A coastal blue carbon ecosystem carbon sink early warning method and system

Abstract

The present application relates to the field of environmental protection, especially a kind of coastal blue carbon ecosystem carbon sink early warning method and system, the method comprises: obtaining coastal blue carbon multi-source time series monitoring data, and obtaining historical carbon flux;Build carbon sink early warning model with exclusive door control fusion, generate orthogonal three-dimensional space base vector according to geographical position, obtain projection component by hydrological characteristic projection, built-in state door control unit, combine current projection component with history state to generate transformation coefficient step by step, calculate anisotropy scalar and cooperate with forget mechanism update unit state;After dimension alignment mapping, parameterized transformation, dimensionality reduction reverse reconstruction, and original feature residual error fusion output;Multi-source time series data is used to train model, after convergence, predict carbon sink change and solve stability index, when index continuous multiple steps are lower than threshold value, automatically trigger carbon sink early warning.

Description

Technical Field

[0001] This application belongs to the field of environmental protection, and in particular relates to a method and system for early warning of carbon sinks in coastal blue carbon ecosystems. Background Technology

[0002] Coastal blue carbon ecosystems, such as mangroves, salt marshes, and seagrass beds, play a crucial role in the global carbon cycle and climate change mitigation. They possess high carbon absorption and storage capacities, serving as natural carbon sinks for maintaining global climate stability. Due to the sensitive and volatile environment of their location in the marine-terrestrial transition zone, coastal blue carbon ecosystems face the dual challenges of climate change and human activities. Under multiple pressures, their ecosystem structures are easily damaged, leading to a decline in carbon sequestration rates and even a shift from sink to source. Carbon flux, as a core indicator of the rate and direction of carbon exchange between the ecosystem and its external environment, most directly reflects the dynamic transformation of sink-source states. Therefore, establishing a scientific early warning method for carbon sinks in coastal blue carbon ecosystems is significant for timely identification of degradation risks and ensuring carbon sink stability. Because the operational mechanisms of these ecosystems are highly variable, the monitoring process relies on multi-source time-series monitoring data encompassing hydrology, hydrochemistry, biological communities, and geographical location. This data exhibits high-dimensional fusion characteristics and high spatial heterogeneity across time and three-dimensional spatial scales, placing high demands on intelligent prediction and early warning technologies.

[0003] With the application of artificial intelligence technologies such as deep learning, data-driven time-series prediction models have been widely used in ecosystem monitoring. For example, the invention patent CN119540803A, "A Method for Monitoring and Measuring Carbon Sequestration of Mangrove Vegetation Based on UAV and AI Technology," utilizes a deep learning framework to identify mangrove area and biomass and calculate carbon sequestration. However, the carbon cycle and carbon flux of blue carbon ecosystems are influenced by geographical location and hydrodynamics, exhibiting anisotropy in different three-dimensional spatial directions. Conventional deep learning lacks the ability to extract directional differences in environmental factors within a local three-dimensional spatial framework. The diurnal and seasonal variations in carbon flux are processes with long-term time-series dependence and internal historical memory accumulation effects. Existing early warning models often sever the connection between spatial multi-dimensional directional feature extraction and temporal state updates, failing to integrate and fuse the changes in the intensity of spatial anisotropy with the state memory and forgetting mechanisms in the temporal evolution process. The lack of spatiotemporal collaborative representation capabilities limits the model's prediction accuracy for carbon sequestration fluctuations, resulting in insufficient sensitivity and accuracy when facing the critical point of declining carbon sequestration stability. Summary of the Invention

[0004] According to one aspect of the present invention, a carbon sink early warning method for coastal blue carbon ecosystems is proposed to address the problem that existing methods lack sufficient sensitivity and accuracy when facing a critical point of declining carbon sink stability, comprising the following steps: Acquire multi-source time-series monitoring data of the coastal blue carbon ecosystem. The multi-source time-series monitoring data includes hydrological parameters, hydrochemical parameters, biological community parameters and geographical location information. A carbon sink early warning model incorporating gated fusion is constructed. The gated fusion process is as follows: based on geographic location information, a set of local and mutually orthogonal three-dimensional spatial basis vectors are generated for the input feature vector; hydrological parameters in the feature vector are extracted as three-dimensional spatial features and projected onto the three-dimensional spatial basis vectors to obtain projection components; the gated fusion incorporates a gated unit with internal states; at each time step, the gated unit generates a set of transformation coefficients based on the projection components at the current time and the internal state at the previous time; based on the numerical distribution of the set of transformation coefficients, a scalar representing the anisotropy intensity is calculated, and the update process of the internal state is modulated using the scalar combined with a forgetting mechanism; After mapping the projected components using a dimension alignment matrix, parameterized control transformations are applied to the mapped projected components using transformation coefficients. The transformed projected components are then dimensionally reduced and reconstructed inversely along the three-dimensional spatial basis vector direction. The reconstructed components are then dimensionally matched with the original input feature vectors and the residuals are added using a broadcast rule or dimension expansion mechanism to obtain the output. Multi-source time-series monitoring data are input into the model for training. The converged model is used to predict carbon sink changes and calculate stability indices. When the stability indices are below a preset threshold for multiple consecutive time steps, an early warning is triggered.

[0005] Optionally, the acquisition of multi-source time-series monitoring data of the coastal blue carbon ecosystem includes hydrological parameters, hydrochemical parameters, biological community parameters, and geographical location information, including: The flow velocity, direction and tilt angle of seawater are acquired by an acoustic Doppler current meter at a fixed sampling period and stored in an environmental database as hydrological parameters. The salinity and dissolved oxygen concentration parameters within a specified seawater volume are acquired using underwater sensor components and stored in an environmental database as water chemistry parameters. By receiving and analyzing communication signals through hardware, the longitude, latitude, and elevation information of each data sampling point in the WGS1984 coordinate system are obtained and stored in the environmental database as geographical location information.

[0006] Optionally, generating a set of local and mutually orthogonal three-dimensional spatial basis vectors for the input feature vector based on geographical location information includes: The system acquires the longitude, latitude, and elevation information of the current sampling point, and transforms the longitude, latitude, and elevation information of the current sampling point into three-dimensional Cartesian coordinates with length dimensions. It also retrieves the three-dimensional Cartesian coordinates of the four nearest sampling points. Subtract the three-dimensional Cartesian coordinates of the current sampling point from the three-dimensional Cartesian coordinates of the four sampling points to obtain four relative position deviation vectors; Perform covariance analysis on all relative position deviation vectors and extract the top 3 principal eigenvectors with the largest corresponding eigenvalues; The first three principal eigenvectors are transformed into mutually orthogonal three-dimensional spatial basis vectors using the Schmitt orthogonalization method.

[0007] Optionally, the gating fusion incorporates a gating unit with internal states; at each time step, the gating unit generates a set of transformation coefficients based on the projection components at the current time and the internal states at the previous time, including: Extract the internal state vector retained by the gating unit in the previous time step, the internal state vector containing 128 nodes; The projected components at the current time step are linearly combined with the internal state vector at the previous time step through a fully connected layer, and the output is restricted to between 0 and 1 by the Sigmoid function to generate transformation coefficients.

[0008] Optionally, the process of calculating a scalar representing the anisotropy intensity based on the numerical distribution of the group of transformation coefficients, and modulating the update process of the internal state using the scalar combined with a forgetting mechanism, includes: Calculate the mean of all elements of the transformation coefficients; The sum of squares of the differences between each element of the transformation coefficient and the mean is calculated to obtain a scalar representing the intensity of anisotropy. Obtain the forgetting coefficient in the open interval between 0 and 1; When updating the internal state, the internal state of the previous time step is multiplied by the forgetting coefficient as the base term, and the scalar, the learnable weight vector and the weight coefficient are multiplied as the gain term. The base term and the gain term are added together to obtain the internal state of the current time step. When the scalar is greater than the anisotropy strength threshold, the weighting coefficient takes the first preset value; When the scalar is less than or equal to the anisotropy strength threshold, the weighting coefficient takes a second preset value, which is less than the first preset value. The learnable weight vector has the same dimension as the internal state.

[0009] Optionally, the step of mapping the projected components using a dimension alignment matrix and then applying a parameterized control transformation to the mapped projected components using transformation coefficients includes: The projection component at the current moment is mapped to an intermediate component with the same dimension as the transformation coefficient using a dimension alignment matrix; The intermediate components are multiplied element by element with the transformation coefficients to obtain the intermediate result after scaling adjustment. By activating the intermediate results using the hyperbolic tangent function, the values ​​are restricted to the positive and negative ranges, resulting in the transformed projection components.

[0010] Optionally, the step of reconstructing the transformed projection components by dimensionality reduction along the direction of the three-dimensional spatial basis vectors, and then performing dimensionality matching and residual addition with the original input feature vector through a broadcast rule or dimensionality expansion mechanism to obtain the output includes: The transformed projection components are restored to three-dimensional components using dimensionality reduction mapping. The three-dimensional components are reconstructed in reverse space using the transpose of the orthogonal matrix formed by the three-dimensional spatial basis vectors. The reconstructed spatial features are then matched with the original input feature vectors by dimension matching and residual addition using broadcast rules or dimension expansion mechanisms to obtain the final output. The final output is equal to the sum of the reconstructed spatial features after dimensional expansion and the feature vector of the original input.

[0011] According to another aspect of the present invention, a carbon sink early warning system for coastal blue carbon ecosystems is proposed, comprising the following modules: The acquisition module is used to acquire multi-source time-series monitoring data of the coastal blue carbon ecosystem. The multi-source time-series monitoring data includes hydrological parameters, hydrochemical parameters, biological community parameters and geographical location information. The module is used to construct a carbon sink early warning model incorporating gated fusion. The gated fusion process is as follows: based on geographic location information, a set of local and mutually orthogonal three-dimensional spatial basis vectors are generated for the input feature vector; hydrological parameters in the feature vector are extracted as three-dimensional spatial features and projected onto the three-dimensional spatial basis vectors to obtain projection components; the gated fusion incorporates a gated unit with internal states; at each time step, the gated unit generates a set of transformation coefficients based on the projection components at the current time and the internal state at the previous time; based on the numerical distribution of the set of transformation coefficients, a scalar representing the anisotropy intensity is calculated, and the update process of the internal state is modulated using the scalar combined with a forgetting mechanism; The calculation module is used to map the projected components using a dimension alignment matrix, and then apply parameterized control transformation to the mapped projected components using transformation coefficients. After the transformed projected components are reduced in dimension, they are reconstructed inversely along the direction of the three-dimensional spatial basis vectors. The output is obtained by dimension matching and residual addition with the feature vector of the original input through a broadcast rule or dimension expansion mechanism. Multi-source time-series monitoring data is input into the model for training. The converged model is used to predict carbon sink changes and calculate stability indicators. When the stability indicators are lower than the preset threshold for multiple consecutive time steps, an early warning is triggered.

[0012] Preferably, the acquisition of multi-source time-series monitoring data of the coastal blue carbon ecosystem includes hydrological parameters, hydrochemical parameters, biological community parameters, and geographical location information, including: The flow velocity, direction and tilt angle of seawater are acquired by an acoustic Doppler current meter at a fixed sampling period and stored in an environmental database as hydrological parameters. The salinity and dissolved oxygen concentration parameters within a specified seawater volume are acquired using underwater sensor components and stored in an environmental database as water chemistry parameters. By receiving and analyzing communication signals through hardware, the longitude, latitude, and elevation information of each data sampling point in the WGS1984 coordinate system are obtained and stored in the environmental database as geographical location information.

[0013] Preferably, generating a set of local and mutually orthogonal three-dimensional spatial basis vectors for the input feature vector based on geographical location information includes: The system acquires the longitude, latitude, and elevation information of the current sampling point, and transforms the longitude, latitude, and elevation information of the current sampling point into three-dimensional Cartesian coordinates with length dimensions. It also retrieves the three-dimensional Cartesian coordinates of the four nearest sampling points. Subtract the three-dimensional Cartesian coordinates of the current sampling point from the three-dimensional Cartesian coordinates of the four sampling points to obtain four relative position deviation vectors; Perform covariance analysis on all relative position deviation vectors and extract the top 3 principal eigenvectors with the largest corresponding eigenvalues; The first three principal eigenvectors are transformed into mutually orthogonal three-dimensional spatial basis vectors using the Schmitt orthogonalization method.

[0014] Preferably, the gating fusion incorporates a gating unit with internal states; at each time step, the gating unit generates a set of transformation coefficients based on the projection components at the current time and the internal states at the previous time, including: Extract the internal state vector retained by the gating unit in the previous time step, the internal state vector containing 128 nodes; The projected components at the current time step are linearly combined with the internal state vector at the previous time step through a fully connected layer, and the output is restricted to between 0 and 1 by the Sigmoid function to generate transformation coefficients.

[0015] Preferably, the process of calculating a scalar representing the anisotropy intensity based on the numerical distribution of the set of transformation coefficients, and modulating the update process of the internal state using the scalar combined with a forgetting mechanism, includes: Calculate the mean of all elements of the transformation coefficients; The sum of squares of the differences between each element of the transformation coefficient and the mean is calculated to obtain a scalar representing the intensity of anisotropy. Obtain the forgetting coefficient in the open interval between 0 and 1; When updating the internal state, the internal state of the previous time step is multiplied by the forgetting coefficient as the base term, and the scalar, the learnable weight vector and the weight coefficient are multiplied as the gain term. The base term and the gain term are added together to obtain the internal state of the current time step. When the scalar is greater than the anisotropy strength threshold, the weighting coefficient takes the first preset value; When the scalar is less than or equal to the anisotropy strength threshold, the weighting coefficient takes a second preset value, which is less than the first preset value. The learnable weight vector has the same dimension as the internal state.

[0016] Preferably, the step of mapping the projected components using a dimension alignment matrix and then applying a parameterized control transformation to the mapped projected components using transformation coefficients includes: The projection component at the current moment is mapped to an intermediate component with the same dimension as the transformation coefficient using a dimension alignment matrix; The intermediate components are multiplied element by element with the transformation coefficients to obtain the intermediate result after scaling adjustment. By activating the intermediate results using the hyperbolic tangent function, the values ​​are restricted to the positive and negative ranges, resulting in the transformed projection components.

[0017] Preferably, the step of reconstructing the transformed projection components by dimensionality reduction along the direction of the three-dimensional spatial basis vectors, and then performing dimensionality matching and residual addition with the original input feature vector through a broadcast rule or dimensionality expansion mechanism to obtain the output includes: The transformed projection components are restored to three-dimensional components using dimensionality reduction mapping. The three-dimensional components are reconstructed in reverse space using the transpose of the orthogonal matrix formed by the three-dimensional spatial basis vectors. The reconstructed spatial features are then matched with the original input feature vectors by dimension matching and residual addition using broadcast rules or dimension expansion mechanisms to obtain the final output. The final output is equal to the sum of the reconstructed spatial features after dimensional expansion and the feature vector of the original input.

[0018] This invention constructs a carbon sink early warning model with high-dimensional spatiotemporal feature extraction advantages by integrating multi-source monitoring data from hydrology, biology, and geography, thereby improving the accuracy of early warnings. The core lies in improved gating fusion, which generates a three-dimensional spatial basis vector based on geographical location and projects hydrological parameters to identify spatial heterogeneity characteristics under changing ecological environments. By inputting a gating unit, transformation coefficients are generated by combining historical internal states with current features. State updates are modulated based on anisotropic intensity scalars combined with a forgetting mechanism, removing redundant historical noise and enhancing the ability to identify long-term and short-term time-series dependencies. Furthermore, through mapping transformation, inverse spatial reconstruction, and residual addition, the model's representational power is enhanced while retaining original key information. This enables the model to predict the changing trends and stability of carbon sinks, achieving early warning of ecological degradation risks. Attached Figure Description

[0019] Figure 1A flowchart of the first embodiment; Figure 2 This is a schematic diagram of the coastal blue carbon ecosystem architecture of the present invention; Figure 3 This is a schematic diagram of the gating fusion structure; Figure 4 The scree plot shows the amplitude of the eigenvalues ​​of each principal component; Figure 5 This is a scatter plot showing the distribution of nonlinear coefficients at each node of the network. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] In the first embodiment, the present invention proposes a method for early warning of carbon sinks in coastal blue carbon ecosystems, such as... Figure 1 It includes the following steps: S1 acquires multi-source time-series monitoring data of the coastal blue carbon ecosystem, including hydrological parameters, hydrochemical parameters, biological community parameters, and geographical location information.

[0022] Read the dataset collected by the sensor network, which includes hydrological parameters such as tidal level and current velocity, water chemical parameters such as seawater salinity and dissolved oxygen concentration, biological community parameters such as mangrove canopy density and benthic biomass, and geographic location information such as longitude and latitude. Figure 2 As shown. In one embodiment, the training samples also include actual monitored carbon sinks, which include carbon flux values ​​obtained through carbon flux observation stations or periodic sampling analysis. The numerical data is then Z-score standardized. Missing values ​​in the time series data are imputed using piecewise cubic Hermitian interpolation to obtain a regularized multivariate time series tensor input matrix.

[0023] In an optional embodiment, the acquisition of multi-source time-series monitoring data of the coastal blue carbon ecosystem includes hydrological parameters, hydrochemical parameters, biological community parameters, and geographical location information, including: The flow velocity, direction and tilt angle of seawater are acquired by an acoustic Doppler current meter at a fixed sampling period and stored in an environmental database as hydrological parameters. The salinity and dissolved oxygen concentration parameters within a specified seawater volume are acquired using underwater sensor components and stored in an environmental database as water chemistry parameters. By receiving and analyzing communication signals through hardware, the longitude, latitude, and elevation information of each data sampling point in the WGS1984 coordinate system are obtained and stored in the environmental database as geographical location information.

[0024] An acoustic Doppler current meter was deployed to address the variable hydrological environment of the coastal area. This meter was permanently positioned in a typical sedimentary profile at a depth of 1.5 to 2.5 meters. The operating frequency was set to 1200 kHz, and measurements were taken at a fixed sampling period of 15 minutes to acquire current velocity data ranging from 0.05 to 3.50 m / s, absolute current direction data from 0 to 360 degrees, and inclination data within a spatial range of ±25 degrees. After acquisition, the hydrological parameters were packaged and structured into a PostgreSQL relational database using RS485 serial communication protocol combined with an encrypted transparent transmission module. The underwater sensor components were integrated and installed on a fixed platform at the same depth, including a conductivity probe with built-in titanium alloy electrodes and a fluorescence dissolved oxygen sensor probe. The sensor assembly is capable of continuously reading parameters for 24 hours within a specified volume of seawater per cubic meter, acquiring salinity data with an accuracy of 0.01 PSU and a range of 20 to 40 PSU, and dissolved oxygen concentration parameters with a detection range of 3.0 to 9.0 mg / L. These parameters, representing the hydrochemical environment of microbial carbon sequestration, are stored in the aforementioned environmental database. The environmental monitoring station, located above the water surface, integrates a GNSS RTK dual-frequency receiving hardware module. By connecting to the base station's NTRIP network and parsing real-time differential communication signals in NMEA 0183 format, it acquires centimeter-level three-dimensional position data. It obtains the spatial positions of various multi-source data sampling points in the WGS1984 coordinate system, specifically extracting longitude (e.g., between 118.5° and 120.5° E), latitude (e.g., between 25.5° and 27.5° N), and elevation information relative to the geoid from -5.0 to +10.0 meters, achieving centimeter-level accuracy. The geographic location information is linked to a timestamp and stored in the environmental database. Outliers are removed from the multi-source time-series monitoring data by Kalman filtering to ensure that the feature tensors input into the carbon sink early warning model have authenticity, multidimensional fusion and spatiotemporal consistency.

[0025] S2, Construct a carbon sink early warning model incorporating gated fusion. The gated fusion process is as follows: Based on geographical location information, generate a set of local and mutually orthogonal three-dimensional spatial basis vectors for the input feature vector; extract hydrological parameters from the feature vector as three-dimensional spatial features and project them onto the three-dimensional spatial basis vectors to obtain projection components; the gated fusion incorporates a gated unit with internal states; at each time step, the gated unit generates a set of transformation coefficients based on the projection components at the current time and the internal state at the previous time; calculate a scalar representing the anisotropy intensity based on the numerical distribution of the set of transformation coefficients, and use the scalar combined with a forgetting mechanism to modulate the update process of the internal state.

[0026] The backbone network structure is built based on Long Short-Term Memory (LSTM) networks, and the model includes an input layer, hidden layers, and an output layer. A temporal feature extraction module is constructed using the `torch.nn.LSTMCell` class, and an improved parameterized learnable anisotropic gating fusion class, inherited from `torch.nn.Module`, is embedded between the hidden layer neurons to replace the traditional ReLU gating fusion. In one embodiment, in the forward propagation sequence of the network, the gating fusion module is located between the feature input layer and the inverse spatial reconstruction layer, receiving the projected three-dimensional components of the hydrological environment. In terms of neuron topology, this module is not independently attached to a single neuron node, but rather functions as a unified hierarchical operation module spanning the 128 hidden layer neurons in the same layer. It achieves mutual interaction and fusion of historical state information of all neurons in the same layer by configuring a 128×128 dimensional weight matrix. This replaces the original independent activation functions such as ReLU in the hidden layers. Figure 3 As shown. The inverse spatial reconstruction layer transforms the features, which are in a high-dimensional abstract state and based on a local coordinate system after being processed by the deep learning network gated fusion module, back into a global three-dimensional space with real geographical significance, and fuses them with the original environmental data. Specifically, it can be a fully connected layer.

[0027] The longitude and latitude information is extracted from the input tensor and converted into an initial three-dimensional coordinate vector using the mapping formula from spherical coordinates to Cartesian coordinates. QR decomposition is performed on the matrix containing the initial three-dimensional coordinate vector and its two small random orthogonal perturbation vectors, and the orthogonal matrix Q obtained by decomposition is used as the local mutually orthogonal three-dimensional spatial basis vectors at the geographical location.

[0028] Three hydrological parameters, such as tidal level, current velocity, and wave height, are extracted from the standardized feature vector to form a three-dimensional spatial feature vector. Those skilled in the art should understand that these three parameters are merely illustrative and not limited to just three hydrological parameters. The dot product of this three-dimensional spatial feature vector and the previously generated three-dimensional spatial basis vector matrix is ​​calculated to achieve vector space projection, obtaining a projection component tensor containing the magnitudes of three orthogonal directions. In the gated fusion class, the internal state tensor is initialized as a zero vector, preferably using a multilayer perceptron as the gate unit. During the cyclic forward propagation of the time series, the projection component acquired at the current moment is concatenated with the cached internal state tensor from the previous moment along the feature dimension. The concatenation result is input into the multilayer perceptron, sequentially processed through linear mapping and Sigmoid gated fusion, outputting a set of transformation coefficient vectors with values ​​between 0 and 1. The unbiased sample variance of this set of transformation coefficients, or the sum of squares of the differences between each element and the mean, is calculated, and the variance or sum of squares is used as a scalar representing the intensity of anisotropy. A forget gate based on the scalar is used to calculate the negative exponential decay value of the scalar and generate the forget weight. The internal state of the previous time step is multiplied by the forget weight using the bitwise multiplication operator, and then the multiplication result is added to the transformation coefficient of the current time step using the addition operator to complete the update of the internal state tensor at the current time step.

[0029] In an optional embodiment, generating a set of local and mutually orthogonal three-dimensional spatial basis vectors for the input feature vector based on geographic location information includes: The system acquires the longitude, latitude, and elevation information of the current sampling point, and transforms the longitude, latitude, and elevation information of the current sampling point into three-dimensional Cartesian coordinates with length dimensions. It also retrieves the three-dimensional Cartesian coordinates of the four nearest sampling points. Subtract the three-dimensional Cartesian coordinates of the current sampling point from the three-dimensional Cartesian coordinates of the four sampling points to obtain four relative position deviation vectors; Perform covariance analysis on all relative position deviation vectors and extract the top 3 principal eigenvectors with the largest corresponding eigenvalues; The first three principal eigenvectors are transformed into mutually orthogonal three-dimensional spatial basis vectors using the Schmitt orthogonalization method.

[0030] The preprocessing module extracts the longitude, latitude, and elevation arrays of the monitoring stations output by the GNSS sensor network in real time, calls the WGS84 reference ellipsoid standard model, sets the Earth's semi-major axis a to 6378137.0 meters, and sets the first eccentricity squared... The coordinate transformation interface is 0.00669437999. Using the spatial mapping equation from the geodetic coordinate system to the geocentric coordinate system, spherical angles and elevations are converted into three-dimensional Cartesian coordinates with a unified length dimension. For example, mapped as , , These are equal to -2,867,543.12, 4,893,121.45, and 2,901,324.78, respectively. Using a KD-Tree spatial index data structure based on Euclidean distance, a nearest neighbor search is performed. Within a set radius constraint threshold of 500 meters, the three-dimensional Cartesian coordinates of the four nearest discrete sampling points are retrieved. to The coordinates of these four neighboring stations are obtained by subtraction. With current center coordinates Subtracting them yields the relative position deviation vector. For example, calculated The values ​​are 12.5, -8.3, and 4.2. Four 1×3 row vectors are concatenated row-wise to form a 4×3 positional deviation matrix D. To extract the dominant dimension of topographic changes, the result of multiplying the transpose of deviation matrix D by itself is calculated, generating a 3×3 covariance matrix. The Jacobi eigenvalue solving algorithm is used to solve this covariance matrix. The eigenvalues ​​are sorted in descending order, and the three column vectors corresponding to the largest eigenvalue are extracted as the principal eigenvectors. , , During the feature orthogonalization stage, the Schmitt orthogonalization processing logic is followed to calculate... The L2 norm, and with Dividing by the L2 norm to perform normalization generates the first spatial basis. In generating the second base When calculating using the vector inner product The scalar projected length will Along The projected components of the directions are removed, and the L2 norm of the difference is calculated and normalized. Similarly, a third basis is generated. Remove at time In already determined and The projection interference in the direction is eliminated and normalized. The output is a set of three-dimensional spatial basis vectors through the above processing. , and In a locally orthogonal space, the anisotropic directions of tidal erosion and organic carbon particle deposition and diffusion can be identified, such as... Figure 4 As shown.

[0031] In an optional embodiment, the gating fusion incorporates a gating unit with internal states; at each time step, the gating unit generates a set of transformation coefficients based on the projection component at the current time and the internal state at the previous time, including: Extract the internal state vector retained by the gating unit in the previous time step, the internal state vector containing 128 nodes; The projected components at the current time step are linearly combined with the internal state vector at the previous time step through a fully connected layer, and the output is restricted to between 0 and 1 by the Sigmoid function to generate transformation coefficients.

[0032] The temporal feature processing neural network model comprises, in series, an input layer, a gated unit layer, an internal state modulation layer, a parameterized feature control layer, an inverse spatial reconstruction layer, and a residual output layer. The model's input consists of the hydrodynamic environment projection components at the current moment and the original multi-source environmental monitoring feature vectors. The model's output is the final environmental reconstruction fusion feature tensor obtained by adding the inverse spatial reconstruction results to the original features via residual concatenation. When the model is forward-propagated to a set time resolution step, such as... When the time interval is 1 hour, the input at the current moment is extracted and recorded as the projected component. This component is a 3×1 column vector obtained by projecting the original three-dimensional hydrodynamic features onto the spatial base. The module extracts the internal state vector retained by the gating unit at the end of the previous cycle t-1. The state vector is instantiated as a one-dimensional array containing 128 compute nodes. For the first time step, The node parameters are randomly initialized using a Gaussian normal distribution with a mean of 0.0 and a standard deviation of 0.02. To integrate external input and historical states, the model is configured with two learnable parameter matrices: one for the input... weight matrix Dimensions 128×3 and adjustment status weight matrix The network has dimensions of 128×128. During the network initialization phase... and Initialization is completed according to the Xavier uniform distribution strategy. Execution and , and The multiplication operation sums the resulting feature tensors and then adds them to a bias vector of size 128×1 with initial values ​​of 0.0. The linear combination of the two data streams is then performed. After the linear combination, to prevent gradient vanishing or numerical overflow caused by abrupt changes in flow rate, numerical control logic is input to truncate the elements within the linear combination term by -15.0 to 15.0. The 128 truncated components are then input element-wise into the Sigmoid function, and summed according to the negative power rule of the natural constant e. The result of the linear combination is mapped to the interval 0.0 to 1.0, and a 128×1 dimensional transformation coefficient tensor is output. This is used to activate the scaling adjustment of the channel. Figure 5 The scatter plot shows the distribution of nonlinear coefficients of each node in the gated unit, representing the differences in response intensity among different nodes.

[0033] In an optional embodiment, the process of calculating a scalar representing the anisotropy intensity based on the numerical distribution of the group of transformation coefficients, and modulating the update process of the internal state using the scalar combined with a forgetting mechanism, includes: Calculate the mean of all elements of the transformation coefficients; The sum of squares of the differences between each element of the transformation coefficient and the mean is calculated to obtain a scalar representing the intensity of anisotropy. Obtain the forgetting coefficient in the open interval between 0 and 1; When updating the internal state, the internal state of the previous time step is multiplied by the forgetting coefficient as the base term, and the scalar, the learnable weight vector and the weight coefficient are multiplied as the gain term. The base term and the gain term are added together to obtain the internal state of the current time step. When the scalar is greater than the anisotropy strength threshold, the weighting coefficient takes the first preset value; When the scalar is less than or equal to the anisotropy strength threshold, the weighting coefficient takes a second preset value, which is less than the first preset value. The learnable weight vector has the same dimension as the internal state.

[0034] Upon receiving the transformation coefficient vector with a dimension of 128×1 generated by the gating unit Then, the summation of the 128 elements along the feature channels is divided by the total number of nodes, 128, to extract the mean scalar μ. (The process is repeated.) Each element in the vector is subtracted from its mean μ. The resulting difference is then squared element-wise and summed to calculate a scalar representing the intensity of anisotropy. This scalar represents the degree of spatial heterogeneity of the current flow field. In the internal state update module, the trigger threshold for anisotropy intensity is pre-set to 1.6. The forgetting coefficient parameter γ is input, with its value limited to the range of 0.0 to 1.0; the actual set value is 0.92. When the calculated... When the value exceeds the threshold of 1.6, the environment is determined to be in a state of significant disturbance, and the gain coefficient is configured to 1.0, according to the update equation. Update the feature states inside the hidden layer. Conversely, when computing the output... When the threshold is less than or equal to 1.6, the marine environment is considered to be in a stable cycle. To avoid observational noise disturbing the state update, the gain coefficient is configured to 0.3, according to the update equation. This generates the latent variable outcome for the current cycle. The learnable adjustable weight vector in the update formula is then used. It is initialized to a 128×1 dimension. This weight will be continuously updated in each round of backpropagation by the AdamW optimizer, which incorporates momentum and weight decay mechanisms.

[0035] S3: After mapping the projected components using a dimension alignment matrix, parameterized control transformation is applied to the mapped projected components using transformation coefficients. The transformed projected components are then dimensionally reduced and restored, and inverse spatial reconstruction is performed along the direction of the three-dimensional spatial basis vectors. The output is obtained by dimension matching and residual addition with the feature vector of the original input through broadcasting rules or dimension expansion mechanisms. Multi-source time-series monitoring data is input into the model for training. The converged model is used to predict carbon sink changes and calculate stability indicators. When the stability indicators are lower than the preset threshold for multiple consecutive time steps, an early warning is triggered.

[0036] In the model initialization phase, a learnable weight matrix is ​​set as the dimension alignment matrix, and `torch.nn.Parameter` is used to register this matrix as a model parameter. The projection component at the current time step is multiplied by the dimension alignment matrix to expand the feature dimension to match the dimension of the transformation coefficients, completing the dimension mapping. The Hadamard product is used to calculate the element-wise product of the mapped projection component and the transformation coefficients, completing the parameterized control transformation. A dimension reduction linear transformation layer `torch.nn.Linear` is set to reduce the dimension of the tensor feature dimension after parameterized control transformation to three dimensions. The product of this three-dimensional tensor and the transpose of the three-dimensional spatial basis vectors is calculated to complete the inverse spatial reconstruction. The `torch.unsqueeze` function is used to add spatial dimension to the reconstructed tensor to trigger an automatic broadcasting mechanism. The reconstructed features after broadcasting are added to the original input feature vector element-wise by residual calculation to generate the final activation output tensor. The final activation output tensor is input into a regression output network composed of fully connected layers to reduce the dimension of the high-dimensional feature mapping and output the predicted carbon sink.

[0037] The AdamW optimization algorithm is used to update model parameters. An initial learning rate is set, and the mean square error between the predicted carbon sink and the actual monitored carbon sink is calculated. Backpropagation training is performed until the loss function converges. During the inference phase, real-time multi-source time-series monitoring data is fed into the converged model, outputting a predicted future carbon sink numerical sequence. The reciprocal of the sliding variance of the predicted carbon sink sequence is calculated as a stability index within a rolling time window of twelve time steps. If this stability index is less than a set absolute threshold for five consecutive time steps, an email containing the latitude and longitude of the warning stations and the predicted values ​​is sent, or... Figure 2 The user terminal display shows that a carbon sequestration warning has been triggered.

[0038] In an optional embodiment, the step of mapping the projected components using a dimension alignment matrix and then applying a parameterized control transformation to the mapped projected components using transformation coefficients includes: The projection component at the current moment is mapped to an intermediate component with the same dimension as the transformation coefficient using a dimension alignment matrix; The intermediate components are multiplied element by element with the transformation coefficients to obtain the intermediate result after scaling adjustment. By activating the intermediate results using the hyperbolic tangent function, the values ​​are restricted to the positive and negative ranges, resulting in the transformed projection components.

[0039] During the parametric control and spatial feature transformation process, the model represents the extracted projection components as 3×1 column vectors. To ensure compatibility with gating states, a weight matrix responsible for tensor dimension alignment is initialized. The vector has dimensions of 128×3, and its initial elements are randomly generated using a standard normal distribution. By performing matrix multiplication, the low-dimensional projected components are mapped to a 128×1 dimensional intermediate eigenvector, achieving the same result as the transformation coefficients. The dimensions are aligned. Then, the 128×1 intermediate feature vector is matched with the coefficient vector. Perform element-wise multiplication. Through element-wise multiplication, the current hydrodynamic projection features are constrained by gating coefficients, achieving spatial anisotropy weighting and outputting a 128×1 dimensional intermediate result. To ensure good fitting and maintain the stability of the calculated values, the network model uses hyperbolic tangent gating fusion. Intermediate results... The 128 components within are calculated independently: a positive value is obtained based on the natural logarithm constant e. Power and negative The difference of powers is divided by the sum of the two terms. This fractional operation logic maps the feature values ​​to the interval between -1.0 and 1.0. The resulting 128×1 transformed projection tensor It avoids gradient explosion that may occur in deep training and restores the response characteristics of carbon sink sediment deposition process to tidal scouring.

[0040] In an optional embodiment, the step of reconstructing the transformed projection components by dimensionality reduction along the direction of the three-dimensional spatial basis vectors, and then performing dimensionality matching and residual addition with the original input feature vector through a broadcast rule or dimensionality expansion mechanism to obtain the output includes: The transformed projection components are restored to three-dimensional components using dimensionality reduction mapping. The three-dimensional components are reconstructed in reverse space using the transpose of the orthogonal matrix formed by the three-dimensional spatial basis vectors. The reconstructed spatial features are then matched with the original input feature vectors by dimension matching and residual addition using broadcast rules or dimension expansion mechanisms to obtain the final output. The final output is equal to the sum of the reconstructed spatial features after dimensional expansion and the feature vector of the original input.

[0041] During the model feature fusion stage, a transformation tensor with a dimension of 128×1 is extracted. To achieve spatial inverse reconstruction, a 3×128 linear projection matrix is ​​configured. Matrix multiplication is then performed on the two vectors to reduce the dimensionality of the higher-order vector to a 3×1 three-dimensional component. The three spatial basis vectors previously generated by orthogonalization are then retrieved. , , The spatial basis vectors are combined to construct a 3×3 orthogonal matrix. By utilizing the algebraic property that the transpose of an orthogonal matrix is ​​its inverse, we can extract the orthogonal matrix. The transpose of the matrix is ​​multiplied on the left by the 3×1 three-dimensional components obtained from the previous reconstruction, and inverse spatial reconstruction is performed. This process remaps the features from the local orthogonal coordinate system to the global three-dimensional coordinate space, generating a 3×1 reconstructed spatial feature tensor. To input this feature into the network, the retained original multi-source monitoring input feature tensor is extracted. The channel width is set to D equal to 64 dimensions. For channel differences, an automatic shape broadcasting mechanism is invoked, such as performing one-dimensional expansion and lateral tiling instructions, or expanding the dimension through a 1D point convolution operation with 3 input channels, 64 output channels, and a kernel size of 1. This expands the 3×1 spatial reconstruction features in the channel dimension to the same level as... The same 64-dimensional array. Perform residual join stacking operation, i.e. Equals the reconstructed features after the dimensional expansion operation and The sum of these residual connections prevents gradient vanishing during deep training while ensuring the model retains original environmental detection information during the extraction of temporal features.

[0042] In the second embodiment, the present invention also proposes a coastal blue carbon ecosystem carbon sink early warning system, comprising the following modules: The acquisition module is used to acquire multi-source time-series monitoring data of the coastal blue carbon ecosystem. The multi-source time-series monitoring data includes hydrological parameters, hydrochemical parameters, biological community parameters and geographical location information. The module is used to construct a carbon sink early warning model incorporating gated fusion. The gated fusion process is as follows: based on geographic location information, a set of local and mutually orthogonal three-dimensional spatial basis vectors are generated for the input feature vector; hydrological parameters in the feature vector are extracted as three-dimensional spatial features and projected onto the three-dimensional spatial basis vectors to obtain projection components; the gated fusion incorporates a gated unit with internal states; at each time step, the gated unit generates a set of transformation coefficients based on the projection components at the current time and the internal state at the previous time; based on the numerical distribution of the set of transformation coefficients, a scalar representing the anisotropy intensity is calculated, and the update process of the internal state is modulated using the scalar combined with a forgetting mechanism; The calculation module is used to map the projected components using a dimension alignment matrix, and then apply parameterized control transformation to the mapped projected components using transformation coefficients. After the transformed projected components are reduced in dimension, they are reconstructed inversely along the direction of the three-dimensional spatial basis vectors. The output is obtained by dimension matching and residual addition with the feature vector of the original input through a broadcast rule or dimension expansion mechanism. Multi-source time-series monitoring data is input into the model for training. The converged model is used to predict carbon sink changes and calculate stability indicators. When the stability indicators are lower than the preset threshold for multiple consecutive time steps, an early warning is triggered.

[0043] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for early warning of carbon sinks in coastal blue carbon ecosystems, characterized in that, include: Multi-source time-series monitoring data of the coastal blue carbon ecosystem were acquired. The multi-source time-series monitoring data included hydrological parameters, hydrochemical parameters, biological community parameters and geographical location information. The multi-source time-series monitoring data were standardized and missing values ​​were interpolated and filled to obtain the input feature vector. A carbon sequestration early warning model incorporating gated fusion is constructed. The gated fusion process is as follows: Based on geographic location information, a set of local and mutually orthogonal three-dimensional spatial basis vectors are generated for the input feature vector. This includes: acquiring the longitude, latitude, and elevation information of the current sampling point; unifying the longitude, latitude, and elevation information of the current sampling point into geocentric and geofixed three-dimensional Cartesian coordinates with length dimensions through coordinate transformation; and retrieving the three-dimensional Cartesian coordinates of the four nearest sampling points; subtracting the three-dimensional Cartesian coordinates of the current sampling point from the three-dimensional Cartesian coordinates of the four sampling points to obtain four relative position deviation vectors; and processing all relative position deviation vectors... Covariance analysis is performed to extract the top three principal eigenvectors with the largest corresponding eigenvalues. The Schmitt orthogonalization method is used to transform these three principal eigenvectors into mutually orthogonal three-dimensional spatial basis vectors. Hydrological parameters from the eigenvectors are extracted as three-dimensional spatial features and projected onto the three-dimensional spatial basis vectors to obtain projection components. Gated fusion incorporates a gated unit with internal states. At each time step, the gated unit generates a set of transformation coefficients based on the projection components at the current time and the internal state at the previous time. A scalar representing the anisotropy intensity is calculated based on the numerical distribution of the transformation coefficients, and the update process of the internal state is modulated using a scalar combined with a forgetting mechanism. After mapping the projected components using a dimension alignment matrix, a parameterized control transformation is applied to the mapped projected components using transformation coefficients. The transformed projected components are then dimension-reduced and reconstructed inversely along the direction of the three-dimensional spatial basis vectors. Finally, the output is obtained by dimension matching and residual addition with the feature vector of the original input through a broadcasting rule or dimension expansion mechanism. Multi-source time-series monitoring data are input into the model for training. The converged model is used to predict carbon sink changes and calculate stability indicators. When the stability indicators are lower than the preset threshold for multiple consecutive time steps, an early warning is triggered.

2. The method according to claim 1, characterized in that, The acquisition of multi-source time-series monitoring data of the coastal blue carbon ecosystem includes hydrological parameters, hydrochemical parameters, biological community parameters, and geographical location information, including: The flow velocity, direction and tilt angle of seawater are acquired by an acoustic Doppler current meter at a fixed sampling period and stored in an environmental database as hydrological parameters. The salinity and dissolved oxygen concentration parameters within a specified seawater volume are acquired using underwater sensor components and stored in an environmental database as water chemistry parameters. By receiving and analyzing communication signals through hardware, the longitude, latitude, and elevation information of each data sampling point in the WGS1984 coordinate system are obtained and stored in the environmental database as geographical location information.

3. The method according to claim 1, characterized in that, The gating fusion has a built-in gating unit with internal states; At each time step, the gating unit generates a set of transformation coefficients based on the projection components at the current time and the internal state at the previous time, including: Extract the internal state vector retained by the gating unit in the previous time step, the internal state vector containing 128 nodes; The projected components at the current time step are linearly combined with the internal state vector at the previous time step through a fully connected layer, and the output is restricted to between 0 and 1 by the Sigmoid function to generate transformation coefficients.

4. The method according to claim 1, characterized in that, The calculation of a scalar representing the anisotropy intensity based on the numerical distribution of the transformation coefficients, and the modulation of the internal state update process using the scalar combined with a forgetting mechanism, includes: Calculate the mean of all elements of the transformation coefficients; The sum of squares of the differences between each element of the transformation coefficient and the mean is calculated to obtain a scalar representing the intensity of anisotropy. Obtain the forgetting coefficient in the open interval between 0 and 1; When updating the internal state, the internal state of the previous time step is multiplied by the forgetting coefficient as the base term, and the scalar, the learnable weight vector and the weight coefficient are multiplied as the gain term. The base term and the gain term are added together to obtain the internal state of the current time step. When the scalar is greater than the anisotropy strength threshold, the weighting coefficient takes the first preset value; When the scalar is less than or equal to the anisotropy strength threshold, the weighting coefficient takes a second preset value, which is less than the first preset value. The learnable weight vector has the same dimension as the internal state.

5. The method according to claim 1, characterized in that, The step of mapping the projected components using a dimension alignment matrix and then applying a parameterized control transformation to the mapped projected components using transformation coefficients includes: The projection component at the current moment is mapped to an intermediate component with the same dimension as the transformation coefficient using a dimension alignment matrix; The intermediate components are multiplied element by element with the transformation coefficients to obtain the intermediate result after scaling adjustment. By activating the intermediate results using the hyperbolic tangent function, the values ​​are restricted to the positive and negative ranges, resulting in the transformed projection components.

6. The method according to claim 1, characterized in that, The process of reducing the dimensionality of the transformed projection components and then reconstructing the space inversely along the direction of the three-dimensional basis vectors, followed by dimensionality matching and residual addition with the original input feature vectors via broadcasting rules or dimensionality expansion mechanisms, yields the output, including: The transformed projection components are restored to three-dimensional components using dimensionality reduction mapping. The three-dimensional components are reconstructed in reverse space using the transpose of the orthogonal matrix formed by the three-dimensional spatial basis vectors. The reconstructed spatial features are then matched with the original input feature vectors by dimension matching and residual addition using broadcast rules or dimension expansion mechanisms to obtain the final output. The final output is equal to the sum of the reconstructed spatial features after dimensional expansion and the feature vector of the original input.

7. A coastal blue carbon ecosystem carbon sink early warning system, characterized in that, include: The acquisition module is used to acquire multi-source time-series monitoring data of the coastal blue carbon ecosystem. The multi-source time-series monitoring data includes hydrological parameters, hydrochemical parameters, biological community parameters and geographical location information. The module also performs standardization processing and interpolation to fill in missing values ​​in the multi-source time-series monitoring data to obtain the input feature vector. The module is used to construct a carbon sequestration early warning model incorporating gated fusion. The gated fusion process is as follows: Based on geographical location information, a set of local and mutually orthogonal three-dimensional spatial basis vectors are generated for the input feature vectors. This includes: obtaining the longitude, latitude, and elevation information of the current sampling point; unifying the longitude, latitude, and elevation information of the current sampling point into geocentric and geofixed three-dimensional Cartesian coordinates with length dimensions through coordinate transformation; and retrieving the three-dimensional Cartesian coordinates of the four nearest sampling points; subtracting the three-dimensional Cartesian coordinates of the current sampling point from the three-dimensional Cartesian coordinates of the four sampling points to obtain four relative position deviation vectors; and applying these vectors to all relative positions. Covariance analysis is performed on the deviation vectors to extract the top three principal eigenvectors with the largest corresponding eigenvalues. The Schmitt orthogonalization method is used to transform the top three principal eigenvectors into mutually orthogonal three-dimensional spatial basis vectors. Hydrological parameters in the eigenvectors are extracted as three-dimensional spatial features and projected onto the three-dimensional spatial basis vectors to obtain projection components. The gating fusion has a built-in gating unit with internal state. At each time step, the gating unit generates a set of transformation coefficients based on the projection components at the current time and the internal state at the previous time. Based on the numerical distribution of the transformation coefficients, a scalar representing the anisotropy intensity is calculated, and the update process of the internal state is modulated using the scalar combined with a forgetting mechanism. The calculation module is used to map the projected components using a dimension alignment matrix, and then apply parameterized control transformation to the mapped projected components using transformation coefficients. After the transformed projected components are reduced in dimension, they are reconstructed inversely along the direction of the three-dimensional spatial basis vectors. The output is obtained by dimension matching and residual addition with the feature vector of the original input through a broadcast rule or dimension expansion mechanism. Multi-source time-series monitoring data is input into the model for training. The converged model is used to predict carbon sink changes and calculate stability indicators. When the stability indicators are lower than the preset threshold for multiple consecutive time steps, an early warning is triggered.

8. The system according to claim 7, characterized in that, The acquisition of multi-source time-series monitoring data of the coastal blue carbon ecosystem includes hydrological parameters, hydrochemical parameters, biological community parameters, and geographical location information, including: The flow velocity, direction and tilt angle of seawater are acquired by an acoustic Doppler current meter at a fixed sampling period and stored in an environmental database as hydrological parameters. The salinity and dissolved oxygen concentration parameters within a specified seawater volume are acquired using underwater sensor components and stored in an environmental database as water chemistry parameters. By receiving and analyzing communication signals through hardware, the longitude, latitude, and elevation information of each data sampling point in the WGS1984 coordinate system are obtained and stored in the environmental database as geographical location information.

Images