Method for monitoring effect of marine ecological restoration in estuary and sea area based on multi-period remote sensing

Abstract

This invention discloses a method for dynamic monitoring of the effects of marine ecological restoration in estuaries based on multi-period remote sensing, relating to the field of marine ecological restoration technology. The method includes: acquiring multi-period satellite and UAV remote sensing images before and after the implementation of restoration projects, and constructing a long-term standardized remote sensing dataset; interpreting the multi-period remote sensing images to generate multi-period ecological element data; inputting the multi-period ecological element data into an intelligent simulation model of water and sediment transport and topographic evolution to simulate the patterns of water and sediment transport and topographic evolution at different restoration stages and evaluate the effectiveness of the project; constructing an ecological network topology map with each ecological patch in the restoration area as a node, and using a graph neural network to learn the spatiotemporal characteristics and predict the state evolution of multi-period ecological indicators to generate a dynamic change curve of the comprehensive restoration effect index; triggering a graded early warning when the comprehensive index deviates from the preset restoration target trajectory. This invention achieves quantitative monitoring and dynamic early warning of the effects of ecological restoration in estuaries.

Description

Technical Field

[0001] This invention relates to the field of marine ecological restoration technology, and more specifically, to a method for dynamic monitoring of the effects of marine ecological restoration in estuaries based on multi-period remote sensing. Background Technology

[0002] Estuaries, as key areas of land-sea interaction, possess unique ecosystem service functions, playing an irreplaceable role in maintaining biodiversity, purifying water quality, regulating floods, and providing bird habitats. Scientifically and accurately assessing the effectiveness of ecological restoration projects is crucial for ensuring restoration goals are achieved, optimizing project plans, and guiding subsequent management. Traditional ecological monitoring methods primarily rely on ground-based quadrat surveys and manual patrols. While these methods can acquire localized high-precision data, they have inherent limitations such as limited coverage, long cycles, high costs, and difficulty in simultaneously monitoring changes in multiple elements. For dynamic ecosystems like estuaries—vast areas with complex topography and alternating land and water zones—traditional methods struggle to meet the demands for large-scale, long-term, and high-frequency monitoring. Remote sensing technology, with its advantages of macroscopic, rapid, and repeatable observations, has been widely applied in areas such as land use change, vegetation cover retrieval, and water dynamics monitoring. In particular, with the continuous enrichment of remote sensing data sources, the combination of multispectral, hyperspectral, and radar satellite data with UAV low-altitude remote sensing provides multi-level and multi-scale data support for monitoring the ecological restoration of estuarine wetlands. However, the current application of remote sensing technology in ecological restoration monitoring is mostly limited to the comparison of changes in a single period or a single indicator. It lacks a systematic, continuous, and index-based dynamic monitoring method based on multi-period remote sensing data. It is difficult to fully reveal the synergistic change patterns of multiple factors such as landform evolution, hydrological connectivity, vegetation restoration, and improvement of bird habitats during the implementation of restoration projects, and it is also impossible to achieve quantitative assessment and trend prediction of restoration effects.

[0003] Therefore, it is urgent to establish a dynamic monitoring method for the marine ecological restoration effect in estuaries and sea areas based on multi-period remote sensing. This method should make full use of long-term multi-source remote sensing data, combined with ecological indicators and spatial analysis methods, to achieve continuous tracking and quantitative assessment of key elements such as topography, hydrological patterns, and vegetation cover in the restoration area. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a dynamic monitoring method for the ecological restoration effect of estuaries and sea areas based on multi-period remote sensing. This method enables quantitative monitoring and dynamic early warning of the ecological restoration effect throughout the entire process, providing a scientific basis for the adaptive management of restoration projects.

[0005] The first aspect of this invention provides a method for dynamic monitoring of the effects of marine ecological restoration in estuaries and sea areas based on multi-period remote sensing, comprising the following steps: Multiple satellite and UAV remote sensing images were acquired before and after the restoration project was implemented, and radiometric correction and geometric registration were performed to construct a long-term standardized remote sensing dataset. Interpret the multi-period remote sensing images in the aforementioned remote sensing dataset, extract the spatiotemporal distribution of water areas, tidal channel networks, vegetation cover areas, and artificial structures, and generate multi-period ecological element data based on the interpretation results; The multi-phase ecological element data are input into a pre-constructed intelligent simulation model of water and sediment transport and topographic evolution to simulate the water and sediment transport and topographic evolution patterns at different restoration stages. The effectiveness of hydrodynamic restoration and micro-topographic remediation projects is evaluated by dynamically fitting the model simulation results with actual remote sensing monitoring data. An ecological network topology map is constructed using each ecological patch in the restoration area as a node and ecological flow as an edge. A graph neural network model is used to learn the spatiotemporal features and predict the state evolution of preset multi-period ecological indicators, generating a dynamic change curve of the comprehensive index of restoration effect. When the overall index of the repair effect deviates from the preset repair target trajectory, a graded warning is triggered.

[0006] In this scheme, multiple remote sensing images from the aforementioned remote sensing dataset are interpreted to extract the spatiotemporal distribution of water areas, tidal channel networks, vegetation cover areas, and man-made structures. Specifically, this includes: A deep learning semantic segmentation network integrating multi-scale dilated convolution and deformable convolution is constructed. Multiple remote sensing images are input into the semantic segmentation network one after another, and the land cover classification probability map corresponding to each image is output to obtain the initial land cover classification results including water, vegetation and artificial structures. Instantaneous tide data at the time of imaging of each remote sensing image is obtained, a tide level-inundation range response relationship is established, the water area in the initial land cover classification result is calculated based on the response relationship, and the water area in the land cover classification result is updated through morphological dilation and erosion smoothing. The binary image of the water body region is extracted from the updated land cover classification results. The skeleton line is extracted from the binary image of the water body region. The tidal channel network topology map is constructed based on the intersection and endpoint of the skeleton line. The tidal channel topology map is used to learn features by graph convolutional network, the main channel and tributaries of each level of the tidal channel are identified, and the tidal channel density, bifurcation ratio and connectivity index are calculated. The vegetation cover area is extracted from the updated land cover classification results, a phenological feature template library of different vegetation types in a preset period is constructed, the temporal spectral curve of each pixel in multiple remote sensing images is obtained, the similarity between the temporal spectral curve and the phenological feature template library is measured, the vegetation type identification result is output, and the coverage of each vegetation type is inverted. Based on the multi-period interpretation results, a spatiotemporal evolution trend field of land cover categories is constructed. After filtering, the value is input into a Markov random field model for iterative optimization, and the interpretation results with spatiotemporal consistency correction are output.

[0007] In this scheme, multi-period ecological element data are generated based on the interpretation results, specifically including: Extract tidal channel network topology parameters, shoreline location temporal information, vegetation cover parameters, and dynamic change data of artificial structures from the interpretation results after spatiotemporal consistency correction. The spatial and temporal references of the tidal channel network topology parameter set, shoreline location time series dataset, vegetation cover parameter set, and dynamic change dataset of artificial structures are unified. All vector data are rasterized to a uniform spatial resolution and spatially resampled. Using the key nodes of ecological restoration projects as the basis for timeline division, remote sensing observation data with different sampling frequencies are aggregated to a unified time node to construct a standardized dataset of ecological elements with equal time intervals across multiple periods.

[0008] In this scheme, a pre-constructed intelligent simulation model of water and sediment transport and topographic evolution is built by integrating neural operators and neural symbol systems to simulate the patterns of water and sediment transport and topographic evolution at different restoration stages, specifically including: A Fourier neural operator architecture is constructed, feature mapping is performed by lifting operators, fast Fourier transform is performed on the feature field in the Fourier layer and global convolution kernel is learned in the frequency domain space, and the initial water and sediment transport prediction field is output through projection operators after inverse Fourier transform. For the initial water and sediment transport prediction field, a mass conservation residual field, a momentum conservation residual field, and a sediment continuity residual field are constructed as symbolic constraint terms and added to the loss function. The weights of each constraint term are dynamically adjusted according to the input characteristics, and the Fourier neural operator is fed back for iterative optimization to output the corrected water and sediment transport prediction field. A neural network is used to construct a constant-time differential equation to model the modified water and sediment transport prediction field, and the derivative function is parameterized by a neural network to output the prediction sequence of water and sediment transport state at future times. A multilayer perceptron network with spatiotemporal coordinates as input is constructed to fit multi-period topographic elevation data, outputting a continuous topographic field as the bed boundary condition of the Fourier neural operator, and the topographic erosion and deposition volume output by the Fourier neural operator is fed back to update the multilayer perceptron network.

[0009] In this scheme, the dynamic fitting of model simulation results with actual remote sensing monitoring data specifically includes: The water and sediment transport prediction field of the simulated grid nodes is interpolated to the regular raster of the remote sensing image. The continuous topographic field is sampled according to the remote sensing observation time to generate a simulated digital elevation model sequence. The remote sensing interpretation results are aggregated into the simulated grid cells. The simulated values ​​of the corresponding time are extracted with the remote sensing observation time as the anchor point to form a paired sample set. The paired sample set is input into an adversarial discriminant network that uses a replica of a Fourier neural operator model as a generator and a binary classification neural network as a discriminant, and outputs a fused simulation-monitoring consistency verification field. The encoder extracts features of the consistency verification field to construct the spatial probability distribution of latent variables. The decoder outputs the mean and variance fields of the performance indicators at each spatial location. The confidence interval of the evaluation conclusion is determined based on the size of the variance field. The simulated time-series data and the monitored time-series data in the paired sample set are input into two feature encoders with the same structure and mapped to a low-dimensional manifold space. The distribution difference between the simulated data manifold and the monitored data manifold in the low-dimensional space is minimized, while the mutual information of samples at the same spatiotemporal location in the manifold space is maximized. The alignment accuracy of the simulated data and the monitored data in the manifold space is output as the physical consistency coefficient. Dynamic mode decomposition is performed on the simulated evolution spatiotemporal field and the monitored evolution spatiotemporal field in the paired sample set, respectively, to extract the spatial distribution and temporal evolution characteristics of each mode. The matching degree between the simulated mode and the monitored mode is calculated by the structural similarity index and the multiple correlation coefficient, and the energy proportion of the reliable mode is output.

[0010] This plan evaluates the effectiveness of the hydrodynamic restoration and micro-topography improvement projects, specifically including: Based on the mean field, hydrodynamic improvement indicators, topographic stability indicators and ecological response indicators of each restoration zone are extracted. Combined with the physical consistency coefficient and the energy proportion of the credible mode, a multi-zone, multi-indicator engineering performance evaluation matrix is ​​constructed. The engineering performance evaluation matrix is ​​input into a self-organizing map neural network, which maps each repair partition to a two-dimensional topological grid, outputs the clustering category and the cluster center vector of each partition, and automatically classifies the performance level according to the Euclidean distance of the cluster center vector. Using the category center vector of each performance level as the reference sequence and the index vector of each repair zone as the comparison sequence, the gray target decision algorithm is used to calculate the gray correlation degree between each zone and the target center of each level. The performance level assignment result of each zone is determined according to the principle of maximum correlation. Based on the performance level attribution results, a spatial distribution map of the engineering performance levels in the repair area is generated.

[0011] In this scheme, an ecological network topology is constructed using ecological patches within the restoration area as nodes and ecological flows as edges, specifically including: Ecological patches are divided based on multi-period ecological element data, and multi-dimensional attribute labels are matched for each ecological patch to form the initial feature vector of the graph neural network node. Identify patch pairs with hydrological connectivity and construct hydrological connectivity edges based on the identification results; Based on vegetation type distribution data and bird habitat survey data, we identified patches with species dispersal potential and constructed species dispersal edges. Based on the water and sediment transport prediction field, patch pairs with nutrient exchange relationships are identified, and nutrient transport edges are constructed; The hydrological connectivity edges, species diffusion edges, and nutrient transport edges together form a multi-edge structure of the ecological network, and the ecological network topology graph is output.

[0012] In this scheme, a graph neural network model is used to learn the spatiotemporal features and predict the state evolution of preset multi-period ecological indicators, generating a dynamic change curve of the comprehensive index of restoration effect, specifically including: A spatiotemporal graph neural network, comprising a spatial graph convolution module, a temporal convolution module, and cross-scale jump connections, is constructed to perform representation learning on the ecological network topology graph. In the spatial graph convolution module, independent graph attention heads are set for hydrological connectivity edges, species diffusion edges, and nutrient transport edges, respectively, and neighborhood information is aggregated to generate spatial feature representations of each node at the current time; The spatial feature representation is input into the temporal convolution module, and the temporal series of node features is modeled using an extended causal convolution architecture to obtain the state prediction value of each node at future time. The fine-grained spatiotemporal features of the shallow network are directly transferred to the deep network through the cross-scale skip connection, and the node state prediction sequence with multi-scale information is fused and output. Construct the health threshold range and the upper limit of the rate of change for each preset ecological indicator, and add the health threshold range and the upper limit of the rate of change as inequality constraints to the network output layer. Based on the node state prediction sequence, a weighted summation method is used to integrate the preset ecological indicators into a comprehensive index of restoration effect, generating a dynamic change curve of the comprehensive index over time.

[0013] In this scheme, the mass conservation residual field, momentum conservation residual field, and sediment continuous residual field output by the intelligent simulation model of water and sediment transport and topographic evolution are projected onto each ecological patch through spatial mapping to generate a physical conservation residual index for each patch. The physical conservation residual index is then added as a regularization term to the loss function of the graph neural network.

[0014] In this plan, when the comprehensive index of the repair effect deviates from the preset repair target trajectory, a tiered warning is triggered, specifically including: Based on the current ecological network topology graph, the degree centrality and betweenness centrality of nodes are calculated as local graph topology indicators, and key hub nodes are selected for key monitoring based on the local graph topology indicators. A graph comparison learning framework is constructed to compare the current ecological network topology graph with the reference graph in the same health state in history, and to calculate the cosine similarity between graph-level feature vectors. The dynamic change curve of the comprehensive index of repair effect is compared with the preset repair target trajectory to calculate the deviation. Combined with the cosine similarity and the status of key hub nodes, the warning is divided into three levels: attention level, warning level and emergency level.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: By integrating multi-scale deep learning networks with tidal phase constraints, high-precision automated extraction of water area, tidal channel network, and vegetation cover was achieved, eliminating temporal comparison biases caused by tidal level differences and providing a unified benchmark for dynamic monitoring of ecological elements. A fusion architecture of Fourier neural operators and physical information neural networks significantly improved the computational efficiency of hydrodynamic simulations. Physical symbol constraints ensured that simulation results conformed to conservation laws, and implicit neural representations enabled continuous terrain modeling at the sub-grid scale, providing accurate physical field data for engineering performance evaluation.

[0016] Multi-model fusion is achieved through adversarial discriminant networks, making the statistical distribution of simulated data approximate that of real monitoring data. The neural process architecture outputs evaluation results with confidence intervals, possessing the ability to quantify uncertainty. Dynamic mode decomposition technology objectively verifies the physical consistency of the models by comparing the dominant evolution modes. A heterogeneous graph network is constructed with ecological functional patches as nodes and multiple ecological flows as edges. The spatiotemporal graph neural network sets independent attention mechanisms for different ecological flow types, and physical constraints are used as regularization terms to guide network learning, significantly improving the predictive generalization ability in sparse data regions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0018] Figure 1 A flowchart of a method for dynamic monitoring of the effects of marine ecological restoration in estuaries based on multi-period remote sensing is shown. Figure 2 A flowchart simulating the water and sediment transport and topographic evolution patterns at different restoration stages is shown; Figure 3 The flowchart shows the dynamic change curve of the comprehensive index of the repair effect. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0021] like Figure 1 As shown, this embodiment provides a method for dynamic monitoring of the marine ecological restoration effect in estuaries based on multi-period remote sensing, including: Multiple satellite and UAV remote sensing images were acquired before and after the restoration project was implemented, and radiometric correction and geometric registration were performed to construct a long-term standardized remote sensing dataset. Interpret the multi-period remote sensing images in the aforementioned remote sensing dataset, extract the spatiotemporal distribution of water areas, tidal channel networks, vegetation cover areas, and artificial structures, and generate multi-period ecological element data based on the interpretation results; The multi-phase ecological element data are input into a pre-constructed intelligent simulation model of water and sediment transport and topographic evolution to simulate the water and sediment transport and topographic evolution patterns at different restoration stages. The effectiveness of hydrodynamic restoration and micro-topographic remediation projects is evaluated by dynamically fitting the model simulation results with actual remote sensing monitoring data. An ecological network topology map is constructed using each ecological patch in the restoration area as a node and ecological flow as an edge. A graph neural network model is used to learn the spatiotemporal features and predict the state evolution of preset multi-period ecological indicators, generating a dynamic change curve of the comprehensive index of restoration effect. When the overall index of the repair effect deviates from the preset repair target trajectory, a graded warning is triggered.

[0022] It should be noted that a deep learning semantic segmentation network integrating multi-scale dilated convolution and deformable convolution is constructed. Dilated convolutions with different dilation rates are set in parallel at different layers of the network, and deformable sampling points are introduced into the convolution kernels. The weighted sum of the weighted focus loss function and the boundary-aware loss function is used as the optimization objective for network training, resulting in a semantic segmentation model for land cover classification in the restoration area. Multiple periods of remote sensing images are successively input into the semantic segmentation network, and the land cover classification probability map corresponding to each period of image is output to obtain the initial land cover classification results including water bodies, vegetation, and man-made structures.

[0023] Instantaneous tide level data at the imaging time of each remote sensing image is acquired, and a tide level-inundation range response relationship is established. Based on this response relationship, the water area in the initial land cover classification results is recalculated, and the instantaneous water area extracted at different tide levels is uniformly recalculated to the theoretical water area corresponding to the reference tide level. Morphological dilation and erosion smoothing are then applied to the recalculated water area boundaries to update the water area in the land cover classification results. Binary images of water bodies are extracted from the updated land cover classification results. Skeleton lines are extracted from these binary images, and a tidal channel network topology is constructed based on the intersections and endpoints of the skeleton lines. A graph convolutional network is used to learn features from the tidal channel topology, based on S... The Trahler classification rule automatically identifies the main channels and tributaries of tidal channels, calculates tidal channel density, bifurcation ratio, and connectivity index; extracts vegetation cover areas from the updated land cover classification results, constructs a phenological feature template library for different vegetation types in a preset period, obtains the temporal spectral curve of each pixel in multi-period remote sensing images, performs similarity measurement between the temporal spectral curve and the phenological feature template library, outputs vegetation type identification results, and inversely retrieves the coverage of each vegetation type; based on the output multi-period interpretation results, constructs a spatiotemporal evolution trend field of land cover categories, inputs the filtered values ​​into a Markov random field model for iterative optimization, and outputs the interpretation results after spatiotemporal consistency correction.

[0024] The spatiotemporal consistency correction results are used to extract the tidal channel network skeleton lines, perform node identification and topology reconstruction on the skeleton lines, and automatically classify the tidal channel network according to the Strahler classification rule. The length, width, and bifurcation ratio of each level of tidal channel are statistically analyzed to generate a tidal channel network topology parameter set containing the tidal channel classification system and geometric parameters. The spatiotemporal consistency correction results are used to extract the boundary lines between water and non-water areas as instantaneous shorelines. The bilinear interpolation method is used to perform sub-pixel-level positioning of the shorelines. Combined with the tidal level data at the imaging time of each image, a shoreline location time-series dataset containing shoreline location coordinates and corresponding tidal level information is generated. The spatiotemporal consistency correction results are used to extract vegetation type distribution and coverage inversion results. The spatial distribution range and area proportion of each vegetation type in the restoration area are statistically analyzed. The coverage time-series curves of each vegetation type are constructed, and the coverage during the spring greening period and the peak coverage during the summer lush period are extracted. Vegetation cover parameter set is generated by analyzing vegetation cover parameters, including those for the autumn withering period. Spatial distribution ranges of artificial structures such as aquaculture ponds, weirs, and dikes are extracted from the interpretation results after spatiotemporal consistency correction. Overlay analysis of distribution maps of artificial structures from adjacent periods is performed to identify newly added, demolished, and unchanged structures. The total area of ​​artificial structures and their proportion of the total restoration area in each period are statistically analyzed to generate a dynamic change dataset of artificial structures. The spatial and temporal references of the tidal channel network topology parameter set, shoreline location time series dataset, vegetation cover parameter set, and dynamic change dataset of artificial structures are unified. All vector data are rasterized to a uniform spatial resolution and spatially resampled. Based on key nodes in the implementation of ecological restoration projects as the time axis, remote sensing observation data with different sampling frequencies are aggregated to a unified time node to construct a standardized dataset of multi-period ecological elements with equal time intervals.

[0025] It should be noted that a pre-constructed intelligent simulation model of water and sediment transport and topographic evolution is used, integrating neural operators and neural symbol systems, to simulate the patterns of water and sediment transport and topographic evolution at different restoration stages. For example... Figure 2 As shown, based on the standardized dataset of the multi-period ecological elements, a sample set for model training and validation is constructed. The sample set feature inputs include topographic elevation field, tidal channel network topology parameters, vegetation cover type and coverage distribution, distribution of artificial structures, upstream runoff and sediment transport, and tidal harmonic constant of the outer sea area. The sample set label outputs include time-period velocity field, water level field, suspended sediment concentration field, and topographic erosion and deposition changes. To further improve the operator learning efficiency, a multi-scale Fourier feature encoding technique is introduced. Fourier modes of different frequency ranges are grouped: low-frequency modes correspond to large-scale topographic evolution and tidal propagation, and are learned using deeper network layers; high-frequency modes correspond to local turbulence and detailed hydrodynamic features of tidal channels, and are learned using shallower network layers.

[0026] A Fourier neural operator architecture is constructed, which performs feature mapping through boosting operators, performs Fast Fourier Transform on the feature field in the Fourier layer, and learns a global convolution kernel in the frequency domain. After inverse Fourier Transform, the initial water and sediment transport prediction field is output through a projection operator. A differentiable physical symbolic inference module is constructed to receive the initial prediction results output by the Fourier neural operator, and performs physical conservation law verification on the velocity field, water level field, and suspended sediment concentration field in the initial prediction results. Mass conservation residual fields, momentum conservation residual fields, and sediment continuity residual fields are constructed as symbolic constraint terms and added to the loss function. The weights of each constraint term are dynamically adjusted according to the input features, and the Fourier neural operator is fed back for iterative optimization to output the corrected water and sediment transport prediction field. The loss function consists of three parts: data fitting loss, physical symbolic constraint loss, and adaptive balance loss. The data fitting loss calculates the mean square error between the initial prediction results output by the Fourier neural operator and the label values ​​in the training sample set, and is expressed as the sum of the predicted velocity field and the mean square error between the initial prediction results and the label values ​​in the training sample set. The weighted sum of the squared deviations of the measured velocity field, the predicted water level field and the measured water level field, the predicted suspended sediment concentration field and the measured suspended sediment concentration field, and the predicted topographic scour and deposition change and the measured scour and deposition amount; the mass conservation loss term in the physical symbol constraint loss is constructed based on the water balance of each patch node, calculating the squared residual between the water entering the patch minus the water flowing out of the patch and the rate of change of water storage within the patch; the momentum conservation loss term is constructed based on the momentum balance of each hydrological connectivity edge, calculating the squared residual between the resultant force acting on the water body and the rate of change of momentum; the sediment continuity loss term is constructed based on the sediment mass balance of each patch node, calculating the squared residual between the rate of change of suspended sediment concentration and the sum of sediment transport divergence and bed scour and deposition rate; the adaptive balance loss calculates the dominant weights of each physical process in the simulation based on the current input features, and dynamically adjusts the weight coefficients of each sub-item in the data fitting loss term and the physical symbol constraint loss term through a learnable symbol weight adjustment module.

[0027] A neural ordinary differential equation (NDE) is constructed to model the modified water and sediment transport prediction field in continuous time. The core idea of ​​the NDE is to define a derivative function parameterized by a neural network. This function takes the current state field (topography, flow velocity, water level, etc.) as input and outputs the derivative of the state with respect to time. Starting from the initial state, the derivative function is integrated through an NDE solver to obtain the state prediction for any future time. Using the neural network parameterized derivative function and the current water and sediment transport prediction field as the initial state, the derivative function is integrated through an NDE solver to output the water and sediment transport state prediction sequence for any future time. A multilayer perceptron network with spatiotemporal coordinates as input is constructed to fit multi-period topographic elevation data, outputting a continuous topographic field as the bed boundary condition for the Fourier neural operator. The topographic field generated by the implicit neural representation is bidirectionally coupled with the hydrodynamic field predicted by the Fourier neural operator: the topographic field provides the bed boundary for the hydrodynamic simulation, and the scouring and deposition volume predicted by the hydrodynamic operator updates the topographic field, forming a closed-loop co-evolution.

[0028] Furthermore, based on the water and sediment transport state prediction sequence output by the continuous time evolution modeling sub-step and the continuous topographic field output by the implicit neural representation topographic modeling sub-step, a multi-scale spatiotemporal causal graph is constructed. The spatiotemporal convergence cross-mapping algorithm is used to calculate the strength and direction of causal relationships between different variables, and the transfer entropy algorithm is used to calculate the contribution of each engineering measure to the system state evolution. The causal attribution graph containing causal relationship direction, strength and contribution components is output.

[0029] It should be noted that the predicted water and sediment transport field and continuous topographic field generated by the Fourier neural operator simulation are dynamically fitted with multi-period ecological element data obtained by remote sensing interpretation to evaluate the implementation effectiveness of hydrodynamic restoration and micro-topography improvement projects.

[0030] The velocity and water level fields on the unstructured grid output by the Fourier neural operator are mapped to the regular grid coordinate system of the remote sensing image using the radial basis function interpolation method. The interpolation radius is dynamically adjusted according to the tidal channel scale. A smaller radius is used in the dense tidal channel area to preserve local hydrodynamic details. The continuous topographic field is sampled according to the remote sensing observation time to generate a sequence of simulated digital elevation models. The water area, vegetation cover data and multi-temporal digital elevation models interpreted by remote sensing are aggregated into simulated grid cells. Simulated values ​​at the corresponding time are extracted with the remote sensing observation time as the anchor point to form a paired sample set. To eliminate the impact of single-model structural bias on the evaluation results, an adversarial discriminative fusion network is constructed. The paired sample set is input into the adversarial discriminative network, which uses multiple replicas of Fourier neural operators as generators and a binary classification neural network as a discriminator. Through minimax game training, the output weights of each model replica are adaptively adjusted. While the discriminator parameters are fixed, the output weights of each model replica are adjusted to make it difficult for the discriminator to distinguish between the fused prediction field and the combination of remote sensing monitoring data. The generator weights are fixed, and the discriminator is trained to improve its discriminative ability. After multiple iterations, the weights of each model replica are adaptively adjusted until the fused prediction field statistically most closely approximates the remote sensing monitoring data. After training, the fused simulation-monitoring consistency verification field is output.

[0031] The encoder extracts features from the consistency verification field to construct the spatial probability distribution of latent variables. The decoder performs upsampling reconstruction and outputs the mean and variance fields of the performance indicators at each spatial location. The confidence interval of the evaluation conclusion is determined based on the size of the variance field. The mean field reflects the point estimate of the engineering performance indicators, and the variance field reflects the degree of uncertainty of the estimate.

[0032] The simulated time-series data and monitored time-series data from the paired sample set are input into two structurally identical feature encoders, mapped to a low-dimensional manifold space. A manifold alignment loss function is constructed to minimize the distribution difference between the simulated and monitored data manifolds in the low-dimensional space, while maximizing the mutual information of samples at the same spatiotemporal location in the manifold space. A contrastive loss function is used to bring positive sample pairs (at the same location) closer together and negative sample pairs (at different locations) further apart in the manifold space. After training, the reciprocal of the average Euclidean distance between the simulated and monitored data manifolds is used as the manifold alignment accuracy and the physical consistency coefficient.

[0033] Dynamic mode decomposition is performed on the simulated and monitored spatiotemporal fields of the paired sample set. High-dimensional spatiotemporal data matrices are constructed using the simulated spatiotemporal field (including long-term flow velocity and topographic data at various locations in the tidal channel) and the monitored spatiotemporal field of the paired sample set as inputs. After dimensionality reduction through singular value decomposition, the best-fit linear operator is solved to extract dynamic modes of each order. The matching degree between the simulated and monitored modes is calculated using the structural similarity index and the multiple correlation coefficient. Modes with spatial similarity and eigenvalue matching degrees exceeding preset thresholds are marked as reliable modes. The proportion of reliable mode energy to the sum of squared amplitudes of all modes is calculated, and the reliable mode energy proportion is output.

[0034] It should be noted that the implementation effectiveness of hydrodynamic restoration and micro-topography improvement projects is comprehensively rated based on the mean field, physical consistency coefficient, and credible modal energy ratio. Based on the mean field, hydrodynamic improvement indicators, topographic stability indicators, and ecological response indicators are extracted for each restoration zone. Combined with the physical consistency coefficient and credible modal energy ratio, a multi-zone, multi-indicator engineering effectiveness evaluation matrix is ​​constructed. This matrix is ​​then input into a self-organizing map neural network to map each restoration zone onto a two-dimensional topological grid, outputting the clustering category and category center vector for each zone. Based on the Euclidean distance of the category center vectors, four effectiveness levels—excellent, good, qualified, and needing improvement—are automatically assigned. Using the category center vectors of each effectiveness level as a reference sequence and the indicator vectors of each restoration zone as a comparison sequence, a gray target decision algorithm is used to calculate the gray correlation degree between each zone and the target center of each level. The effectiveness level assignment result for each zone is determined based on the principle of maximum correlation. Based on the effectiveness level assignment results, a spatial distribution map of the engineering effectiveness levels in the restoration area is generated.

[0035] It should be noted that, based on vegetation cover maps, water distribution maps, and tidal channel network maps from multiple ecological element data periods, a watershed segmentation algorithm and regional merging were employed to divide the restoration area into several ecological patches with internal homogeneity and significant boundaries, avoiding the artificial boundary effect caused by grid division. The segmentation process comprehensively considered spectral characteristics, spatial adjacency relationships, and ecological type labels, ensuring relatively consistent ecological attributes within each patch and significant differences in ecological attributes between patches. Furthermore, multi-dimensional attribute labels were matched to each ecological patch, including: tidal channel density, tidal channel bifurcation ratio, tidal channel connectivity, vegetation cover, vegetation type, distance from the shoreline, and influence factors of artificial structures in the area where the patch is located, forming the initial feature vector of the graph neural network nodes.

[0036] Based on the tidal channel network topology and simulated hydrodynamic field, patch pairs with hydrological connectivity are identified, and hydrological connectivity edges are constructed according to the identification results. If two patches are directly connected through the tidal channel system, or if there is a periodically flooded hydrological path between them, a hydrological connectivity edge is constructed between them. The weight of the edge is determined by a combination of the hydraulic gradient, flow velocity, and water transport capacity between the two points. Based on vegetation type distribution data and bird habitat survey data, patch pairs with species dispersal potential are identified, and species dispersal edges are constructed. For vegetation communities, considering seed dispersal distance and direction, species dispersal edges are constructed between adjacent or close patches of the same type. For bird habitats, considering the maximum flight distance and habitat preferences of major bird species, potential dispersal edges are constructed between suitable habitat patches, and the weight of the edge is determined by a combination of dispersal distance, vegetation type similarity, and human disturbance intensity. Based on the water and sediment transport prediction field, patch pairs with nutrient exchange relationships are identified, and nutrient transport edges are constructed. Nutrient transport edges are also constructed between upstream and downstream patches, as well as between adjacent patches with significant concentration gradients. The weight of each edge is determined by a combination of the mass flux and the magnitude of the concentration gradient between the two points. These hydrological connectivity edges, species dispersal edges, and nutrient transport edges together form a multi-edge structure of the ecological network, and the ecological network topology is output.

[0037] It should be noted that, as Figure 3As shown, a spatiotemporal graph neural network is constructed, comprising a spatial graph convolution module, a temporal convolution module, and cross-scale skip connections, to perform representation learning on the ecological network topology. The input of the spatiotemporal graph neural network is the feature vector sequence of each patch node at multiple historical moments, and the output is the predicted state value of each node at future moments, including key ecological indicators such as tidal channel density, tidal channel connectivity, vegetation cover, vegetation type diversity, shoreline stability index, and artificial structure impact index. In the spatial graph convolution module, independent graph attention heads are set for hydrological connectivity edges, species diffusion edges, and nutrient transport edges, respectively, and the attention weights of different types of neighboring nodes to the central node are calculated, enabling the network to distinguish the differentiated impacts of different types of ecological flows on the evolution of node states, and to aggregate neighborhood information to generate spatial feature representations of each node at the current moment. The spatial feature representations are input into the temporal convolution module, and a dilated causal convolution architecture is used to model the temporal series of node features. The dilated convolution skips sampling points of a fixed step size in the time dimension, allowing the convolution kernel to cover historical information over a larger time span while keeping the number of parameters constant. By stacking multiple layers of dilated causal convolutions to capture the multi-scale temporal dependencies from short-term fluctuations to long-term trends, the state prediction values ​​of each node at future moments are obtained. Through the cross-scale skip connections, the fine-grained spatiotemporal features of the shallow network are directly transferred to the deep network and fused with the coarse-grained abstract features of the deep network. The shallow network retains local detail information at the patch scale, while the deep network extracts global pattern information at the regional scale. This enables the network to accurately predict the state changes of local patches while ensuring the coordination and consistency of the overall ecological network, outputting a node state prediction sequence that integrates multi-scale information.

[0038] A physical symbol constraint module for training Fourier neural operators and a physical consistency coefficient for dynamic fitting are introduced as prior physical knowledge to guide the learning process of the graph neural network. The mass conservation residual field, momentum conservation residual field, and sediment continuity residual field output from the intelligent simulation model of water and sediment transport and topographic evolution are projected onto each ecological patch through spatial mapping to generate a physical conservation residual index for each patch. This physical conservation residual index is then added as a regularization term to the loss function of the graph neural network, ensuring that the network's state evolution process satisfies fundamental physical laws. The mass conservation loss is constructed based on the water balance of each patch node; the amount of water entering the patch minus the amount of water flowing out of the patch should equal the rate of change of water storage within the patch. The momentum conservation loss is constructed based on the momentum balance of each hydrologically connected edge; the resultant force acting on the water body should equal the rate of change of momentum. The sediment continuity loss is constructed based on the sediment mass balance of each patch node; the sum of the rate of change of suspended sediment concentration, sediment transport rate divergence, and bed sedimentation rate should be zero. The dynamically fitted physical consistency coefficient is used as the sample weight for each patch prediction task, guiding the network to learn the state evolution process that conforms to physical laws. For regions with high physical consistency coefficients, higher weights are assigned to make the network pay more attention to the fitting accuracy of these regions; for regions with low physical consistency coefficients, lower weights are assigned to avoid being misled by unreliable data.

[0039] Based on ecological literature and project baseline survey data, health threshold ranges and upper limits of change rates for each preset ecological indicator are constructed. These health threshold ranges and upper limits of change rates are added as inequality constraints to the network output layer. The comprehensive weight of each ecological indicator is dynamically determined using the entropy weight method or the analytic hierarchy process (AHP). Based on the node state prediction sequence, the preset ecological indicators are integrated into a comprehensive restoration effect index using a weighted summation method, generating a dynamic curve of the comprehensive index over time. Furthermore, using the area of ​​each patch as a weight, the restoration effect index of the patch nodes in the entire restoration area is averaged by area to obtain the overall comprehensive restoration effect index of the restoration area. .

[0040] It should be noted that, based on the current ecological network topology, node degree centrality and betweenness centrality are calculated as local graph topology indicators. Key pivot nodes are then selected for focused monitoring based on these local graph topology indicators, such as patches ranking in the top 10% in degree centrality or betweenness centrality. These patches play crucial roles in the ecological network, including material transport, energy flow, and species dispersal; degradation in these patches will trigger a chain reaction. A graph comparison learning framework is constructed, comparing the current ecological network topology with a reference graph from the same historical period in a healthy state. Graph-level feature vectors are extracted from the two graphs using a graph isomorphic network, and the cosine similarity between these feature vectors is calculated. The dynamic change curve of the comprehensive restoration effect index is compared with the preset restoration target trajectory, and the deviation is calculated. Combining the cosine similarity and the status of key pivot nodes, the warning is divided into three levels: attention level, warning level, and emergency level.

[0041] When an early warning is triggered, the attention weight distribution embedded in the graph neural network is extracted, and the key nodes and edges that led to the warning are traced back. The ecological patches and ecological flow types that contribute most to the current abnormal state are identified. Based on the warning level and the source analysis results, a pre-built engineering measures knowledge base is input. This knowledge base includes information such as the applicable conditions, expected effects, and implementation costs of engineering measures such as micro-topography restoration, hydrodynamic restoration, and vegetation replanting. A case-based reasoning method is used to generate recommendations, matching the current abnormal state with historical successful cases to suggest the optimal combination of control schemes. Adaptive control suggestions, including the type of control measure, implementation location, expected effects, and implementation costs, are output. The early warning information and control suggestions are then output to the management platform in the form of a visual graph.

[0042] The second embodiment of the present invention provides a computer-readable storage medium, which includes a program for a dynamic monitoring method of the effect of marine ecological restoration in estuary areas based on multi-period remote sensing. When the program for the dynamic monitoring method of the effect of marine ecological restoration in estuary areas based on multi-period remote sensing is executed by a processor, it implements the steps of the method for the dynamic monitoring method of marine ecological restoration in estuary areas based on multi-period remote sensing.

[0043] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring of the effects of marine ecological restoration in estuaries and sea areas based on multi-period remote sensing, characterized in that, Includes the following steps: Multiple satellite and UAV remote sensing images were acquired before and after the restoration project was implemented, and radiometric correction and geometric registration were performed to construct a long-term standardized remote sensing dataset. Interpret the multi-period remote sensing images in the aforementioned remote sensing dataset, extract the spatiotemporal distribution of water areas, tidal channel networks, vegetation cover areas, and artificial structures, and generate multi-period ecological element data based on the interpretation results; The multi-phase ecological element data are input into a pre-constructed intelligent simulation model of water and sediment transport and topographic evolution to simulate the water and sediment transport and topographic evolution patterns at different restoration stages. The effectiveness of hydrodynamic restoration and micro-topographic remediation projects is evaluated by dynamically fitting the model simulation results with actual remote sensing monitoring data. An ecological network topology map is constructed using each ecological patch in the restoration area as a node and ecological flow as an edge. A graph neural network model is used to learn the spatiotemporal features and predict the state evolution of preset multi-period ecological indicators, generating a dynamic change curve of the comprehensive index of restoration effect. When the overall repair effect index deviates from the preset repair target trajectory, a tiered warning is triggered; A pre-constructed intelligent simulation model of water and sediment transport and topographic evolution is used, integrating neural operators and neural symbolic systems, to simulate the patterns of water and sediment transport and topographic evolution at different restoration stages. Specifically, this includes: A Fourier neural operator architecture is constructed, feature mapping is performed by lifting operators, fast Fourier transform is performed on the feature field in the Fourier layer and global convolution kernel is learned in the frequency domain space, and the initial water and sediment transport prediction field is output through projection operators after inverse Fourier transform. For the initial water and sediment transport prediction field, a mass conservation residual field, a momentum conservation residual field, and a sediment continuity residual field are constructed as symbolic constraint terms and added to the loss function. The weights of each constraint term are dynamically adjusted according to the input characteristics, and the Fourier neural operator is fed back for iterative optimization to output the corrected water and sediment transport prediction field. A neural network is used to construct a constant-time differential equation to model the modified water and sediment transport prediction field, and the derivative function is parameterized by a neural network to output the prediction sequence of water and sediment transport state at future times. A multilayer perceptron network with spatiotemporal coordinates as input is constructed to fit multi-period topographic elevation data, outputting a continuous topographic field as the bed boundary condition of the Fourier neural operator, and the topographic erosion and deposition volume output by the Fourier neural operator is fed back to update the multilayer perceptron network.

2. The method for dynamic monitoring of marine ecological restoration effects in estuaries and sea areas based on multi-period remote sensing as described in claim 1, characterized in that, Interpreting multiple remote sensing images from the aforementioned remote sensing dataset to extract the spatiotemporal distribution of water areas, tidal channel networks, vegetation cover areas, and man-made structures, specifically including: A deep learning semantic segmentation network integrating multi-scale dilated convolution and deformable convolution is constructed. Multiple remote sensing images are input into the semantic segmentation network one after another, and the land cover classification probability map corresponding to each image is output to obtain the initial land cover classification results including water, vegetation and artificial structures. Instantaneous tide data at the time of imaging of each remote sensing image is obtained, a tide level-inundation range response relationship is established, the water area in the initial land cover classification result is calculated based on the response relationship, and the water area in the land cover classification result is updated through morphological dilation and erosion smoothing. The binary image of the water body region is extracted from the updated land cover classification results. The skeleton line is extracted from the binary image of the water body region. The tidal channel network topology map is constructed based on the intersection and endpoint of the skeleton line. The tidal channel topology map is used to learn features by graph convolutional network, the main channel and tributaries of each level of the tidal channel are identified, and the tidal channel density, bifurcation ratio and connectivity index are calculated. The vegetation cover area is extracted from the updated land cover classification results, a phenological feature template library of different vegetation types in a preset period is constructed, the temporal spectral curve of each pixel in multiple remote sensing images is obtained, the similarity between the temporal spectral curve and the phenological feature template library is measured, the vegetation type identification result is output, and the coverage of each vegetation type is inverted. Based on the multi-period interpretation results, a spatiotemporal evolution trend field of land cover categories is constructed. After filtering, the value is input into a Markov random field model for iterative optimization, and the interpretation results with spatiotemporal consistency correction are output.

3. The method for dynamic monitoring of marine ecological restoration effects in estuaries and sea areas based on multi-period remote sensing as described in claim 2, characterized in that, Based on the interpretation results, multi-period ecological element data are generated, specifically including: Extract tidal channel network topology parameters, shoreline location temporal information, vegetation cover parameters, and dynamic change data of artificial structures from the interpretation results after spatiotemporal consistency correction. The spatial and temporal references of the tidal channel network topology parameter set, shoreline location time series dataset, vegetation cover parameter set, and dynamic change dataset of artificial structures are unified. All vector data are rasterized to a uniform spatial resolution and spatially resampled. Using the key nodes of ecological restoration projects as the basis for timeline division, remote sensing observation data with different sampling frequencies are aggregated to a unified time node to construct a standardized dataset of ecological elements with equal time intervals across multiple periods.

4. The method for dynamic monitoring of marine ecological restoration effects in estuaries and sea areas based on multi-period remote sensing according to claim 1, characterized in that, Dynamic fitting of model simulation results with actual remote sensing monitoring data specifically includes: The water and sediment transport prediction field of the simulated grid nodes is interpolated to the regular raster of the remote sensing image. The continuous topographic field is sampled according to the remote sensing observation time to generate a simulated digital elevation model sequence. The remote sensing interpretation results are aggregated into the simulated grid cells. The simulated values ​​of the corresponding time are extracted with the remote sensing observation time as the anchor point to form a paired sample set. The paired sample set is input into an adversarial discriminant network that uses a replica of a Fourier neural operator model as a generator and a binary classification neural network as a discriminant, and outputs a fused simulation-monitoring consistency verification field. The encoder extracts features of the consistency verification field to construct the spatial probability distribution of latent variables. The decoder outputs the mean and variance fields of the performance indicators at each spatial location. The confidence interval of the evaluation conclusion is determined based on the size of the variance field. The simulated time-series data and the monitored time-series data in the paired sample set are input into two feature encoders with the same structure and mapped to a low-dimensional manifold space. The distribution difference between the simulated data manifold and the monitored data manifold in the low-dimensional space is minimized, while the mutual information of samples at the same spatiotemporal location in the manifold space is maximized. The alignment accuracy of the simulated data and the monitored data in the manifold space is output as the physical consistency coefficient. Dynamic mode decomposition is performed on the simulated evolution spatiotemporal field and the monitored evolution spatiotemporal field in the paired sample set, respectively, to extract the spatial distribution and temporal evolution characteristics of each mode. The matching degree between the simulated mode and the monitored mode is calculated by the structural similarity index and the multiple correlation coefficient, and the energy proportion of the reliable mode is output.

5. The method for dynamic monitoring of marine ecological restoration effects in estuaries and sea areas based on multi-period remote sensing according to claim 4, characterized in that, The effectiveness of hydrodynamic restoration and micro-topography improvement projects will be evaluated, including: Based on the mean field, hydrodynamic improvement indicators, topographic stability indicators and ecological response indicators of each restoration zone are extracted. Combined with the physical consistency coefficient and the energy proportion of the credible mode, a multi-zone, multi-indicator engineering performance evaluation matrix is ​​constructed. The engineering performance evaluation matrix is ​​input into a self-organizing map neural network, which maps each repair partition to a two-dimensional topological grid, outputs the clustering category and the cluster center vector of each partition, and automatically classifies the performance level according to the Euclidean distance of the cluster center vector. Using the category center vector of each performance level as the reference sequence and the index vector of each repair zone as the comparison sequence, the gray target decision algorithm is used to calculate the gray correlation degree between each zone and the target center of each level. The performance level assignment result of each zone is determined according to the principle of maximum correlation. Based on the performance level attribution results, a spatial distribution map of the engineering performance levels in the repair area is generated.

6. The method for dynamic monitoring of marine ecological restoration effects in estuaries and sea areas based on multi-period remote sensing according to claim 1, characterized in that, An ecological network topology is constructed using ecological patches within the restoration area as nodes and ecological flows as edges, specifically including: Ecological patches are divided based on multi-period ecological element data, and multi-dimensional attribute labels are matched for each ecological patch to form the initial feature vector of the graph neural network node. Identify patch pairs with hydrological connectivity and construct hydrological connectivity edges based on the identification results; Based on vegetation type distribution data and bird habitat survey data, we identified patches with species dispersal potential and constructed species dispersal edges. Based on the water and sediment transport prediction field, patch pairs with nutrient exchange relationships are identified, and nutrient transport edges are constructed; The hydrological connectivity edges, species diffusion edges, and nutrient transport edges together form a multi-edge structure of the ecological network, and the ecological network topology graph is output.

7. The method for dynamic monitoring of marine ecological restoration effects in estuaries and sea areas based on multi-period remote sensing according to claim 1, characterized in that, A graph neural network model is used to learn the spatiotemporal features and predict the state evolution of preset multi-period ecological indicators, generating a dynamic change curve of the comprehensive index of restoration effect, specifically including: A spatiotemporal graph neural network, comprising a spatial graph convolution module, a temporal convolution module, and cross-scale jump connections, is constructed to perform representation learning on the ecological network topology graph. In the spatial graph convolution module, independent graph attention heads are set for hydrological connectivity edges, species diffusion edges, and nutrient transport edges, respectively, and neighborhood information is aggregated to generate spatial feature representations of each node at the current time; The spatial feature representation is input into the temporal convolution module, and the temporal series of node features is modeled using an extended causal convolution architecture to obtain the state prediction value of each node at future time. The fine-grained spatiotemporal features of the shallow network are directly transferred to the deep network through the cross-scale skip connection, and the node state prediction sequence with multi-scale information is fused and output. Construct the health threshold range and the upper limit of the rate of change for each preset ecological indicator, and add the health threshold range and the upper limit of the rate of change as inequality constraints to the network output layer. Based on the node state prediction sequence, a weighted summation method is used to integrate the preset ecological indicators into a comprehensive index of restoration effect, generating a dynamic change curve of the comprehensive index over time.

8. The method for dynamic monitoring of marine ecological restoration effects in estuaries based on multi-period remote sensing according to claim 7, characterized in that, The mass conservation residual field, momentum conservation residual field, and sediment continuity residual field output by the intelligent simulation model of water and sediment transport and topographic evolution are projected onto each ecological patch through spatial mapping to generate a physical conservation residual index for each patch. The physical conservation residual index is then added as a regularization term to the loss function of the graph neural network.

9. The method for dynamic monitoring of marine ecological restoration effects in estuaries and sea areas based on multi-period remote sensing according to claim 1, characterized in that, When the overall repair effect index deviates from the preset repair target trajectory, a tiered warning is triggered, specifically including: Based on the current ecological network topology graph, the degree centrality and betweenness centrality of nodes are calculated as local graph topology indicators, and key hub nodes are selected for key monitoring based on the local graph topology indicators. A graph comparison learning framework is constructed to compare the current ecological network topology graph with the reference graph in the same health state in history, and to calculate the cosine similarity between graph-level feature vectors. The dynamic change curve of the comprehensive index of repair effect is compared with the preset repair target trajectory to calculate the deviation. Combined with the cosine similarity and the status of key hub nodes, the warning is divided into three levels: attention level, warning level and emergency level.

Images