A remote sensing image marine green tide monitoring method and system fusing physical priori and space-time evolution

Abstract

The present application relates to the technical field of remote sensing monitoring, in particular to a remote sensing image marine green tide monitoring method and system fusing physical priori and space-time evolution. The method comprises the following steps: acquiring multi-modal remote sensing monitoring images; performing multi-modal feature extraction on the acquired images, including spectral reflectance feature extraction, marine dynamics feature extraction, and feature alignment and unified representation; establishing physical priori of green tide characteristic band by using marine optical radiation transfer model; constructing dynamic space-time graph based on the extracted multi-modal features to obtain node global feature vector and dynamic adjacency matrix; performing adaptive graph convolution feature coding based on the physical priori and the dynamic space-time graph; and realizing consistency processing and high-precision extraction of multi-source features by fusing multi-platform multi-spectral images such as satellites and unmanned aerial vehicles and marine dynamic data, combining atmospheric correction and band resampling, thereby significantly improving the comprehensiveness and reliability of green tide feature recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing monitoring, in particular to a remote sensing image marine green tide monitoring method and system fusing physical priori and spatio-temporal evolution. BACKGROUND

[0002] As an important form of marine ecological disaster, green tide is often caused by the outbreak of large-area floating algae such as Enteromorpha, and has the characteristics of strong suddenness, fast diffusion speed and wide coverage. Green tide can lead to depletion of marine dissolved oxygen, deterioration of water quality, loss of fishery resources and damage to coastal economy, and has become a major challenge to marine environmental safety monitoring and ecological management.

[0003] At present, marine green tide monitoring and early warning mainly rely on multispectral remote sensing images. By obtaining the reflectivity information of the sea area in visible light, near-infrared and other multi-band, the distribution and evolution process of marine green tide are identified. Traditional green tide detection methods can be divided into three categories: first, the detection method based on spectral index, which uses the reflectivity difference between green tide and surrounding water in near-infrared and visible light bands to realize detection by constructing green tide sensitive index (such as normalized vegetation index NDVI, floating algae detection index FAI, normalized green algae index NGI, etc.). This method is simple to implement and efficient to calculate, but it relies on fixed thresholds and is difficult to adapt to changes in water turbidity, suspended particulate matter concentration and observation conditions in different sea areas, and the detection accuracy is significantly limited. Second, the detection method based on physical model, which uses marine radiation transfer model or water optical model, combined with the physical priori knowledge of green tide such as chlorophyll absorption peak and spectral reflectance characteristics, to derive the judgment rule of the existence of green tide. This method has strong physical interpretability, but the model construction is complex, the parameters are difficult to obtain, and the model accuracy is easily affected by different sensors, different times and variable marine dynamic conditions, and the generalization ability is insufficient. Third, the detection method based on data-driven, which uses machine learning or deep learning model to extract features and classify remote sensing images, to realize automatic detection of green tide. In recent years, convolutional neural networks, Transformers and other methods have made significant progress in intelligent analysis of remote sensing images, and can mine complex patterns from high-dimensional spectral and spatial features. However, deep learning methods have high dependence on large-scale high-quality labeled samples, and are prone to overfitting in small sample sea areas and complex observation conditions, and it is difficult to fuse marine dynamics and physical priori, resulting in a lack of spatio-temporal consistency and interpretability of the detection results.

[0004] Overall, the existing marine green tide detection methods still have obvious shortcomings in physical interpretability, timing consistency, and cross-regional generalization ability: lack of physical prior constraints, most methods rely only on data statistical characteristics, ignoring the specific spectral absorption mechanism of green tide and the optical properties of marine water, resulting in insufficient robustness of the algorithm in different sea areas and different sensors; lack of spatio-temporal evolution modeling, the occurrence and spread of green tide are driven by multiple marine dynamic factors such as ocean currents, wind fields, and tides, and existing methods are mostly single-time, static detection, which is difficult to depict the propagation path and life cycle evolution law of green tide; lack of multi-source heterogeneous data fusion, multi-spectral images, historical time series data, and marine dynamic models cannot be effectively integrated, resulting in limited detection accuracy and prediction ability.

[0005] Therefore, it is urgent to propose a multi-spectral remote sensing image marine green tide detection method that integrates physical prior and spatio-temporal evolution constraints. SUMMARY

[0006] In order to solve the problems of insufficient utilization of physical prior, weak spatio-temporal evolution modeling ability, and insufficient fusion of multi-source and multi-temporal information in the process of marine green tide remote sensing detection, the present application provides a remote sensing image marine green tide monitoring method and system that integrates physical prior and spatio-temporal evolution.

[0007] In the first aspect, the present application provides a remote sensing image marine green tide monitoring method that integrates physical prior and spatio-temporal evolution, which adopts the following technical solution:

[0008] A remote sensing image marine green tide monitoring method that integrates physical prior and spatio-temporal evolution, comprising:

[0009] Obtaining multi-modal remote sensing monitoring images;

[0010] Performing multi-modal feature extraction on the obtained images, including spectral reflectance feature extraction, marine dynamics feature extraction, and feature alignment and unified representation;

[0011] Using a marine optical radiation transfer model to establish physical prior of green tide characteristic bands;

[0012] Based on the extracted multi-modal features, constructing a dynamic spatio-temporal graph to obtain node global feature vectors and a dynamic adjacency matrix;

[0013] Based on the physical prior and the dynamic spatio-temporal graph, performing adaptive graph convolution feature coding;

[0014] Based on the physical prior and the time series Transformer, modeling the multi-modal features in time series;

[0015] Performing marine green tide prediction by fusing the adaptive graph convolution feature coding and the time series modeling results;

[0016] Outputting the prediction results.

[0017] Further, the multi-modal feature extraction on the acquired image comprises spectral reflectance feature extraction, wherein the acquired multi-band original radiation value is R raw , and first, atmospheric correction and band resampling are performed to obtain apparent water body reflectance R(λ); after obtaining the water body reflectance, key vegetation indices are extracted for green tide sensitive bands; for ocean dynamics feature extraction, the original ocean dynamics data is R d , and an interpolation and re-projection method is used to calculate an ocean dynamics feature vector; in order to dynamically depict the trend of green tide change with the dynamics condition, a multi-temporal sequence {X d t} is used to calculate a time sequence gradient; finally, a cross-platform alignment method based on adversarial learning is used, wherein the high-resolution spectral feature extracted by the unmanned aerial vehicle is X X u , the low-resolution spectral feature extracted by the satellite is X X s , and the multi-platform feature is unified by minimizing an alignment loss function to realize the consistency of the multi-platform features, so that the multi-source features are uniformly mapped to a common representation space to obtain a global feature vector, wherein the global feature vector comprises spectral information and ocean dynamics prior.

[0018] Further, the physical prior of the green tide feature band is established by using the ocean optical radiation transfer model, which comprises introducing an ocean optical radiation transfer model RTM, defining a theoretical reflectance vector of the green tide pixel in different bands R phy , constructing a theoretical prior feature of the typical spectrum of the green tide based on the RTM, and in the process of training the depth model, the theoretical prior curve and the model predicted spectrum are constrained, the difference in shape between the two is minimized, so that the prediction result still conforms to the physical law under different observation conditions, and the spectral shape consistency loss is:

[0019] ,

[0020] wherein, R pred represents the reflectance predicted by the depth model, R phy represents the theoretical spectrum generated by the RTM; the linear mixing abundance constraint based on the endmember theory represents the reflectance of each pixel as a linear combination of the reflectance of multiple endmembers; the theoretical reflectance curve and the prior abundance distribution of different endmembers are obtained according to the ocean optical radiation transfer model; then in the process of training the depth model, the prediction result is consistent with the theoretical combination result, and the linear mixing abundance constraint loss is introduced:

[0021] ,

[0022] wherein, a phy represents the theoretical abundance distribution calculated by RTM, a pred represents the theoretical abundance distribution predicted by depth.

[0023] Further, the dynamic spatio-temporal graph is constructed based on the extracted multi-modal features, including taking the remote sensing pixels as the graph nodes, and constructing a dynamic spatio-temporal graph dynamically reflecting the spatial diffusion relationship and the temporal evolution law, wherein the remote sensing image is divided into N nodes, and the geographic coordinates of each node are P i ( lon i , lat i ), the Euclidean distance between nodes is d ij =∥p i -p j ∥, the spatial similarity weight between nodes i and j is defined according to the Gaussian kernel function, and when the two nodes are blocked by the coastline or island topography, the shielding is performed through the mask matrix Bij; the ocean current driving weight and the wind field driving weight are calculated based on the marine dynamic process, and the ocean current and wind field effects are weighted and fused, the historical outbreak data is used to construct a temporal evolution prior graph, and finally the geographical similarity, the dynamic driving and the historical prior are fused to construct a dynamic adjacency matrix at time t:

[0024] ,

[0025] wherein α, β, δ are preset weights, A geo represents the spatial adjacency relationship matrix, A dyn represents the dynamic driving feature matrix, P t represents the temporal transition probability matrix, and is normalized to obtain:

[0026] ,

[0027] wherein, D t represents the diagonal matrix, and the output {A t} is a dynamic spatio-temporal graph sequence, which is a graph structure prior for deep spatio-temporal modeling.

[0028] Further, the adaptive graph convolution feature coding is performed based on the physical prior and the dynamic spatio-temporal graph, including designing a feature encoder that propagates stably across scales on the dynamic graph, wherein the relationship vector of edge i-j at time t is , the relationship vector is mapped to the convolution kernel parameter of the edge through a small kernel generator MLP K , and the dynamic adjacency A tThe edge-conditioned aggregation on the upper layer is represented as:

[0029] ,

[0030] where σ(·) is a nonlinear activation function, A t denotes the dynamic adjacency matrix, B (l) is a self-connection transformation; then the spectral domain learnable filtering and the time domain attention convolution are performed, wherein the polynomial approximation of the Laplacian spectral filtering is adopted in the frequency domain to avoid the eigen-decomposition of the Laplacian matrix, and high-order neighborhood information aggregation is realized; in the time domain, the attention-based time sequence representation is calculated for a time window H t - Tw+1 ,…, H t to capture the diffusion time delay and speed information, and finally the two representations obtained in the spectral domain and the time domain are fused through a learnable gate t :

[0031] ,

[0032] where pool(·) represents a global pooling operation, U represents a parameter matrix, σ is a sigmoid function, and ⊙ is an element-wise multiplication.

[0033] Further, the adaptive graph convolution feature encoding based on the physical prior and the dynamic spatio-temporal graph further comprises constructing a multi-scale propagation for modeling the regional diffusion mode while preserving the pixel-level boundary details, wherein the edge-conditioned convolution and the spectral-time domain filtering are applied to obtain a fine-scale representation ; a plurality of adjacent and dynamically similar pixels are aggregated into super nodes, and a learnable allocation matrix S is used, each row s i represents a soft coefficient of the pixel i allocated to M super nodes, satisfying row normalization, and the coarse-scale representation and adjacency are obtained by aggregation using S; the fine-scale and the coarse-scale are aggregated to obtain a final representation with both fine details and coarse-scale consistency, and the uncertainly weighted fusion is performed, and there are observation noise differences and physical interpretability differences between different pixels, so as to avoid the interference of high-noise pixels on the overall loss in the training process, and the loss function is set as:

[0034] ,

[0035] wherein represents the uncertainty probability parameter of the i-th pixel, represents the spectral consistency loss, represents the abundance constraint loss, represents the prediction task loss.

[0036] Further, the physical prior-based and time-series Transformer model is used to model the time-series characteristics of the multi-modal features, including a physical prior-guided time-series Transformer, which realizes high-precision modeling of the whole process of green tide by explicitly introducing spectral prior, chlorophyll concentration dynamic characteristics and ocean dynamics gradient constraints in the self-attention mechanism, wherein the self-attention mechanism based on the physical prior guide improves the modeling ability of the green tide life cycle by explicitly introducing the chlorophyll concentration estimation and spectral gradient characteristics as a bias term in the attention calculation, and introduces a physical prior bias B phy , to obtain an improved formula; after the physical prior-guided Transformer encoding of the layer, a deep feature representation of the time series is obtained, and finally a residual correction mechanism based on the physical prior is introduced to calculate the residual between the predicted spectrum and the endmember mixed reconstruction spectrum by a linear mixing model: L

[0037] ,

[0038] wherein R phy represents a physical prior residual term, S represents a predicted spectrum, k represents the total number of spectral bands, k represents a physical weight coefficient; the residual is mapped back to the deep feature space for dynamic correction:

[0039] ,

[0040] wherein a is a learnable parameter, finally, the model obtains the features and the features based on the time-series Transformer and the adaptive convolution to predict the whole process of the occurrence, development and disappearance of the multi-time green tide.

[0041] Further, the green tide prediction is performed by fusing the adaptive graph convolution feature encoding and the time-series modeling results, including fusing the feature vectors generated by the graph adaptive convolution and the physical prior-guided time-series Transformer, combining the spatial diffusion and time-series evolution characteristics, and jointly modeling in the probability space, wherein the graph adaptive convolution network is responsible for modeling the spatial diffusion process of the green tide, and the PP-Transformer extracts the evolution law on the long time series, wherein the spatial diffusion prediction extracted by the graph adaptive convolution network is denoted as , the output time-series prediction is denoted as , and the fusion prediction probability map P f is denoted as:

[0042] ,

[0043] ​wherein, a represents an adaptive fusion coefficient, σ(·) is a Sigmoid activation function, H and W represent the height and width of the remote sensing image respectively, and the local spatial diffusion pattern and global temporal evolution trend of green tide are captured by multi-source result fusion.

[0044] Further, the marine green tide prediction by fusing the adaptive graph convolution feature encoding and the temporal modeling result further comprises introducing an adaptive prediction correction method based on prediction uncertainty, by dynamically adjusting the contribution degree of different models in the high uncertainty area, first, the prediction variance of the graph adaptive convolution network and the physically prior guided temporal Transformer at each pixel position is calculated respectively, to quantify the uncertainty of the prediction value in multiple sampling or multi-model prediction, and the weight is adaptively allocated according to the uncertainty, and finally the prediction map after adaptive correction is obtained:

[0045] ,

[0046] wherein, P (A) represents the original prediction probability output by the graph adaptive convolution network, P (T) represents the original prediction probability output by the physically prior guided temporal Transformer, P f represents the fusion prediction probability map; a green tide risk index is introduced to comprehensively consider the green tide coverage area, chlorophyll concentration gradient and historical evolution mode, to realize dynamic risk grading and early warning, wherein the green tide coverage mask M is determined by the threshold θ:

[0047] ,

[0048] wherein, M ij represents the green tide coverage label of the pixel point (i,j), P c (i,j) represents the value of the fusion prediction probability map at the pixel (i,j), and the green tide risk index is defined as:

[0049] ,

[0050] wherein, λ1, λ2, λ3 represent the weighting coefficients of the risk index, A represents the green tide coverage area, ∇C represents the chlorophyll concentration gradient, H represents the historical evolution trend factor.

[0051] In a second aspect, a marine green tide monitoring system for remote sensing images fusing physical prior and spatio-temporal evolution comprises:

[0052] A data acquisition module configured to acquire multi-modal remote sensing monitoring images;

[0053] The feature extraction module is configured to perform multi-modal feature extraction on the acquired image, including spectral reflectance feature extraction, ocean dynamics feature extraction, and feature alignment and unified representation.

[0054] The physical prior module is configured to establish a physical prior of a green tide characteristic wave band by using an ocean optical radiation transmission model.

[0055] The space-time module is configured to construct a dynamic space-time graph based on the extracted multi-modal features, and obtain a node global feature vector and a dynamic adjacency matrix.

[0056] The encoding module is configured to perform adaptive graph convolution feature encoding based on the physical prior and the dynamic space-time graph.

[0057] The modeling module is configured to perform time series modeling on the multi-modal features based on the physical prior and the time series Transformer.

[0058] The prediction module is configured to perform ocean green tide prediction by fusing the adaptive graph convolution feature encoding and the time series modeling result; and output the prediction result.

[0059] In a third aspect, the present application provides a computer readable storage medium, wherein a plurality of instructions are stored, the instructions being suitable for being loaded and executed by a processor of a terminal device.

[0060] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer readable storage medium, the processor being used to implement various instructions; and the computer readable storage medium being used to store a plurality of instructions, the instructions being suitable for being loaded and executed by the processor to perform the ocean green tide monitoring method of the present application.

[0061] In summary, the present application has the following beneficial technical effects:

[0062] By fusing multispectral imagery from multiple platforms, including satellites and UAVs, with marine dynamic data, and combining atmospheric correction and band resampling, the system achieves consistent processing and high-precision extraction of multi-source features, significantly improving the comprehensiveness and reliability of green tide feature recognition. It introduces a marine optical radiative transfer model, constructing physical priors such as red band absorption valleys, near-infrared reflection peaks, and green band enhancement. Through spectral shape consistency loss and linear mixture abundance constraints, physical knowledge is embedded into a deep model, effectively enhancing detection robustness across scenes and time phases. Simultaneously, based on an adaptive graph convolutional network and a physical prior-guided temporal Transformer, the system can dynamically model the spatiotemporal evolution characteristics of green tides, achieving accurate predictions of the occurrence, diffusion, and dissipation of green tides. Combined with risk index modeling, it provides high-confidence green tide detection results and early warning information, offering strong support for marine ecological environment monitoring and emergency decision-making. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of a remote sensing image-based method for monitoring ocean green tides that integrates physical priors and spatiotemporal evolution, according to Embodiment 1 of the present invention.

[0064] Figure 2 This is a schematic diagram comparing the ACC and F1 of various models for green tide anomaly detection in Embodiment 1 of the present invention;

[0065] Figure 3 This is a schematic diagram comparing the accuracy of various models under different Dropout ratios for green tide anomaly detection in Embodiment 1 of the present invention;

[0066] Figure 4 This is a thermal schematic diagram of the spatial distribution of green tide anomaly detection according to Embodiment 1 of the present invention. Detailed Implementation

[0067] The present invention will be further described in detail below with reference to the accompanying drawings.

[0068] Example 1

[0069] Reference Figure 1 This embodiment of a remote sensing image-based method for monitoring ocean green tides that integrates physical priors and spatiotemporal evolution includes:

[0070] S1. Multimodal Feature Extraction Module

[0071] In the process of monitoring green tide, the remote sensing images collected by different platforms (satellite, unmanned aerial vehicle, etc.) have significant differences in observation angle, spatial resolution, and band range. In order to achieve high-precision characterization of green tide distribution, it is necessary to extract water reflectivity, vegetation index, and marine dynamics from multi-platform, multi-spectral, and multi-temporal remote sensing images, and to unify their representation. The core goal of this module is to improve the robustness and generalization ability of green tide detection by fusing multi-dimensional features such as spectrum, space, and dynamics, and to provide a reliable feature basis for subsequent physical prior modeling and spatio-temporal evolution prediction. Specifically, it includes the following three parts:

[0072] 1) Spectral reflectance feature extraction: Different bands in multi-spectral remote sensing data have significant sensitivity to green tide biomass changes. Let the multi-band original radiation value obtained from satellite or unmanned aerial vehicle images be R raw . First, perform atmospheric correction and band resampling to obtain the apparent water reflectivity R(λ):

[0073] ,

[0074] wherein L (λ) represents the band radiance received by the sensor, d represents the distance correction factor, E s ( lambda ) represents the solar irradiance at the top of the atmosphere; theta s represents the solar zenith angle. Water reflectivity R includes all the original information of the bands. For green tide sensitive bands (blue, green, red, near-infrared), key vegetation indices NDVI , NDWI such as

[0075] ,

[0076] are extracted, where R NIR , R RED , R GREEN are the near-infrared, red, and green band reflectances, respectively.

[0077] 2) Marine dynamics feature extraction: The occurrence and spread of green tide are strongly driven by marine dynamic environment, including sea surface temperature, ocean current velocity, wind field, and other parameters. Let the original marine dynamics data be R d . Using interpolation and re-projection methods, define the marine dynamics feature vector as:

[0078] ,

[0079] whereinT s representing sea surface temperature, V c , theta c representing ocean current velocity and direction, V w , theta w representing wind speed and direction. To dynamically depict the trend of green tide changing with dynamic conditions, the time series gradient is calculated using the multi-temporal sequence {X d t}.

[0080] ,

[0081] 3) Feature alignment and unified representation. Since satellite and UAV images differ in spatial resolution, spectral range, observation angle, etc., directly fusing features will lead to a decline in green tide detection performance. The present invention uses a cross-platform alignment method based on adversarial learning. The high-resolution image extracted by the UAV is extracted by the convolutional network feature, and the spectral feature obtained by splicing the vegetation index NDVI and NDWI is X u . The low-resolution image extracted by the satellite is extracted by the convolutional network feature, and the spectral feature obtained by splicing the vegetation index NDVI and NDWI is X s . The multi-platform feature is unified by minimizing the alignment loss function:

[0082] ,

[0083] wherein, f θ is the feature mapping function based on the convolution-Transformer hybrid encoder. The multi-source features are uniformly mapped to a common feature space to obtain a global feature vector:

[0084] ,

[0085] The feature vector contains spectral information and also integrates ocean dynamics prior, providing a unified input for subsequent physical constraint modeling and spatio-temporal evolution prediction.

[0086] S2. Physical prior modeling and constraint module

[0087] In green tide detection, a pure data-driven deep learning model is vulnerable to sensor noise, observation condition changes, and multi-scene differences, resulting in insufficient generalization ability. To this end, this module introduces a marine optical radiation transfer model (RTM) to establish physical priors in the green tide feature band and convert them into physical constraint terms for the deep detection network to guide the subsequent deep learning model training. Specifically, it includes the following three parts:

[0088] 1) Physical prior spectral modeling. Green tide shows significant spectral characteristics in water remote sensing reflectance curves, with a significant absorption valley in the red band, a clear reflection peak in the near-infrared band, and enhanced reflection characteristics in the green band. To embed this physical prior of spectral shape in the deep model, we introduce a marine optical radiation transfer model (RTM):

[0089] ,

[0090] where θs represents the solar incident angle, θv represents the sensor observation angle, C chl represents the chlorophyll concentration, C CDOM represents the colored dissolved organic matter concentration, C TSM represents the suspended matter concentration, represents the geometric correction factor, which is used to correct the influence of the sun and sensor angles on reflectance. a(λ) represents the total absorption coefficient, a water (λ), a CDOM (λ), a TSM (λ) represents the specific absorption coefficient of the corresponding component, and b(λ) represents the backscattering coefficient, which is determined by the water body itself.

[0091] Define the theoretical reflectance vector of the green tide pixel in different bands R phy :

[0092] ,

[0093] where, f RTM represents the marine optical radiation transfer model.

[0094] 2) Spectral shape consistency constraint. Based on the marine optical radiation transfer model (RTM), the theoretical prior feature of the typical green tide spectrum is constructed, and the theoretical prior curve is constrained with the model predicted spectrum during the deep model training process. By minimizing the difference in shape between the two, the prediction result still conforms to the physical law under different observation conditions. Spectral shape consistency loss:

[0095] ,

[0096] where, Rpred Represents the reflectance predicted by the depth model. R phy This represents the theoretical spectrum generated by RTM.

[0097] 3) Linear Mixture Abundance Constraint: In spectral remote sensing, an endmember refers to the most representative spectral component, and endmember abundance refers to the proportion of a certain type of endmember in a pixel. Based on endmember theory, the linear mixture abundance constraint represents the reflectance of each pixel as a linear combination of the reflectance of multiple endmember types.

[0098] ,

[0099] in, a i Represents the abundance of the end-point, E i Represents end-member reflectivity. K Represents the number of endmembers.

[0100] The theoretical reflectivity curves of different endmembers were obtained based on the marine optical radiative transfer model, and the predicted prior abundance distribution was obtained by inversion using a linear mixture model. a phy :

[0101] ,

[0102] Among them, R obs E represents the reflectance vector of a pixel in different spectral bands. RTM The endmember theoretical spectral matrix representing RTM generation, a=[a1,a2,a3,…,a…] k ] represents the endmember abundance vector of a pixel, which is a K-dimensional vector, where each a k It represents the proportion of a certain type of endmember in that pixel.

[0103] In deep model training, consistency constraints are imposed between the predicted results and the theoretical combination results, and a linear mixture abundance constraint loss is introduced:

[0104] ,

[0105] in, a phy This represents the theoretical abundance distribution obtained by inversion from a linear mixture model. a pred These represent the theoretical abundance distribution obtained from depth prediction. In the subsequent deep green tide detection network, these two constraints are used as regularization terms and are related to the main task loss. L cls Let's work together to optimize.

[0106] S3. Dynamic Spatiotemporal Graph Construction Module

[0107] In the multispectral remote sensing of ocean monitoring, the occurrence and evolution of green tide are driven by multiple factors, including geographical proximity, ocean current transport, wind field effect, and historical outbreak rules. It is difficult to comprehensively depict the dynamic diffusion process of green tide by relying on single-time satellite images or local observation data. Therefore, the core goal of this module is to construct a dynamic spatio-temporal graph that can dynamically reflect the spatial diffusion relationship and temporal evolution rule by taking remote sensing pixels or regions as graph nodes, and provide graph structure prior constraints for subsequent spatio-temporal depth models.

[0108] 1) Spatial adjacency relationship modeling. In the scenario of ocean remote sensing monitoring, the geographical adjacency determines the possibility of spatial diffusion of green tide. Let the remote sensing image be divided into N nodes, and the geographical coordinates of each node be P i =( lon i , lat i ), and the Euclidean distance between nodes be d ij =∥p i -p j ∥. First, define the spatial similarity weight between nodes i and j according to the Gaussian kernel function:

[0109] ,

[0110] where is the scale parameter for controlling the range of adjacent action. When two nodes are blocked by coastlines, islands, and other topography, the mask matrix B ij is used for shielding:

[0111] ,

[0112] where B ij =0 indicates that there is physical obstruction, B ij =1 indicates that the open sea is connected.

[0113] 2) Dynamic driven feature fusion. The diffusion of green tide not only depends on the spatial distance, but also is significantly affected by the ocean dynamic process, including the ocean current field and the wind field. Let the node i at time t be the ocean current velocity vector u t (i) , the wind velocity vector be w t (i) , and the displacement vector of node i pointing to j be r ij =p j -p iOcean current driven weight calculation formula:

[0114] ,

[0115] where λ curr controls the attenuation effect of ocean current on distance.

[0116] Wind field driven weight calculation formula:

[0117] ,

[0118] where λ wind controls the driving range of wind direction.

[0119] Weighted fusion of ocean current and wind field effects:

[0120] ,

[0121] where, β , gamma is the adjustable weight.

[0122] 3) Time evolution prior modeling: The occurrence, diffusion and extinction of green tide have significant time sequence rules, so the time evolution prior graph is constructed using historical outbreak data, and the time dependence of dynamic adjacency matrix is introduced. Let the historical multi-time green tide distribution sequence be { Y 1, Y 2, …, Y T}, and the state transition frequency between nodes is N i→j , the time transition probability matrix is obtained:

[0123] ,

[0124] To adapt to the needs of sudden events and real-time updates, exponential smoothing is used:

[0125]

[0126] where, represents the transition statistics based on the latest window, and η ∈ (0, 1) is the smoothing coefficient.

[0127] 4) Dynamic adjacency matrix generation: The dynamic adjacency matrix at time t is constructed by integrating geographical similarity, dynamic driving and historical prior:

[0128] ,

[0129] where α, β, δ are learnable or preset weights, A geo represents the spatial adjacency relationship matrix, and A dyndenotes a dynamics-driven feature matrix, P t denotes a temporal transition probability matrix. Normalization is performed as

[0130]

[0131] where, D t denotes a diagonal matrix, and the output {A t} is a dynamic spatio-temporal graph sequence, which serves as a graph structure prior for subsequent deep spatio-temporal modeling.

[0132] S4. Adaptive graph convolutional feature encoding module

[0133] After completing multi-modal feature extraction and obtaining node global feature vectors and dynamic adjacency matrix {At} construction, the goal of the adaptive graph convolutional feature encoding module is to design a physically interpretable, spatio-temporally consistent, and cross-scale robust propagation feature encoder on dynamic graphs, providing high-quality node representations for subsequent green tide detection and time series prediction.

[0134] 1) Set edge conditioning kernel, dynamic adjacency matrix has provided relationship description for each edge (such as geographical proximity, ocean current and wind field projection, historical transition probability, etc.). Let the relationship vector of edge i-j at time t be:

[0135]

[0136] where, denotes a spatial adjacency relationship, denotes a dynamics-driven feature, denotes a temporal evolution relationship. Through a small kernel generator MLP K map the relationship vector to the convolution kernel parameters of the edge:

[0137]

[0138] where, denotes a weight parameter. On the dynamic adjacency A t perform edge-conditioned aggregation (including self-connection residual term):

[0139]

[0140] where, σ(·) is a nonlinear activation function, A t denotes a dynamic adjacency matrix, B (l) is a self-connection transformation. This mechanism allows different physical situations to have differentiated propagation kernels.

[0141] ​​​​​2) Spectrum-time integration: The goal of spectrum-time integration is to suppress high-frequency noise and improve large-scale consistency in the spatial spectrum domain, and to capture the dependence of green tide diffusion speed and life cycle in the time domain, so that spatial propagation and time evolution complement each other. First, perform spectrum domain learnable filtering, then perform time domain attention / convolution, and finally fuse the two through learnable gating to form unified representation. In the spectral domain, we use a polynomial approximation of the Laplacian spectral filter to avoid eigenvalue decomposition of the Laplacian matrix, thereby efficiently implementing high-order neighborhood information aggregation:

[0142] ,

[0143] where, h t is the node feature matrix at time t, is the normalized graph Laplacian, T k (·) is the k th Chebyshev polynomial, which controls the receptive field size of spectral filtering, theta k (t) is the learnable spectral filtering coefficient (which can be fine-tuned over time to adapt to non-stationarity). In the time domain, we calculate the attention-based time sequence representation for a time window { H t - Tw+1 ,…, H t} to capture the time lag and speed information of diffusion:

[0144] ,

[0145] where, W Q , W K , W V is a linear mapping matrix, d q is the dimension scaling factor of the key / query, T w is the length of the time window. The two representations obtained by the spectral domain and the time domain are fused through learnable gating γ t :

[0146] ,

[0147] where pool(·) represents the global pooling operation, U represents the parameter matrix, σ is the sigmoid, and is the element-wise multiplication. The gating can be a scalar, a vector, or a node / channel-specific vector, thereby flexibly determining the weight distribution of the spectral domain and the time domain.

[0148] 3) Multi-scale propagation: Multi-scale propagation is used to model region-level (super-pixel) diffusion patterns while preserving pixel-level boundary details, thus balancing local finesse with global consistency. It is implemented in three steps: fine-scale encoding, coarse-scale convergence, and coarse-fine fusion. Edge-conditional convolution and spectral-temporal filtering are applied to obtain the fine-scale representation. This step primarily handles the fine-grained characterization of local boundaries. To obtain region-level semantics, several neighboring and dynamically similar cells are aggregated into supernodes, using a learnable assignment matrix S, where each row s... i Represents a pixel i The soft coefficients assigned to M supernodes satisfy row normalization (softmax). S-aggregation yields the coarse-scale representation and adjacency:

[0149] ,

[0150] By aggregating the fine and coarse scales, a final representation is obtained that retains both detail and coarse-scale consistency.

[0151] ,

[0152] 4) Uncertainty-weighted fusion: Different pixels exhibit differences in observation noise and physical interpretability. To avoid interference from high-noise pixels on the overall loss, and considering spectral consistency and abundance constraints, the following loss function is set during training:

[0153] ,

[0154] in, Representing the i The uncertainty probability parameters of each pixel are learned by the model. Represents spectral uniformity loss, Represents abundance constraint loss, This represents the predicted task loss.

[0155] S5. Physics-Prior-Guided Timing Transformer Module

[0156] In the occurrence, diffusion and extinction process of marine green tide, its spatio-temporal evolution is influenced by multiple factors, including sea surface temperature (SST), wind speed, tidal current speed, ocean vortex dynamics, chlorophyll concentration changes, etc. Multi-temporal and multi-spectral remote sensing observations can provide rich spectral, radiation and dynamic characteristics. However, due to the stage, non-stationary and time-dependent nature of the life cycle of green tide, it is often difficult to accurately predict the occurrence and development process of green tide by relying solely on static feature extraction. Therefore, this module proposes a physically prior-guided time series Transformer, which explicitly introduces spectral priors, chlorophyll concentration dynamic features and ocean dynamics gradient constraints into the self-attention mechanism, achieving high-precision modeling of the entire process of green tide.

[0157] Introducing physical prior bias into time series self-attention, traditional Transformers ignore the unique spectral and dynamic evolution rules of green tide when modeling long sequence dependencies. We propose a physically prior-guided self-attention mechanism that explicitly introduces chlorophyll concentration estimates and spectral gradient features as bias terms in attention calculation to improve the modeling capability of the life cycle of green tide. Introducing physical prior bias B phy , the improved formula is:

[0158] ,

[0159] where B phy is composed of chlorophyll concentration difference bias and spectral reflectance gradient bias, Q, K, V represent query, key and value matrices, d represent the dimensions.

[0160] Multi-stage prediction and residual correction, after the L layer physically prior-guided Transformer encoding, we get the time series deep feature representation:

[0161] ,

[0162] To further improve the prediction accuracy, we introduce a residual correction mechanism based on physical priors. First, we calculate the residual between the predicted spectrum and the endmember mixed reconstruction spectrum using a linear mixing model:

[0163] ,

[0164] where R phy represents the physical prior residual term, S represents the predicted spectrum, k represents the total number of spectral bands, k represents the physical weight coefficient. Then map the residual back to the deep feature space for dynamic correction:

[0165] ,

[0166] where a is a learnable parameter. Finally, the model gets the features from the time-sequential Transformer and the adaptive convolution to predict the whole process of green tide occurrence, development and disappearance.

[0167] S6. Green tide detection and warning module

[0168] In the green tide prediction and warning, due to the noise of multi-source remote sensing data, cloud cover and ocean dynamics uncertainty, single model prediction has limitations. This paper proposes a green tide detection and warning module, which realizes high-precision prediction through multi-source prediction result fusion, uncertainty perception adaptive prediction correction and green tide risk index modeling and warning. First, the feature vectors generated by the adaptive convolution of the graph and the time-sequential Transformer guided by physical prior are fused, combined with spatial diffusion and time evolution features, and jointly modeled in the probability space. Second, uncertainty perception correction is introduced to improve the prediction robustness and physical rationality. Finally, based on the fusion results, a green tide risk index model is built to generate risk levels and warning information by integrating coverage area, chlorophyll concentration gradient and historical evolution pattern.

[0169] 1) Multi-source prediction result fusion

[0170] In order to fully utilize the complementarity of different models in feature extraction, a multi-source prediction fusion method based on graph adaptive convolution network and time-sequential Transformer guided by physical prior is proposed. The graph adaptive convolution network is responsible for modeling the spatial diffusion process of green tide, and the PP-Transformer extracts the evolution law in long time series. Finally, joint prediction modeling is performed in the probability space. Let the spatial diffusion prediction extracted by the graph adaptive convolution network be , the output time-sequential prediction be , and the fusion prediction probability map P f be represented as:

[0171] ,

[0172] where a represents the adaptive fusion coefficient, which is dynamically learned by the prediction accuracy on the validation set, and σ(·) is the Sigmoid activation function that maps the fusion result to the probability space. H and W represent the height and width of the remote sensing image, respectively. Through multi-source result fusion, the local spatial diffusion pattern and global time evolution trend of green tide can be captured simultaneously, significantly improving the overall prediction accuracy and generalization ability.

[0173] 2) Uncertainty perception adaptive prediction correction

[0174] Remote sensing data will introduce large uncertainty under cloud cover, observation missing and physical process complexity, which leads to the deviation of prediction results in some local areas. Therefore, an adaptive prediction correction method based on prediction uncertainty is introduced to improve the robustness of prediction by dynamically adjusting the contribution of different models in high uncertainty areas. First, the prediction variance of graph adaptive convolution network and physically prior guided time series Transformer at each pixel position is calculated respectively:

[0175] ,

[0176] where Var(·) represents the variance operator, which is used to quantify the uncertainty of the prediction value in multiple sampling or multi-model prediction. According to the uncertainty, the weight is adaptively allocated:

[0177] ,

[0178] where w (A) represents the weight of graph adaptive convolution network model in the final fusion prediction. w (T) represents the weight of physically prior guided time series Transformer model in the final fusion prediction. U (A) , U (T) represents the prediction variance of the corresponding model, and the greater the uncertainty, the lower the weight.

[0179] The final prediction map after adaptive correction is obtained:

[0180] ,

[0181] where P (A) represents the original prediction probability output by the graph adaptive convolution network, P (T) represents the original prediction probability output by the physically prior guided time series Transformer, P f represents the fusion prediction probability map.

[0182] 3) Modeling and early warning of green tide risk index

[0183] After obtaining the high-precision green tide prediction map Pc, the green tide risk index is further introduced to comprehensively consider the green tide coverage area, chlorophyll concentration gradient and historical evolution mode, and realize dynamic risk classification and early warning. Let the prediction probability map be P c , and the green tide coverage mask M is determined by the threshold θ:

[0184] ,

[0185] where M ijGreen tide coverage label P representing the pixel point of position (i, j) c (i, j) represents the value of the fusion prediction probability map at pixel (i, j), and θ represents the prediction probability threshold.

[0186] The green tide risk index is defined as:

[0187] ,

[0188] where λ1, λ2, and λ3 represent the weighted coefficients of the risk index, which are learned through the validation set, A represents the green tide coverage area (statistically obtained through the mask M), ∇C represents the chlorophyll concentration gradient, reflecting the trend of green tide intensity change, H representing the historical evolution trend factor. According to the numerical range of GCRI, different risk levels are set.

[0189] Experimental verification:

[0190] To verify the effectiveness of the green tide spatio-temporal modeling and early warning system proposed in this study based on multi-source remote sensing, unmanned aerial vehicle images, and physical prior modeling, this paper constructs an experimental platform for field scene simulation and multi-source data driving in a typical nearshore green tide high-risk sea area. Satellite remote sensing images, low-altitude unmanned aerial vehicle high-resolution images, and multi-modal data such as temperature, salinity, dissolved oxygen, pH value, and chlorophyll concentration from ocean dynamic observation buoys, totaling 12,500 samples, fully cover the green tide evolution period (occurrence-diffusion-development-remission), and have rich spatial and temporal dynamic characteristics. To enhance the adaptability of the experiment to the actual green tide monitoring and emergency early warning needs, complex environmental variables such as typhoon disturbance, cloud cover, remote sensing image quality degradation, ocean current mutation, and partial monitoring node data missing are introduced in the design process to comprehensively test the robustness and generalization ability of the model under high uncertainty scenarios.

[0191] The comparative method selects the current representative green tide detection and prediction models, including the basic Transformer, the green tide detection and prediction model GCN-LSTM combined with graph convolution and time series modeling, ASTGCN combined with graph attention mechanism, ST-Transformer based on spatio-temporal attention mechanism, Cross-ModalMatching Transformer (CMMT) based on cross-modal feature alignment, and the AGCN-PPTransformer joint modeling method proposed in this paper. All models are compared under the condition of uniform data set division (training set: validation set: test set = 6:2:2), consistent optimization strategy and training rounds to ensure the fairness and comparability of the evaluation results.

[0192] The performance evaluation indicators include accuracy (ACC), F1-Score, spatial positioning error, early warning lead time, false positive rate, and model inference delay. The model is evaluated from three dimensions: detection accuracy, timeliness, and deployment feasibility. The experimental results are shown in Figures 2, 3, and Table 1. The proposed method significantly outperforms the comparative models in all indicators, fully verifying its effectiveness and advantages in multi-modal feature fusion, physical prior modeling, and green tide evolution trend prediction tasks.

[0193] Table 1 Comparison of data of different methods in six indicators

[0194] Model name ACC F1 Early warning lead time Positioning error False alarm rate Inference delay Transformer 81.9% 82.3% 1.6 days 11.4 km 9.8% 2.5 min ASTGCN 86.4% 85.0% 2.1 days 10.5 km 8.1% 3.2 min GCN-LSTM 85.5% 84.1% 1.8 days 13.5 km 8.9% 4.0 min CMMT 87.4% 86.8% 2.4 days 11.7 km 7.2% 4.6 min ST-Transformer 83.2% 80.7% 1.5 days 12.1 km 10.4% 3.0 min The method of the present application 91.5% 89.9% 3.1 days 6.3 km 5.1% 3.4 min

[0195] From Figure 2 , Figure 3 and Table 1, it can be seen that the traditional Transformer, ASTGCN, MMFN, CMMT and ConvLSTM methods have certain prediction ability in the green tide anomaly detection task, but there are still obvious shortcomings in the key performance dimensions. Transformer has certain advantages in global modeling, which can model long-term dependencies in time series, but due to its weak ability to model regional spatial details, the positioning error is large, and the false positive rate is as high as 9.8%. ASTGCN introduces a graph structure modeling approach, which better integrates time series and spatial factor information, and outperforms Transformer in early warning lead time and accuracy, but its static graph structure is difficult to adapt to the dynamic nature of the rapid evolution of green tide. GCN-LSTM and CMMT improve prediction accuracy through time series mechanism and multi-modal fusion mechanism, with F1 values of 84.1% and 86.8%, but the fusion method is relatively shallow, lacking cross-modal consistency alignment and semantic enhancement, resulting in a high false positive rate and a significant increase in inference delay. Transformer has certain advantages in time series modeling, but due to the lack of explicit spatial structure expression, the prediction accuracy is low, with the lowest ACC and F1 values (81.9% and 82.3%).

[0196] In contrast, the method of the present application is based on multi-source multi-spectral feature extraction, physical prior constraint, dynamic spatio-temporal graph modeling, adaptive graph convolution feature coding and physical prior guided time series Transformer, etc. key technologies, fully integrating satellite remote sensing images, unmanned aerial high-resolution images and marine dynamic observation and other multi-source information, realizing higher level spatial consistency, temporal sensitivity and physical prior constraint modeling capability. In the experiment, the method is superior to other comparative methods in six evaluation indexes: the accuracy reaches 91.5%, the F1 value reaches 89.9%, the green tide early warning lead (3.1 days) and the spatial positioning accuracy (error 6.3 km) are all significantly ahead, the false positive rate is reduced to 5.1%, and the reasoning delay is controlled within 3.4 minutes. The above results fully verify the practicability of the method in green tide detection, boundary identification and evolution trend prediction, and have good engineering deployment prospect and marine emergency response value.

[0197] In order to verify the robustness and generalization ability of the model under different data Dropout ratios, the system tests the accuracy changes of multiple models and presents the results as shown in Figure 3 The figure shows the trend of the accuracy curve of each model with the change of Dropout ratio (0% - 40%). It can be seen that the accuracy curve corresponding to the method of the present application is always above other comparative models, can maintain a high performance level even at a high Dropout ratio, especially when the Dropout ratio is 40, the accuracy is still high (about 78%), and the downward trend is relatively flat during the whole change process. The experimental results show that the method of the present application performs excellently in resisting overfitting and maintaining prediction stability, which helps to improve the tolerance of the model to data noise and reliability in practical applications.

[0198] In order to verify the spatial anomaly distribution perception ability of the model, the system constructs a spatial heat map of the probability of green tide occurrence, and the results are shown in Figure 4 The different color depths in the figure correspond to the probability of green tide occurrence in different monitoring areas, and the deeper the color, the higher the risk of the area. It can be seen that the abnormal distribution predicted by the method of the present application has obvious spatial aggregation, and the high-risk areas are concentrated in the nearshore sea area affected by human activities, the spatial boundary is clear, and the change trend conforms to the historical monitoring data. The results show that the method not only has high prediction accuracy, but also can effectively locate the abnormal area of green tide in the spatial scale, providing fine reference basis for subsequent early warning and response measures.

[0199] Embodiment 2

[0200] The embodiment provides a remote sensing image marine green tide monitoring system fusing physical prior and spatio-temporal evolution, comprising:

[0201] a data acquisition module configured to acquire multi-modal remote sensing monitoring images;

[0202] a feature extraction module configured to perform multi-modal feature extraction on the acquired images, including spectral reflectance feature extraction, ocean dynamics feature extraction, and feature alignment and unified representation;

[0203] a physical prior module configured to establish a physical prior of a green tide characteristic wave band using an ocean optical radiation transfer model;

[0204] a spatio-temporal module configured to construct a dynamic spatio-temporal graph based on the extracted multi-modal features, to obtain a node global feature vector and a dynamic adjacency matrix;

[0205] an encoding module configured to perform adaptive graph convolution feature encoding based on the physical prior and the dynamic spatio-temporal graph;

[0206] a modeling module configured to perform time series modeling on the multi-modal features based on the physical prior and a time series Transformer;

[0207] a prediction module configured to perform marine green tide prediction by fusing the adaptive graph convolution feature encoding and the time series modeling results; and output the prediction results.

[0208] A computer-readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded and executed by a processor of a terminal device to implement the marine green tide monitoring method based on remote sensing images by fusing physical priors and spatio-temporal evolution.

[0209] A terminal device, comprising a processor and a computer-readable storage medium, the processor being configured to implement instructions, and the computer-readable storage medium being configured to store a plurality of instructions, the instructions being adapted to be loaded and executed by the processor to implement the marine green tide monitoring method based on remote sensing images by fusing physical priors and spatio-temporal evolution.

[0210] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, therefore: any equivalent changes made in the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for monitoring marine green tide in remote sensing images by fusing physical priori and spatio-temporal evolution, characterized in that, The method comprises the following steps: acquiring multi-modal remote sensing monitoring images; performing multi-modal feature extraction on the acquired images, including spectral reflectance feature extraction, ocean dynamics feature extraction, and feature alignment and unified representation, to obtain a global feature vector; establishing a physical prior of a green tide characteristic band by using an ocean optical radiation transfer model; constructing a dynamic spatio-temporal graph based on the extracted multi-modal features to obtain a dynamic adjacency matrix; performing adaptive graph convolution feature coding based on the physical prior and the dynamic spatio-temporal graph; performing time series modeling on the multi-modal features based on the physical prior and a time series Transformer; performing marine green tide prediction by fusing the adaptive graph convolution feature coding and the time series modeling results; outputting the prediction result. The multi-modal feature extraction on the acquired image includes spectral reflectance feature extraction, wherein the acquired multi-band original radiation value is R raw First, atmospheric correction and band resampling are performed to obtain apparent water reflectance R(λ). After obtaining the water reflectance, key vegetation indices are extracted for the green tide sensitive band. For marine dynamics feature extraction, the original marine dynamics data is R d An interpolation and re-projection method is used to calculate the marine dynamics feature vector. To dynamically depict the trend of the green tide changing with the dynamics condition, a multi-time sequence {X d is constructed using the marine dynamics feature vector X d t A cross-platform alignment method based on adversarial learning is used. The spectral feature obtained by splicing the vegetation coefficients NDVI and NDWI after convolution network feature extraction on the high-resolution image extracted by the unmanned aerial vehicle is X u The spectral feature obtained by splicing the vegetation coefficients NDVI and NDWI after convolution network feature extraction on the low-resolution image extracted by the satellite is X s Multi-platform feature uniformization is realized by minimizing the alignment loss function, multi-source features are uniformly mapped to a common representation space, and a global feature vector is obtained, wherein the global feature vector includes spectral information and marine dynamics prior. The physical prior of the green tide characteristic wave band is established by using a marine optical radiation transfer model, including introducing a marine optical radiation transfer model RTM, defining a theoretical reflectivity vector of a green tide pixel in different wave bands R phy Theoretical prior characteristics of a typical green tide spectrum are constructed based on the RTM, and in the depth model training process, the theoretical prior curve and the model predicted spectrum are constrained, the prediction result is still consistent with the physical law under different observation conditions by minimizing the difference in shape, and the spectral shape consistency loss is: ， wherein, R pred representative reflectance predicted by the depth model, R phy representative theoretical spectrum generated by RTM; based on the linear mixing abundance constraint of endmember theory, the reflectance of each pixel is represented as a linear combination of multi-class endmember reflectance; the theoretical reflectance curves of different endmembers and their prior abundance distribution are obtained according to the ocean optical radiation transfer model; then in the depth model training, the predicted result is consistent with the theoretical combination result, and the linear mixing abundance constraint loss is introduced: ， wherein, a phy represents the theoretical abundance distribution calculated from the linear mixture model inversion, a pred represents the theoretical abundance distribution predicted from the depth, i represents the endmember abundance of the pixel, E i represents the endmember reflectance, K represents the number of endmembers, represents the coefficient, KL represents the KL divergence; The extracted multi-modal feature is used to construct a dynamic space-time graph, including taking a remote sensing pixel as a graph node to construct a dynamic space-time graph dynamically reflecting spatial diffusion relationship and time evolution law, wherein the remote sensing image is divided into N nodes, the geographic coordinates of each node are P i =( lon i , lat i ), the Euclidean distance between nodes is d ij =∥p i -p j ∥, the spatial similarity weight between nodes i and j is defined according to the Gaussian kernel function, when two nodes are blocked by the coastline or island topography, the shielding is performed through a mask matrix B ij , the shielding of the blocked node pair is realized through element-level multiplication; the ocean current driving weight and the wind field driving weight are calculated based on the marine dynamic process, and the ocean current and wind field effects are weighted and fused, a time evolution prior graph is constructed using historical outbreak data, and finally the geographic similarity, dynamics driving and historical prior three factors are fused to construct a dynamic adjacency matrix at time t: ， wherein α, β, δ are preset weights, A geo represents a spatial adjacency relation matrix, A dyn represents a dynamics driving feature matrix, P t represents a time series transition probability matrix, and is normalized to obtain: ， where, D t represents a diagonal matrix, and the output {A t} is a dynamic spatio-temporal graph sequence, as a graph structure prior for deep spatio-temporal modeling; The adaptive graph convolution feature coding based on the physical prior and the dynamic space-time graph comprises a feature encoder which can be physically interpretable, space-time consistent and robustly propagated across scales on a dynamic graph, and a relationship vector of an edge i-j at a time t is calculated based on three relationships of geographical proximity, ocean current and wind field projection and historical transition probability A small nuclear generator MLP K The relationship vector is mapped to the convolution kernel parameter of the edge, and the edge conditioning aggregation is performed on the dynamic adjacency matrix A t , which is represented as ， where σ(·) is a nonlinear activation function, A t denotes the dynamic adjacency matrix, B (l) is a self-connection transform, denotes the convolution kernel parameters of the l-th layer, which is generated by a small kernel generator MLP K is generated; then spectral domain learnable filtering and time domain attention convolution are performed, wherein polynomial approximation of Laplacian spectral filtering is adopted in the spectral domain to avoid eigen-decomposition of the Laplacian matrix, high-order neighborhood information aggregation is realized to obtain spectral domain features ; in the time domain, attention-based time sequence representation is calculated for a time window , } to capture the diffusion time lag and speed information to obtain time domain features , and finally the two representations obtained in the spectral domain and the time domain are fused through a learnable gate γ t : ， where pool(·) represents a global pooling operation, U represents a parameter matrix, σ is a sigmoid, and is an element-wise multiplication, and respectively represent the spectral domain feature and the time domain feature; The adaptive graph convolution feature coding based on the physical prior and the dynamic space-time graph further comprises constructing a multi-scale propagation for modeling a regional level diffusion mode while preserving a pixel level boundary detail, wherein a boundary conditioned convolution and a spectral-time domain filtering are applied to obtain a fine scale representation , and a fine scale coding is realized; a plurality of adjacent and dynamically similar pixels are aggregated into super nodes, a learnable allocation matrix S is used, and each row s i represents a soft coefficient of the pixel allocation to M super nodes, and satisfies row normalization i , a coarse scale representation and adjacency are obtained by aggregation using S; and the fine scale and the coarse scale are aggregated to obtain a final representation which has both details and takes into account the consistency of the coarse scale, and a loss function is set during the training process: ， wherein, represents an uncertainty probability parameter of the i-th pixel, represents a spectral shape consistency loss, represents a linear mixture abundance constraint loss, represents a prediction task loss; The physical prior-based and timing Transformer models the timing of the multi-modal features, including a physical prior-guided timing Transformer, which realizes high-precision modeling of the whole process of green tide by explicitly introducing spectral prior, chlorophyll concentration dynamic characteristics and ocean dynamics gradient constraints in the self-attention mechanism, wherein the physical prior-guided self-attention mechanism improves the modeling capability of the green tide life cycle by explicitly introducing chlorophyll concentration estimation and spectral gradient characteristics as a bias term in attention calculation, and introduces a physical prior bias B phy The self-attention mechanism is improved: ， where B phy Q, K, V represent query, key and value matrix respectively, d represents dimension, and the features output by the adaptive graph convolution are used as the input of the Transformer L After the Transformer encoding guided by the physical prior of the layer, the deep feature representation of the time series is obtained, and finally the residual correction mechanism based on the physical prior is introduced. The residual between the predicted spectrum and the endmember mixed reconstruction spectrum is calculated by the linear mixing model: ， wherein R phy represents the physical prior residual term, S represents the predicted spectrum, K represents the total number of spectral bands, k represents the physical weight coefficient; the residual is mapped back to the deep feature space, dynamically corrected, and the features obtained based on the time series Transformer : ， wherein, a is a learnable parameter, H * representing that the deep feature representation is obtained based on the improved attention, and finally, the model obtains the feature based on the time sequence Transformer and the feature obtained by the adaptive graph convolution carrying out the whole process prediction of the occurrence, development and disappearance of the multi-temporal green tide The marine green tide prediction by fusing adaptive graph convolution feature coding and time series modeling results comprises fusing feature vectors generated by adaptive graph convolution and physical prior guided time series Transformer, combining spatial diffusion and time evolution features, and jointly modeling in a probability space, wherein the adaptive graph convolution network is responsible for modeling the diffusion process of the green tide in space, the physical prior guided time series Transformer extracts the evolution law in a long time sequence, the spatial diffusion prediction extracted by the adaptive graph convolution network is , the time series prediction output by the time series Transformer is , and the fusion prediction probability map P f is represented as: ， wherein, α represents an adaptive fusion coefficient, and σ(·) is a Sigmoid activation function, and respectively represent the feature vectors generated by the graph adaptive convolutional network and the physical prior guided time sequence Transformer, and the local spatial diffusion pattern and the global time sequence evolution trend of the green tide are captured by multi-source result fusion. The marine green tide prediction by fusing the adaptive graph convolution feature coding and the time series modeling results further comprises introducing an adaptive prediction correction method based on prediction uncertainty, by dynamically adjusting the contribution degree of different models in high uncertainty areas, first, the prediction variance of the adaptive graph convolution network and the time series Transformer guided by the physical prior at each pixel position is calculated, for quantifying the uncertainty of the prediction value in multiple sampling or multi-model prediction, the uncertainty is adaptively assigned a weight, and finally an adaptive corrected prediction map is obtained: ， where P (A) represents the original prediction probability output by the adaptive graph convolution network, P (T) represents the original prediction probability output by the physically prior guided time series Transformer, P f represents the fused prediction probability graph, w (A) represents the weight of the graph adaptive convolution network model in the final fused prediction, w (T) represents the weight of the physically prior guided time series Transformer model in the final fused prediction; a green tide risk index is introduced, which comprehensively considers the green tide coverage area, chlorophyll concentration gradient and historical evolution mode, realizes dynamic risk grading and early warning, wherein the green tide coverage mask M is determined by the threshold θ: ， where M ij a green tide coverage label, P c (i,j) represents the value of the fused prediction probability map at pixel (i,j), and the green tide risk index is defined as: ， Wherein, λ1, λ2, λ3 represent the weighted coefficients of risk index, A represents the green tide coverage area, and ∇C represents the chlorophyll concentration gradient, H represent the historical evolution trend factor.

2. A system for monitoring marine green tide from remote sensing images by combining physical priori and spatio-temporal evolution, which executes the method for monitoring marine green tide from remote sensing images by combining physical priori and spatio-temporal evolution as claimed in claim 1, characterized in that, The method comprises the following steps: a data acquisition module configured to acquire multi-modal remote sensing monitoring images; a feature extraction module configured to perform multi-modal feature extraction on the acquired images, including spectral reflectance feature extraction, ocean dynamics feature extraction, and feature alignment and unified representation; a physical prior module configured to establish a physical prior of a green tide characteristic band by using an ocean optical radiation transfer model; a spatio-temporal module configured to construct a dynamic spatio-temporal graph based on the extracted multi-modal features to obtain a dynamic adjacency matrix; an encoding module configured to perform adaptive graph convolution feature coding based on the physical prior and the dynamic spatio-temporal graph; a modeling module configured to perform time series modeling on the multi-modal features based on the physical prior and a time series Transformer; a prediction module configured to perform marine green tide prediction by fusing the adaptive graph convolution feature coding and the time series modeling results; and output the prediction result.

Images