within adjacent time steps, if there is an abrupt change in the magnitude of carbon change within a continuous time scale that cannot be explained by known ecological disturbance events, it is determined that the temporal continuity constraint is not met. When the model output violates any quantization judgment rule, the mechanism model parameters of the corresponding region are backtracked and corrected, or the region is marked as an abnormal region. Step 409: Mechanism Model Output Standardization: [This section appears to be incomplete and requires further context.] Perform scale normalization; Step 410: Finally, generate the mechanistic model output dataset, including the following data: theoretical carbon change. Updated carbon storage It serves as the main input and physical benchmark for the GAN error compensation model.

7. The method according to claim 6, characterized in that, In step 5, a Generative Adversarial Network (GAN) is introduced as the GAN error compensation model.

8. The method according to claim 7, characterized in that, Step 5 specifically includes the following steps: Step 501: Construct error samples; During the model training phase, an error sample set is constructed using historical time-series data: Calculating theoretical carbon reserves using a carbon cycle mechanism model ; Obtaining the apparent carbon state using a lightweight Swin Transformer ; To obtain the true carbon storage, combine sample site surveys or long-term monitoring data. ; Construct training sample pairs for input: Output: ,in The high-dimensional carbon state feature vector output by the Swing Transformer; Let be the carbon change error term for the i-th spatial unit at time t; This represents the actual change in carbon storage in the i-th spatial unit during the period from time t to t+1. Step 502: Generator input structure design; The generator input consists of the following two parts: The mechanism error characteristics section is used to reflect the systematic deviation trend of the mechanism model at different time scales; The carbon state space characteristic part is used to reflect the structural differences in error in different spatial regions; The two parts are concatenated and used as the unified input vector for the generator. Step 503: Generator function definition: Given the current carbon state space characteristics and historical mechanism error characteristics, output a carbon storage error compensation amount to make the mechanism model prediction results closer to the actual observation results; Step 504: Discriminator function definition; The discriminator is used to determine whether the input error compensation result originates from the true error distribution. The input includes: True error compensation amount Generator output error compensation amount ; The discriminator makes a comprehensive judgment based on the error magnitude, error distribution characteristics, and degree of matching with spatial features, and outputs the probability value that the sample comes from the true error distribution. Step 505: Construct an adversarial training mechanism; During training, the generator and discriminator are optimized alternately through an adversarial process: the discriminator continuously improves its ability to distinguish between the real error and the generated error; the generator continuously improves its ability to approximate the real error distribution with its output. Step 506: Mechanism Consistency Constraint Mechanism: Apply mechanism consistency constraints to the GAN error compensation results, requiring that the error compensation direction be consistent with the actual carbon change direction, that is, the error compensation results must not change the positive or negative trend of carbon storage change; at the same time, the error compensation magnitude must not exceed a certain proportion of the change predicted by the mechanism model. Step 507: Error compensation amplitude control rules; Output results of GAN error compensation model An amplitude limit is applied, and when the error compensation amplitude exceeds the set proportional threshold of the change in the output of the mechanism model, automatic reduction or rollback processing is performed. The error compensation range must meet the following constraints: ， When the constraints are not met, the error compensation result is scaled proportionally to the allowable range, or it is rolled back to the output result of the mechanism model. Step 508: GAN error compensation model inference and execution process; For each time step t and spatial grid cell i, perform the following operation: Obtain the theoretical carbon change output by the mechanistic model. Obtain the high-dimensional carbon state feature vector output by the Swing Transformer network. and will and These features serve as input to the GAN error compensation model; the generator network outputs the error compensation amount. Then, the mechanism consistency constraint check and error compensation magnitude control rules are executed sequentially, and the constrained compensation results are sent to the fusion inference module; wherein, the fusion inference module is a logical function module used to perform weighted fusion of the mechanism model output and the GAN error compensation results to generate the final carbon state change estimation result; Step 509: Perform time smoothing on the output of the GAN error compensation model; Step 510: Anomalous Disturbance Enhancement Mode; When abnormal perturbation features are detected in the carbon state estimation results, the system automatically enters the abnormal perturbation enhancement mode, increases the GAN error compensation weight, and marks the current region as abnormal. Step 511: Output the error compensation result; The final output error compensation result dataset includes the following data: mechanistic model error correction amount. Abnormal disturbance identification information and error confidence assessment indicators.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images