A Deep Learning-Based Dynamic Monitoring Method for Wetland Ecological Restoration

Abstract

This invention discloses a deep learning-based dynamic monitoring method for wetland ecological restoration, relating to the field of ecological restoration technology. The method includes the following steps: real-time acquisition of multi-source wetland ecological sensor data and remote sensing image streams; construction of a spatiotemporal fusion data cube; extraction of ecological feature tensors; degradation pattern analysis of the ecological feature tensors to generate a dynamic topological map of the ecological state; and calculation of the ecological connectivity index; identification of restoration needs based on the ecological connectivity index; location of degradation hotspots using a multimodal graph convolutional network; and generation of a coordinate set of priority restoration areas. Through spatiotemporal fusion of multi-source data and deep integration of deep learning and landscape ecology, a complete technical system is constructed, encompassing dynamic perception of ecological state and intelligent optimization of restoration plans. This effectively addresses pain points in traditional wetland restoration, such as data scale mismatch, ambiguous location of degradation areas, poor ecological adaptability of restoration paths, and difficulty in multi-objective coordination.

Description

Technical Field

[0001] This invention relates to the field of ecological restoration technology, and more specifically, to a dynamic monitoring method for wetland ecological restoration based on deep learning. Background Technology

[0002] Against the backdrop of escalating wetland degradation globally, ecological restoration has become a core means of wetland protection. However, with the expansion of restoration scale and the increasing complexity of ecosystems, traditional monitoring methods face severe challenges in terms of data fusion accuracy, real-time restoration response, and multi-objective synergistic optimization.

[0003] Existing technologies struggle to effectively integrate heterogeneous data sources such as satellite remote sensing, ground sensors, and drone aerial photography, resulting in a lack of spatiotemporal alignment for key ecological parameters like vegetation cover, hydrological fluctuations, and soil moisture, leading to biased assessments of restoration effectiveness. Traditional static monitoring models fail to capture the nonlinear evolution of wetland ecosystems, especially under dynamic disturbances such as tidal forces and rainy season floods, resulting in high error rates in boundary identification between restored and degraded areas. Current methods lack synergistic optimization mechanisms for biodiversity enhancement, restoration cost control, and engineering sustainability, leading to a negative effect of "local improvement, overall degradation" in restoration projects.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a dynamic monitoring method for wetland ecological restoration based on deep learning. By integrating multi-source data in time and space and deeply integrating deep learning with landscape ecology, a complete technical system is constructed, from dynamic perception of ecological status to intelligent optimization of restoration plans. This effectively solves the pain points in traditional wetland restoration, such as data scale mismatch, ambiguous positioning of degraded areas, poor ecological adaptability of restoration paths, and difficulty in multi-objective coordination.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A deep learning-based dynamic monitoring method for wetland ecological restoration includes the following steps: real-time acquisition of multi-source wetland ecological sensor data and remote sensing image streams, construction of a spatiotemporal fusion data cube, and extraction of ecological feature tensors; degradation mode analysis of the ecological feature tensors to generate a dynamic topological map of the ecological state, and calculation of the ecological connectivity index; identification of restoration needs based on the ecological connectivity index, location of degradation hotspots through a multimodal graph convolutional network, and generation of a set of coordinates for restoration priority areas; ecological corridor modeling of the set of coordinates for restoration priority areas, and generation of initial restoration path boundaries using a cellular automata evolutionary algorithm; construction of an ecological benefit optimization function, collaborative optimization of the initial restoration path boundaries using an adaptive multi-objective evolutionary algorithm, outputting the optimal restoration scheme and dynamically adjusting restoration equipment.

[0008] In a preferred embodiment, the degradation pattern analysis of the ecological feature tensor to generate a dynamic topological map of the ecological state and the calculation of the ecological connectivity index specifically involves: analyzing the three-dimensional data of hydrological fluctuations, vegetation cover, and soil moisture in the ecological feature tensor to construct a dynamic attribute map; calculating the material exchange matrix between ecological units in the attribute map and extracting patch structure features by introducing landscape ecology theory; fusing the material exchange matrix and patch structure features to construct a weighted dynamic topological map of the ecological state; and calculating the spatial permeability change gradient of the topological map during the continuous restoration stage as the ecological connectivity index.

[0009] In a preferred embodiment, the identification of restoration needs based on the ecological connectivity index, and the location of degradation hotspots through a multimodal graph convolutional network to generate a set of coordinates for restoration priority areas, specifically involves: mapping the ecological connectivity index to a Riemannian manifold space to construct an ecological degradation distance matrix; using an attention-weighted graph convolutional network to aggregate spatial features of the ecological degradation distance matrix; using a variational autoencoder to perform latent space clustering on the feature vectors to identify degradation hotspots exceeding the degradation threshold; and extracting the coordinates of hydrological nodes, vegetation patch vectors, and soil pollution ranges within the degradation hotspots to generate a set of coordinates for restoration priority areas.

[0010] In a preferred embodiment, the step of modeling ecological corridors for the coordinate set of priority restoration areas and generating initial restoration path boundaries using a cellular automata evolutionary algorithm specifically involves: analyzing the ecological correlation of the coordinate set of priority restoration areas and constructing a biological migration corridor model; using an ant colony optimization algorithm to simulate species diffusion paths and generate candidate ecological corridors; calculating the landscape pattern similarity between candidate ecological corridors and historical restoration achievements, and selecting paths that meet the similarity criteria as initial restoration path boundaries.

[0011] In a preferred embodiment, the ecological correlation of the coordinate set of the priority restoration area and the construction of a biological migration corridor model are specifically as follows: each coordinate point of the priority restoration area is taken as a cell node, and the node attributes include the ecological resistance coefficient; the ecological transition probability between nodes is calculated based on the species migration pattern and hydrological connectivity; the corridor construction cost between nodes is quantified using the minimum cumulative resistance model; and the ecological transition probability and construction cost are integrated to generate a multi-scale ecological corridor network.

[0012] In a preferred embodiment, the construction of the ecological benefit optimization function, employing an adaptive multi-objective evolutionary algorithm to collaboratively optimize the initial restoration path boundary, outputting the optimal restoration scheme and dynamically controlling the restoration equipment, specifically involves: constructing an ecological benefit optimization function that includes biodiversity gain, restoration cost, and engineering sustainability; encoding the initial restoration path boundary as a gene population and decomposing and optimizing it using an improved MOEA / D algorithm; introducing a simulated annealing mechanism for dynamic neighborhood search to generate a non-dominated solution set; and selecting the optimal restoration scheme from the non-dominated solution set based on the fuzzy integral method to drive the restoration equipment to execute the scheme.

[0013] In a preferred embodiment, the initial repair path boundary is encoded as a gene population and decomposed and optimized using an improved MOEA / D algorithm. Specifically, the weight vector is initialized with a Sobol sequence to ensure uniform distribution of the solution set; a two-layer LSTM network is designed to predict the ecological response trend and adaptively adjust the neighborhood size; and a landscape shape index is introduced as a diversity maintenance operator to prevent premature convergence of the solution set.

[0014] The technical effects and advantages of the deep learning-based dynamic monitoring method for wetland ecological restoration proposed in this invention are as follows:

[0015] This invention constructs a comprehensive technical system encompassing the entire process from dynamic perception of ecological status to intelligent optimization of restoration plans through multi-source data spatiotemporal fusion and deep integration of deep learning and landscape ecology. It effectively addresses pain points in traditional wetland restoration, such as data scale mismatch, ambiguous location of degraded areas, poor ecological adaptability of restoration paths, and difficulty in multi-objective coordination. Its dynamic topology graph analysis, manifold spatial clustering, and adaptive multi-objective optimization technologies enable accurate identification of degradation hotspots, dynamic evolution simulation of ecological corridors, and synergistic optimization of ecological, economic, and sustainable restoration plans, forming a closed loop of "monitoring-decision-execution." Compared to traditional methods, this significantly improves the accuracy, foresight, and efficiency of wetland restoration, providing systematic technical support for the scientific protection and efficient restoration of wetland ecosystems. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a deep learning-based dynamic monitoring method for wetland ecological restoration.

[0017] Figure 2 This is a schematic diagram illustrating the spatiotemporal fusion data cube construction process for a deep learning-based dynamic monitoring method for wetland ecological restoration.

[0018] Figure 3 This is a schematic diagram of the multi-objective optimization and dynamic regulation process of a deep learning-based dynamic monitoring method for wetland ecological restoration.

[0019] Figure 4 This is a schematic diagram of a dynamic monitoring system for wetland ecological restoration based on deep learning. Detailed Implementation

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

[0021] Example 1, Figures 1 to 3 This invention presents a deep learning-based dynamic monitoring method for wetland ecological restoration, comprising the following steps:

[0022] Real-time acquisition of multi-source wetland ecological sensor data and remote sensing image streams, construction of spatiotemporal fusion data cube, and extraction of ecological feature tensors;

[0023] It should be noted that "multi-source wetland ecological sensor data" and "remote sensing image stream" together constitute a three-dimensional sensing data source for wetland ecological status. Its core is to achieve comprehensive capture of ecological processes through multi-dimensional coverage of "point-area-time".

[0024] Multi-source wetland ecological sensor data: refers to micro-ecological data collected through a ground-based sensor network deployed in key wetland areas, specifically including:

[0025] Hydrological sensors: Real-time monitoring of water level changes, water flow velocity, water quality parameters (such as dissolved oxygen, pH value, nitrogen and phosphorus concentration), etc., with sampling frequency up to the minute level, accurately capturing the instantaneous characteristics of wetland hydrological fluctuations;

[0026] Soil sensors: measure soil moisture (water content), electrical conductivity (reflecting salinity), organic matter content, etc., and are deployed in soil profiles at different depths to reflect the microscopic processes of soil-vegetation interaction;

[0027] Vegetation sensors: Real-time data collection of spectral reflectance, chlorophyll content, transpiration rate, etc. of vegetation leaves using spectrometers, chlorophyll meters, etc., directly reflecting the vegetation growth status.

[0028] Meteorological sensors: record meteorological elements such as rainfall, wind speed, temperature, and humidity, serving as external driving factors for wetland ecological processes.

[0029] These data are characterized by "high temporal resolution and low spatial coverage", which can accurately depict the dynamic changes of local ecological units, but they have the limitation of "limited spatial representativeness" when used alone.

[0030] Remote sensing image stream: refers to the continuous acquisition of macroscopic wetland image data through remote sensing platforms such as satellites and drones, including:

[0031] Satellite remote sensing: such as Landsat (30m resolution, 16-day revisit) and Sentinel-2 (10m resolution, 5-day revisit) optical images, can extract large-scale vegetation coverage, water area, land use type, etc.; Sentinel-1 radar images can penetrate clouds and monitor wetland water level changes and vegetation biomass in rainy weather.

[0032] Drone remote sensing: It acquires centimeter-level high-resolution images through low-altitude flight, and conducts supplementary monitoring for sensor coverage blind spots or key restoration areas, capturing details such as vegetation distribution and corridor connectivity at the patch scale.

[0033] Remote sensing images are characterized by "high spatial coverage and low temporal resolution". They can reflect the macroscopic pattern of wetland ecosystems, but it is difficult to capture the instantaneous changes of microscopic processes.

[0034] Real-time assurance mechanism: Sensor data is preprocessed locally through edge computing nodes (such as noise reduction and outlier removal) and then uploaded to the cloud via wireless transmission technologies such as 5G / NB-IoT; Remote sensing images are received and preprocessed in near real-time (such as atmospheric correction and geometric correction) by establishing real-time data links with satellite / UAV ground stations, ensuring that the delay from data acquisition to usability is controlled within hours, meeting the timeliness requirements of "dynamic monitoring".

[0035] Spatiotemporal Fusion Data Cube: Solving the Problems of "Scale Fragmentation" and "Spatiotemporal Asynchrony" in Multi-Source Data

[0036] The "Spatiotemporal Fusion Data Cube" is a structured data carrier that fuses the aforementioned multi-source heterogeneous data within a unified spatiotemporal framework. Its core is to eliminate differences in spatial resolution, temporal frequency, and observation dimensions through mathematical modeling, constructing a three-dimensional integrated data structure of "space-time-feature" (essentially a four-dimensional tensor: x, y spatial dimension + t temporal dimension + f feature dimension). The specific construction process includes:

[0037] Data spatiotemporal registration: Using a high-precision wetland digital elevation model (DEM) as the spatial reference, the "point coordinates" of sensor data and the "pixel coordinates" of remote sensing images are unified into the same coordinate system (such as UTM projection); in the time dimension, linear interpolation, spline interpolation and other methods are used to align the high-frequency data (minute level) of the sensor and the low-frequency data (day / week level) of the remote sensing to a unified time axis (such as hour level or day level) to solve the problem of "time asynchrony".

[0038] Multi-scale data fusion: To address the scale differences between "point data (sensor data)" and "area data (remote sensing data)," a fusion strategy combining "top-down" and "bottom-up" approaches is adopted.

[0039] For sensor data: Point data are expanded into raster data (e.g., 10m×10m) with the same resolution as remote sensing images by methods such as Kriging interpolation and inverse distance weighting, while preserving microscopic dynamic features;

[0040] For remote sensing imagery: Enhance the details of low-resolution imagery through super-resolution reconstruction (such as the SRGAN algorithm based on deep learning), and at the same time, use sensor data to correct the remote sensing inversion results (such as using field soil moisture data to optimize the soil moisture content retrieved from remote sensing) to improve the accuracy of macroscopic data.

[0041] The data cube's structural design: The final cube uses "spatial grid (x, y)" as the horizontal axis, "time series (t)" as the vertical axis, and "ecological element characteristics (f)" as the depth axis. Each grid cell (x, y, t) corresponds to a set of multi-dimensional feature values ​​(such as water level, vegetation cover, soil moisture, etc. at that location at time t). This structure preserves the high-frequency dynamics of sensor data while integrating the macroscopic spatial pattern of remote sensing imagery, providing a unified data foundation for subsequent "spatiotemporal correlation analysis".

[0042] Ecological feature tensor extraction: extracting the "state fingerprint" of the ecosystem from the data cube;

[0043] The "ecological feature tensor" is a high-dimensional feature set extracted from a spatiotemporal fusion data cube. Essentially, it transforms raw data into "computable features" that directly reflect the ecological state of wetlands, providing input for subsequent degradation pattern analysis. The extraction process must balance "ecological saliency" (the correlation between features and ecological processes) and "computational operability" (suitability for deep learning model processing), specifically including:

[0044] Feature Dimension Design: Based on the core elements of wetland ecosystems—water, soil, and vegetation—three feature dimensions are designed, each containing multiple levels of features:

[0045] Hydrological characteristics include water level change rate, flood pulse frequency, and water connectivity index (based on remote sensing water boundary changes), reflecting the driving effect of wetland hydrological dynamics on the ecosystem.

[0046] Vegetation characteristics include Normalized Difference Vegetation Index (NDVI), vegetation cover, vegetation phenology (extracted from remote sensing time series data), and vegetation community diversity index (combined with patch classification from high-resolution UAV imagery), which characterize vegetation growth status and community structure.

[0047] Soil characteristic dimensions include the spatiotemporal variation coefficient of soil moisture, the gradient of soil organic matter content, and the salinization index (combined with remote sensing salinity inversion), which reflect the soil environment's ability to support vegetation and organisms.

[0048] Tensor extraction method: Employing a hybrid extraction strategy of "statistical features + deep learning features":

[0049] For structured data (such as water level and soil moisture from sensors): extract time series features (such as mean, variance, trend term, and periodicity) and spatial correlation features (such as water level gradient between a certain grid and surrounding grids).

[0050] For remote sensing image data: spatial features (such as the texture features of vegetation patches and the morphological features of water body boundaries) are extracted using convolutional neural networks (such as U-Net and ResNet), and temporal features (such as the seasonal fluctuation patterns of NDVI) are extracted by combining LSTM.

[0051] Finally, the above multi-dimensional features are integrated into a high-dimensional tensor (dimension: number of spatial grids × number of time steps × number of feature dimensions), each element of which corresponds to a certain ecological feature value of a certain spatial location and a certain time of the wetland.

[0052] The degradation pattern of the ecological feature tensor is analyzed to generate a dynamic topological graph of the ecological state, and the ecological connectivity index is calculated, specifically as follows:

[0053] We analyze the three-dimensional data of hydrological fluctuations, vegetation cover, and soil moisture in the ecological characteristic tensor and construct a dynamic attribute map.

[0054] It should be noted that the ecological characteristic tensor includes three core dimensions: hydrological fluctuations, vegetation cover, and soil moisture. The core of this step is to transform the abstract tensor data into a concrete "graph structure" to achieve the correlation modeling of ecological units: extracting indicators such as water level fluctuations, flood inundation frequency, and water flow path connectivity, focusing on "the driving role of hydrodynamic processes on ecological units" (such as how periodic floods maintain the habitat diversity of wetland vegetation); analyzing vegetation types (emergent / submerged plants), spatiotemporal changes in cover, and succession of dominant species in the community, reflecting "the structural support role of vegetation as an ecosystem engineer"; and paying attention to soil moisture gradient, spatial distribution of organic matter, and degree of salinization, reflecting "the basic carrying capacity of soil for vegetation growth and material cycling".

[0055] Wetland ecological units (such as vegetation patches, water bodies, and soil units) are used as "nodes". The attributes of each node are composed of the above three-dimensional data (for example, the attributes of a vegetation patch node are: [water level fluctuation 0.5m, coverage 80%, soil organic matter 3%]). The "edges" between nodes initially represent the spatial adjacency relationship (for example, patch A is adjacent to water body B).

[0056] The key innovation lies in "dynamics": by slicing time series data (e.g., one time step every 30 days), attribute maps at different times are constructed to capture the temporal changes in node attributes (e.g., a decrease in vegetation cover) and adjacent relationships (e.g., shrinkage of water bodies causing a patch to disconnect from the water source), thus solving the problem that traditional static maps cannot reflect the dynamics of ecological processes.

[0057] The material exchange matrix between ecological units in the attribute map is calculated, and landscape ecology theory is introduced to extract patch structure features.

[0058] It should be noted that material exchange refers to the flow of energy, nutrients, and water between ecological units (such as vegetation patches releasing litter into water bodies and water bodies replenishing water to surrounding soils). It is calculated in the following ways: for example, the amount of water exchanged between adjacent soil units and vegetation patches is calculated using the water diffusion equation; the intensity of nitrogen and phosphorus exchange between water bodies and wetland vegetation is estimated using the nutrient migration model; the intensity of material exchange is indirectly quantified by fitting the attribute correlations of different units in the ecological feature tensor through machine learning (such as random forests) to the correlation between vegetation NDVI changes and surrounding water level changes; finally, a "material exchange matrix" (an n×n matrix, where n is the number of ecological units) is formed, and the matrix element M(i,j) represents the intensity of material exchange from unit i to unit j, with a larger value indicating a closer correlation.

[0059] Introducing the "patch-corridor-matrix" theory from landscape ecology, key indicators reflecting the integrity of ecological structure are extracted from attribute maps: the area and perimeter of each ecological unit (such as a vegetation patch) are calculated to reflect its carrying capacity for organisms (the larger the area, the stronger the carrying capacity); the "perimeter / area ratio" is used to measure the complexity of patch shape (the larger the ratio, the more irregular the shape and the stronger the edge effect); the number of segments of a certain type of ecological unit (such as water area) is calculated to reflect the basic structure of its connectivity (high fragmentation indicates poor connectivity); and the proportion of a certain type of patch (such as dominant vegetation) in the overall landscape reflects the stability of the ecosystem (an excessively high proportion of dominant species may reduce diversity).

[0060] By integrating the material exchange matrix and patch structure characteristics, a weighted dynamic topology map of ecological state is constructed.

[0061] It should be noted that the "edge weight" in the topological graph consists of two parts: one is taken from the material exchange matrix M(i,j), which reflects the actual material flow intensity between units (e.g., M(i,j)=0.8 indicates strong exchange); the other is calculated based on the patch structure characteristics, such as the shape index difference between two adjacent patches (the smaller the difference, the easier it is to form a stable connection) and the product of fragmentation (the lower the fragmentation, the higher the structural weight); the final comprehensive weight of the edge W(i,j)=α×functional weight+(1-α)×structural weight (α is the weight coefficient, which is dynamically adjusted according to the wetland type, such as 0.6 for hydrological-dominant wetlands and 0.4 for vegetation-dominant wetlands).

[0062] It retains three-dimensional ecological characteristics (hydrology, vegetation, and soil) and adds patch structure characteristics (such as area and shape index); it comprehensively reflects functional association and structural support capacity. The higher the weight, the stronger the interdependence between the two units in the ecological process; it is updated over time (e.g., once a month). When a unit degrades (e.g., a sharp drop in vegetation cover), the intensity of material exchange between it and the surrounding units decreases, resulting in a decrease in the weight of the edges, and the topological structure is reconstructed accordingly (e.g., some edges disappear, reflecting the break in ecological association).

[0063] The gradient of spatial permeability change in the topology map during the continuous restoration phase is calculated as an ecological connectivity index.

[0064] It should be noted that spatial permeability refers to "the efficiency of a certain ecological flow (such as water or species) spreading from one node to the entire graph in the topology graph". The calculation method is as follows: each node in the topology graph is assigned a "source strength" (such as the water storage capacity of a water source node and the biomass of a vegetation node); the flow diffusion probability is calculated based on the edge weights (the higher the weight, the greater the diffusion probability), simulating the propagation process of the ecological flow in the entire graph; the permeability is quantified by combining the "average diffusion distance" and the "proportion of nodes reached" (the farther the diffusion and the more nodes covered, the higher the permeability).

[0065] Differential calculations were performed on the permeability during continuous restoration phases (e.g., from t1 to t2, from t2 to t3) to obtain the "spatial permeability change gradient": Positive gradient: Permeability increases, indicating improved ecological connectivity (e.g., a corridor recovers after restoration, and species dispersal efficiency increases); Negative gradient: Permeability decreases, indicating deterioration of connectivity (e.g., degradation of a patch leads to obstruction of ecological flow); The magnitude of the absolute value of the gradient reflects the severity of the change in connectivity (the larger the absolute value, the more significant the change).

[0066] Compared to traditional connectivity indices (such as those based on Euclidean distance), this index has three major advantages:

[0067] Balancing function and structure: It reflects both actual material exchange (function) and plaque structural support (structure); Dynamic response: It updates in real time during the restoration process and can capture the immediate impact of short-term interventions (such as water replenishment) on connectivity; Spatial heterogeneity: It can output the permeability gradient of different regions and accurately locate areas with weak connectivity (providing a basis for subsequent identification of degradation hotspots).

[0068] Based on the ecological connectivity index, restoration needs are identified. A multimodal graph convolutional network is used to locate degradation hotspots and generate a set of coordinates for priority restoration areas. Specifically:

[0069] By mapping the ecological connectivity index to the Riemannian manifold space, an ecological degradation distance matrix is ​​constructed.

[0070] It should be noted that the Riemannian manifold is a curved geometric space that can more accurately describe the intrinsic structure of nonlinear data. Mapping the ecological connectivity index to this space essentially transforms the coupling relationship of "hydrological connectivity-vegetation cover-soil health" into curvature changes on the manifold (the greater the curvature, the more complex the interactions between ecological factors in the region); and calculates the "ecological distance" between regions using manifold metrics (rather than Euclidean distance). For example, two geographically distant regions connected by groundwater levels may appear closer in the manifold space.

[0071] The core of the matrix is ​​to quantify the "degree of degradation difference between regions". The specific steps are as follows: using wetland grid cells (e.g., 10m×10m) as the basic unit, each cell corresponds to an ecological connectivity index value; in the Riemannian manifold space, calculate the "geodesic distance" (the shortest path between two points on the manifold, reflecting the true difference in ecological degradation characteristics) between any two cells (i,j); combine the degradation degree of the cell itself (e.g., the degree to which the connectivity index is below the threshold) to perform weighted correction on the geodesic distance (the more severely degraded the cell, the higher the distance weight to other cells); finally, an n×n ecological degradation distance matrix is ​​formed (n is the total number of grid cells). The matrix element D(i,j) represents the degree of difference in ecological degradation characteristics between cells i and j. The larger the value, the more significant the difference in degradation patterns between the two, and the smaller the value, the more similar the degradation characteristics (possibly belonging to the same degradation hotspot area).

[0072] An attention-weighted graph convolutional network is used to aggregate spatial features of the ecological degradation distance matrix;

[0073] It should be noted that: using wetland grid cells as "nodes" and the ecological degradation distance matrix D(i,j) as the "edge weights" (the smaller the D value, the higher the edge weight, indicating a stronger degradation correlation between the two cells), an "ecological degradation correlation graph" is constructed. GCN aggregates features in the following way: the initial features of each node are its ecological connectivity index and hydrological / vegetation / soil features (from the ecological feature tensor of claim 1); through multi-layer convolution operations, the features of adjacent nodes are aggregated to the central node according to the edge weights (degradation correlation) (e.g., a node will "absorb" features of surrounding nodes with similar degradation patterns), realizing the upgrade from "local degradation features to regional degradation patterns".

[0074] Traditional GCNs aggregate all neighboring nodes with equal weights, while attention mechanisms achieve differentiated weighting by learning "attention coefficients": the attention coefficient α(i,j) of node i to its neighbor node j is calculated using the formula: α(i,j)=softmax(similarity(F_i,F_j)), where F_i and F_j are the feature vectors of nodes i and j, and similarity is the feature similarity function (such as cosine similarity); the stronger the degradation correlation (the smaller D(i,j)) and the more similar the features of neighboring nodes, the larger α(i,j) is, and the higher the proportion of its features in the aggregation; for nodes in edge areas (such as the boundary between wetlands and farmland), their weights are enhanced through "spatial location attention" (these areas are often degradation-sensitive areas).

[0075] Ultimately, the aggregated features output by each node contain both its own degradation information and the most relevant features from the surrounding area, forming a high-dimensional feature vector that reflects the "regional degradation pattern".

[0076] The latent space clustering of feature vectors is performed using a variational autoencoder to identify degradation hotspot regions that exceed the degradation threshold.

[0077] It should be noted that VAE consists of an encoder and a decoder. The core is to compress high-dimensional feature vectors into a low-dimensional "latent space" (such as 2-3 dimensions) while preserving the probability distribution characteristics of the data: Encoder: Maps the feature vectors output by GCN to the probability distribution (mean and variance) in the latent space, and obtains the latent vector z (representing the "condensed features" of the degradation mode in the region) by sampling; Decoder: Reconstructs the original feature vector from the latent vector z, and ensures that each point in the latent space corresponds to a typical degradation mode (such as "hydrologically dominated degradation" or "vegetation cover drop degradation") by minimizing the reconstruction error; Clustering: In the latent space, the DBSCAN (density clustering) algorithm is used to cluster latent vectors with similar features into one class, and each class represents the spatial distribution of a degradation type.

[0078] A dual threshold system is set to identify hotspots requiring priority restoration: Degradation threshold: Based on historical data or ecological red lines, a lower limit for the ecological connectivity index is set (e.g., below 0.3 is considered severe degradation), filtering out areas where the index exceeds the standard; Spatial clustering threshold: The clustering density of degraded areas is required to reach a certain standard (e.g., at least 5 adjacent grid units are continuously degraded), excluding isolated, non-systematic degradation points (which may be due to sensor errors or random fluctuations); Finally, clustered areas that simultaneously meet the above thresholds are marked as "degraded hotspot areas," which are the most vulnerable and critical nodes in the ecosystem that have the greatest impact on overall connectivity.

[0079] Extract the coordinates of hydrological nodes, vegetation patch vectors, and soil pollution range in the degradation hotspot areas to generate a coordinate set of priority remediation areas.

[0080] It should be noted that the extraction of multiple types of key elements specifically includes: hydrological node coordinates: identifying water sources (such as small lakes, perennial waterlogged areas) and hydrological connectivity bottlenecks (such as sections of rivers with interrupted flow) within the hotspot area, and extracting the center point coordinates (such as (x1,y1), (x2,y2)) by combining water body indices (such as NDWI) from remote sensing images with sensor positioning; vegetation patch vectors: performing fine classification of vegetation communities within the hotspot area (such as reed beds, sedge areas), and extracting the boundary vector data (polygon vertex coordinate set) of each patch by semantic segmentation of high-resolution UAV images (such as U-Net model); soil pollution range: combining soil sensor data (such as heavy metal content, pH anomaly areas) and soil salinization index retrieved from remote sensing, delineating the spatial range of the pollution area, and outputting its minimum bounding rectangle or polygon coordinates.

[0081] The spatial information of the above elements is standardized into a unified coordinate system (such as WGS84 latitude and longitude) to form a coordinate set containing three levels of information: core restoration points: coordinates of single points such as hydrological nodes and heavily polluted center points; key restoration surfaces: coordinates of polygonal boundaries such as vegetation patches and polluted areas; spatial relationships: marking the topological relationships between each element (such as a vegetation patch being adjacent to a polluted area, or a hydrological node having been connected to another node by water flow).

[0082] This coordinate set contains both precise spatial location information and preserves the relationships between ecological elements, providing a directly usable "target map" for the next step of "ecological corridor modeling".

[0083] Ecological corridor modeling is performed on the coordinate set of the priority restoration area, and the initial restoration path boundary is generated using a cellular automata evolution algorithm, specifically:

[0084] Analyze the ecological correlation of the coordinate set of the priority restoration area and construct a biological migration corridor model;

[0085] It should be noted that ecological connectivity is manifested as a combination of "functional connectivity" and "process connectivity," specifically analyzing the following core relationships: Species migration connectivity: Based on the habitat requirements of target protected species (such as migratory birds and amphibians), determine whether there are potential migration routes in priority areas (e.g., vegetation patch A and water area B are breeding and foraging grounds for a certain bird, requiring a corridor connection); Hydrological connectivity connectivity: Analyze the potential water flow paths (e.g., groundwater connectivity, surface runoff channels) between hydrological nodes (e.g., water sources, catchment areas) in priority areas to determine whether there are natural or restorable hydrological connections; Energy flow connectivity: Assess the potential for material exchange between vegetation patches and soil units (e.g., litter transport, nutrient diffusion) to identify key energy flow channels.

[0086] The analytical method combines "data-driven" and "ecological theory": potential associations are identified through spatiotemporal correlation data in the ecological feature tensor (such as the synchronicity of NDVI changes between two nodes), while the direction of association is clarified by referring to the "source-sink" theory in landscape ecology (the source is the node with strong ecological function and the sink is the degraded node that needs replenishment).

[0087] The model is a mathematical abstraction of the "optimal migration path of species from source nodes to sink nodes," and includes three key elements: Node attributes: Each priority restoration zone node is assigned "ecological source strength" (such as biomass and water conservation capacity) and "ecological resistance coefficient" (reflecting the migration difficulty within the node, such as low resistance in densely vegetated areas and high resistance in bare land); Corridor cost function: quantifies the difficulty of corridor construction connecting two nodes, calculated based on the "minimum cumulative resistance model," with the formula: Cumulative resistance = ∑(path length × unit resistance coefficient); where the unit resistance coefficient is dynamically adjusted according to land cover type (such as water body < grassland < bare land) and intensity of human disturbance (such as no disturbance < slight disturbance < severe disturbance);

[0088] Multi-scale network structure: To address the differences in migration capabilities among different species (such as small insects and large mammals), a three-level corridor network of "micro-meso-macro" is constructed (micro corresponds to the internal channels of patches, and macro corresponds to regional migration routes) to ensure that the model is adaptable to the needs of multiple species.

[0089] Ant colony optimization algorithm is used to simulate species dispersal paths and generate candidate ecological corridors;

[0090] It should be noted that the core elements of the ant colony algorithm correspond to the design requirements of ecological corridors: individual ants: simulate the diffusion behavior of a single species (such as short-distance flight of birds or crawling on the ground of insects); nodes (cities): correspond to key nodes (source nodes or sink nodes) in the coordinate set of priority restoration areas; path pheromones: represent the "preference" of a species for a certain path. The higher the pheromone concentration, the more suitable the path is for species migration (e.g., pheromones accumulate quickly on paths with good vegetation cover and low resistance); heuristic function: designed based on the ecological resistance coefficient. The lower the resistance of an area, the larger the heuristic function value, guiding ants (species) to prioritize that area.

[0091] Initialization: Set the number of ants (simulating the number of individuals in the species), the initial pheromone concentration (uniform distribution), and the number of iterations (simulating diffusion time); randomly place ants at source nodes (priority areas for restoration with strong ecological functions); path search: each ant starts from the source node and selects the next grid cell based on the probability of "pheromone concentration + heuristic function" (e.g., a cell with high pheromone concentration and low resistance has a high probability of being selected), moving towards the sink node (priority areas with severe degradation); pheromone update: after reaching the sink node, ants deposit pheromones along the path (the shorter the path and the lower the resistance, the more pheromones are deposited); at the same time, a small amount of pheromone will evaporate from all paths (to avoid the algorithm getting trapped in local optima); multi-path generation: by adjusting the number of ants and the number of iterations, multiple different paths are generated (to avoid homogenization), each path corresponding to a potential ecological corridor scheme, forming a candidate corridor set.

[0092] The key innovation lies in the "embedded ecological constraints": "insurmountable areas" (such as heavily polluted areas and man-made structures) and "must-pass areas" (such as key water sources) are added to the path search to ensure that candidate corridors meet the mandatory requirements for wetland ecological protection.

[0093] Calculate the landscape pattern similarity between candidate ecological corridors and historical restoration achievements, and select paths that meet the similarity criteria as the initial restoration path boundaries.

[0094] It should be noted that indicators strongly correlated with restoration effectiveness were selected from the two dimensions of "structure and function" to quantify the similarity between candidate corridors and historical cases.

[0095] Structural indicators:

[0096] Corridor width variation coefficient (reflects the uniformity of width; too narrow and it is prone to breakage, too wide and the cost is high).

[0097] Curvature (the degree of tortuosity of a path, which is related to species migration preferences, such as birds preferring straighter paths and small animals preferring tortuous paths);

[0098] Connectivity of surrounding patches (the number of connections between the corridor and surrounding ecological patches, which affects species diversity);

[0099] Functional indicators:

[0100] Mean ecological resistance (the overall migration resistance level of the corridor, compared with the resistance range of historically effective corridors).

[0101] Species pass rate (the proportion of species that successfully cross the corridor based on historical data simulations).

[0102] It should be noted that the weight of functional indicators (such as species throughput) is higher than that of structural indicators (such as the coefficient of variation of corridor width). Based on the theory in "Landscape Ecology" (Forman, 1995) that the core value of ecological corridors lies in maintaining species migration function, and combined with the Analytic Hierarchy Process (AHP) calculation, the following was determined: by inviting 5 experts in the field of wetland ecology to score the importance of indicators such as "species throughput, corridor width, and tortuosity", after consistency test (CR<0.1), the weight of species throughput was found to be 0.3, and the average weight of structural indicators was 0.25.

[0103] Similarity Calculation and Screening Criteria

[0104] Data standardization: Normalize the various indicators of candidate corridors and historical effective corridors (eliminate differences in dimensions).

[0105] Weighted similarity calculation: The overall similarity is calculated using cosine similarity or Euclidean distance, and the formula is as follows:

[0106]

[0107] in, As candidate corridor indicators, As an indicator of historical corridors, The weights are assigned to the indicators (functional indicators have higher weights than structural indicators, such as setting the species pass rate weight to 0.3).

[0108] Threshold screening: Set a similarity threshold (e.g., ≥0.7) to screen out candidate corridors that meet the criteria; if there are multiple compliant paths, select the path with the "lowest cumulative resistance" and "most coverage repair priority areas" as the optimal solution.

[0109] For the selected optimal candidate corridors, expand them to both sides based on their centerline (the expansion width is determined according to the needs of the target species, such as 50-100m for large mammals and 5-10m for insects), forming a strip-shaped area containing a "core area (species migration corridor)" and a "buffer zone (to reduce external interference)". The spatial boundary (polygon vertex coordinates) is the "initial repair path boundary".

[0110] The ecological correlation of the coordinate set of the priority restoration area is analyzed, and a biological migration corridor model is constructed, specifically as follows:

[0111] Each priority restoration zone coordinate point is treated as a cell node, and the node attributes include the ecological resistance coefficient.

[0112] It should be noted that, with the coordinate set of the priority restoration area (such as hydrological nodes and the center point of vegetation patches) as the core, and combined with the spatial continuity of wetland ecological units, the entire study area is divided into regular grid cells (such as 5m×5m or 10m×10m). Each cell node contains two types of information: spatial coordinates: precise geographical location information (such as latitude and longitude) to ensure the node's location in physical space; attribute characteristics: including land cover type (water body, vegetation, bare land, etc.), soil type, intensity of human disturbance (such as distance from roads / buildings), and suitability for target species habitats. These characteristics together determine the node's "ecological resistance coefficient".

[0113] The ecological resistance coefficient is a core indicator for measuring the difficulty of species migration within a given cell node; the greater the resistance, the more difficult it is for the species to pass through. Its calculation is based on an "ecological characteristic-resistance mapping" model: Baseline resistance assignment: A baseline value is set according to land cover type (e.g., water resistance = 10, grassland resistance = 30, bare land resistance = 80, artificial structure resistance = ∞ (insurmountable)).

[0114] It should be noted that the baseline values ​​for ecological resistance are based on the 'Ecological Resistance Coefficient Grading Standard' in *Principles and Applications of Landscape Ecology* (Fu Bojie et al., 2019), adjusted according to the characteristics of wetland ecosystems: water bodies provide drinking water and migration routes for species, so the resistance is set at 10 (low resistance); grasslands are the habitat for most wetland species, so the resistance is set at 30 (medium to low resistance); bare land lacks vegetation protection, so the resistance is set at 80 (high resistance); artificial structures hinder species migration, so the resistance is set at ∞ (insurmountable). These values ​​have been verified through preliminary experiments: in a typical freshwater wetland, the species migration paths simulated using these resistance values ​​show an 82% agreement with field tracking data (GPS-guided migratory bird routes).

[0115] Correction factors: Dynamic variables are introduced for adjustment, such as: vegetation cover correction: the higher the cover, the lower the resistance (e.g., if the cover is >70%, the resistance is multiplied by 0.7); hydrological connectivity correction: if the distance between the cell and the water source is <50m, the resistance is multiplied by 0.5 (species are more likely to migrate in areas near water); disturbance intensity correction: the closer to the human activity area (e.g., <100m), the resistance is multiplied by 1.5 (disturbances increase the difficulty of migration); finally, the ecological resistance coefficient of each cell node = basic resistance × product of each correction factor, and the value range is usually standardized to 1-100 (1 is the minimum resistance, 100 is the maximum resistance).

[0116] Based on species migration patterns and hydrological connectivity, the probability of ecological transition between nodes is calculated;

[0117] It should be noted that, for key wetland species (such as migratory birds, amphibians, and aquatic insects), the core parameters affecting their migration are extracted: migration capacity: maximum single migration distance (small insects <50m, migratory birds >1000m), daily activity range, terrain preference (e.g., birds prefer the edge of open water, frogs prefer moist meadows); environmental sensitivity: tolerance thresholds to water quality, vegetation density, and disturbance intensity (e.g., a certain frog only moves around waters with pH 6.5-7.5); seasonal rhythms: for example, migratory birds need to approach shallow mudflats during the breeding season and approach deep water areas during the wintering season. These parameters are obtained through field surveys, species tracking data, or literature reviews, forming a "species migration parameter database".

[0118] Hydrology is a core driving factor in wetland ecosystems, and its connectivity directly affects the exchange of materials and species migration paths between nodes. It is mainly quantified through the following indicators: Water flow connectivity: Calculate the water flow velocity and water level difference between nodes based on hydrological sensor data to reflect the connectivity potential of surface water; Groundwater gradient: Analyze the direction and intensity of groundwater infiltration between nodes through a soil moisture sensor network to identify hidden hydrological connections; Flood pulse frequency: Statistically count the frequency of nodes being flooded during the rainy season. The higher the frequency, the more likely it is to become a seasonal migration channel.

[0119] Considering both species migration patterns and hydrological connectivity, a "probability-weighted model" is used to calculate the transition probability P(i,j) between adjacent nodes (i,j):

[0120] P(i,j)=α×f+(1-α)×g

[0121] f is the species' suitability score for node j (calculated based on habitat preference, ranging from 0 to 1); g is the hydrological connectivity strength between nodes i and j (normalized to 0-1, with higher values ​​indicating better connectivity); α is the weighting coefficient (adjusted according to the species' dependence on hydrology, such as α=0.3 for fish and α=0.7 for terrestrial birds).

[0122] It should be noted that the α value is determined based on the species' dependence on hydrology, referring to the 'Species Niche Characteristics' in *Wetland Biology* (Mitsch, 2015): fish survival and reproduction are highly dependent on hydrological connectivity (hydrological factors explain >70% of their habitat selection contribution), hence α=0.3 (hydrological connectivity has a higher weight); terrestrial birds migration is more dependent on vegetation habitat (vegetation factors contribute >60%), hence α=0.7 (species adaptability has a higher weight). Specific values ​​were obtained by fitting the correlation between species occurrence probability and environmental factors using a logistic regression model, with all models showing R² > 0.65.

[0123] Ultimately, the higher the P(i,j) value, the greater the likelihood of a species migrating from i to j, making it an important "attractiveness indicator" of corridor pathways.

[0124] The minimum cumulative resistance model is used to quantify the corridor construction cost between nodes;

[0125] It should be noted that the formula for calculating the cumulative resistance cost from node i to node j is:

[0126]

[0127] Wherein: the path consists of n cellular nodes, k is the kth node on the path; D(k) is the spatial distance of the kth node (usually the Euclidean distance or Manhattan distance to the previous node); R(k) is the ecological resistance coefficient of the kth node (from step one); "min" means taking the minimum cumulative resistance among all possible paths (i.e., the optimal path cost).

[0128] The "cost" here refers not only to the economic cost of engineering construction, but specifically to: ecological cost: the higher the cumulative resistance, the more ecologically sensitive areas the corridor traverses (e.g., bare land areas with high resistance require more vegetation restoration measures); economic cost: the intensity of engineering measures is inferred from the cumulative resistance (e.g., bare land with a resistance of 80 requires soil covering + vegetation planting, and the unit cost is 3 times that of grassland with a resistance of 30); time cost: paths with high cumulative resistance require a longer restoration period (e.g., severely degraded areas may require 2-3 years of soil improvement, while mildly degraded areas only require 1 year).

[0129] By integrating the probability of ecological leaps with the construction cost, a multi-scale ecological corridor network is generated.

[0130] It should be noted that for each potential corridor path, its "Comprehensive Suitability Index" is calculated using the following formula:

[0131]

[0132] In the formula, P is the average ecological transition probability of the path (the higher the probability, the more suitable the species migration); C is the cumulative resistance cost of the path. β is the normalized cost (the lower the value, the more economically feasible); β is the function-cost balance coefficient (β=0.7 for ecological priority areas and β=0.5 for cost-sensitive areas).

[0133] Based on the spatial scale of the pathways and the target populations they serve, the corridor network is divided into three levels: Micro-corridors (1-10m wide): serving small species (insects, amphibians), with vegetation patches in priority restoration areas as nodes, and the pathways must meet the requirements of "high vegetation cover + low resistance", with a comprehensive suitability index S≥0.8; Meso-corridors (10-50m wide): serving medium-sized species (small mammals, waterbirds), connecting hydrological nodes and vegetation patches, and the pathways must consider "hydrological connectivity + moderate resistance", with S≥0.7; Macro-corridors (50-200m wide): serving large species (migratory birds, deer), connecting regional priority restoration areas (such as different wetland patches), and the pathways can tolerate a certain level of resistance but must ensure "seasonal connectivity", with S≥0.6.

[0134] The final multi-scale corridor network must meet the following requirements: Connectivity: Any two nodes in the same priority restoration area must be connected by at least one corridor; Redundancy: Key nodes (such as core water sources) must have 2-3 alternative corridors to cope with sudden environmental changes (such as local droughts); Hierarchical correlation: Micro-level corridors must have intersection nodes with meso-level corridors, and meso-level and macro-level corridors must have intersection nodes to form an organic whole of "small-scale nested large-scale".

[0135] An ecological benefit optimization function is constructed, and an adaptive multi-objective evolutionary algorithm is used to collaboratively optimize the initial restoration path boundary, outputting the optimal restoration scheme and dynamically adjusting the restoration equipment. Specifically:

[0136] Construct an ecological benefit optimization function that includes biodiversity gains, restoration costs, and engineering sustainability;

[0137] It should be noted that the core indicators for measuring the degree of improvement in wetland species richness and community structure after restoration include: the species recovery index (SRI). : The ratio of the number of target species (such as endangered waterbirds and endemic plants) after restoration to the historical baseline; Community evenness index ( ): Restore the evenness of the distribution of individuals of each species within the region (avoiding the dominance of a single species); increase the habitat connectivity rate ( ): The difference between the species throughput of the restored ecological corridor and that before restoration.

[0138] The expression for biodiversity gain is: f1(X) = × + × + ×

[0139] Where X represents the decision variables for the restoration plan (such as corridor width, vegetation planting density, etc.). , , Weight (ecologically sensitive area) (Higher weight).

[0140] The economic input for implementing the quantified plan covers both direct and indirect costs: Direct costs include engineering materials (such as vegetation seedlings and soil conditioners), labor costs, and equipment rental (such as wetland excavators and water purification equipment); Indirect costs include ecological compensation (such as compensation for temporary occupation of surrounding land) and monitoring costs (sensor maintenance and data transmission).

[0141] The function expression is: f2(X) = ∑(unit cost × project volume) + annual maintenance cost × repair cycle

[0142] The assessment scheme's ability to cope with natural disturbances (such as floods and droughts) and human activities in the long term (e.g., 10-20 years) is evaluated as follows: Disturbance resistance index: the structural integrity of the restoration area under a 50-year flood (e.g., the proportion of corridors destroyed); Self-sustainability: the natural maintenance time of the restored ecosystem (e.g., vegetation communities, hydrological cycles) without human intervention; Landscape stability index: the interannual change rate of land cover type in the restoration area (the smaller the change, the more stable).

[0143] The expression is: f3(X) = 0.4 × anti-interference index + 0.4 × self-sustaining ability + 0.2 × landscape stability index.

[0144] The three objectives are inherently conflicting: for example, maximizing biodiversity gains may require expanding the restoration area (increasing costs), while minimizing costs may lead to simplified restoration measures (reducing sustainability). Therefore, constraints need to be set: ecological red line constraint: vegetation coverage in the core area after restoration ≥ 60%, water quality meets standards (e.g., COD ≤ 50 mg / L); engineering feasibility constraint: corridor width ≥ minimum requirements of target species (e.g., migratory bird corridor ≥ 50 m); cost budget constraint: total investment does not exceed 120% of the preset threshold.

[0145] The initial repair path boundary is encoded as a gene population, and the improved MOEA / D algorithm is used for decomposition and optimization.

[0146] It should be noted that the key parameters of the initial repair path boundary are transformed into "genes" (decision variables) that the algorithm can process, using a hybrid approach of "real number encoding + binary encoding": Real number encoding describes continuous parameters, such as corridor width (5-200m), vegetation planting density (10-50 plants / m²), and soil improvement depth (0.3-1.5m); Binary encoding describes discrete parameters, such as vegetation type (01=reed, 10=cattail, 11=mixed planting) and engineering measures (0=natural restoration, 1=human intervention). Each "individual" (candidate solution) corresponds to a gene sequence, and the population size is set according to the complexity of the problem (e.g., 50-200 individuals).

[0147] The core of MOEA / D is to decompose a multi-objective optimization problem into N single-objective sub-problems (N being the population size), with each sub-problem corresponding to a weight vector. By collaboratively optimizing all sub-problems, the Pareto optimal front is approximated. The improvements of this invention are reflected in: Weight vector initialization: A uniformly distributed weight vector is generated using a Sobol sequence (avoiding the uneven distribution problem of traditional random generation), ensuring that sub-problems cover all objective combinations (such as "high ecology - medium cost," "medium ecology - low cost," etc.); Neighborhood cooperation mechanism: Each sub-problem only communicates with neighboring sub-problems (with similar weight vectors), reducing computational load; simultaneously, the neighborhood size is dynamically adjusted according to the optimization progress of the sub-problems (the neighborhood of sub-problems that optimize quickly shrinks, focusing on local search); Ecological response prediction integration: A two-layer LSTM network (claim 7) is introduced to predict the ecological response trend after the implementation of the restoration plan (such as changes in vegetation cover after one year), and the prediction results are used as the basis for optimizing sub-problems, improving the foresight of the plan.

[0148] A simulated annealing mechanism is introduced to perform dynamic neighborhood search and generate non-dominated solution sets;

[0149] It should be noted that the temperature parameter setting is: initial temperature. Based on the initial difference setting of the objective function (the larger the difference, the better). The higher the temperature, the more exponentially it decreases with each iteration. = × , =0.95 is the cooling coefficient); Neighborhood search strategy: For the current optimal solution of each subproblem, randomly generate neighborhood solutions (e.g., fine-tune corridor width ±10%, change vegetation type in 10% of the area); Acceptance criterion: If the subproblem objective value of the neighborhood solution is better than the current solution, it is directly accepted; otherwise, it is rejected with probability P=exp(−Δf / Accept (Δf is the difference in the target value, (This refers to the current temperature). The higher the temperature, the greater the probability of accepting a worse solution (encouraging the exploration of new regions); as the temperature decreases, only slightly worse solutions are accepted (focusing on local optimization).

[0150] After multiple iterations (e.g., 100-500 generations), the algorithm selects "non-dominated solutions": solutions for which no other solution is superior in all objectives. For example, if solution A has a higher biodiversity gain than solution B while having a lower cost, then A dominates B, and B is eliminated. The final set of non-dominated solutions contains various solutions that balance "ecology, cost, and sustainability," forming a Pareto optimal frontier.

[0151] The optimal repair scheme is selected from the non-dominated solution set based on the fuzzy integral method, and the repair equipment is driven to execute it.

[0152] It should be noted that fuzzy integrals are suitable for handling decision-making problems in ecological restoration where "incommensurability of indicators and uncertainty of weights" exist. Their core is to comprehensively evaluate schemes through "membership functions" and "fuzzy measures": Membership degree calculation: The three objective values ​​of each scheme are converted into membership degrees of 0-1 (higher values ​​are better). For example, the biodiversity gain f1=0.8 corresponds to a membership degree of... =0.8; Fuzzy measure assignment: Based on expert experience or historical data, set interaction weights between objectives (e.g., "synergistic weight between biodiversity and sustainability" is higher than "cost and sustainability"); Fuzzy integral calculation: For each option, calculate the comprehensive score using Choquet integral (a commonly used fuzzy integral):

[0153]

[0154] Where σ(i) represents the membership ranking, g represents the fuzzy measure, and Aσ(i) represents the set of the first i targets. The solution with the highest score is the optimal repair solution.

[0155] The optimal solution needs to be translated into specific operating instructions for the equipment to achieve real-time dynamic adjustment: Instruction generation: The solution parameters (such as "planting reeds in a 30m wide corridor at a density of 20 plants / ㎡") are parsed into equipment control instructions (such as the travel route of the seeder and the seeding density parameters); Real-time feedback: The ecological perception module (claim 8) continuously collects sensor data of the restoration area (such as vegetation survival rate and soil moisture) and calculates the deviation between the actual effect and the expected solution; Dynamic adjustment: If the deviation exceeds the threshold (such as the survival rate being lower than 70%), the optimization algorithm is triggered to recalculate the local solution and adjust the equipment parameters (such as increasing the irrigation frequency and replanting seedlings), forming a closed-loop control of "monitoring-evaluation-adjustment".

[0156] The initial repair path boundary is encoded as a gene population, and an improved MOEA / D algorithm is used for decomposition and optimization, specifically:

[0157] The weight vector is initialized using a Sobol sequence to ensure a uniform distribution of the solution set;

[0158] It should be noted that the Sobol sequence is a quasi-random sequence with the following core characteristics: uniform distribution: in high-dimensional space, the sequence points can uniformly fill the entire feasible region, avoiding the "clustering" or "blank" phenomenon of random sequences; repeatability: given initial parameters, the sequence generation result is unique, ensuring the stability of the algorithm; low computational complexity: the generation efficiency is higher than other quasi-random sequences (such as Halton sequences), making it suitable for large-scale weight vector generation.

[0159] For the three-objective optimization problem (biodiversity, cost, and sustainability) of wetland restoration, the weight vector is a three-dimensional vector λ=(λ1,λ2,λ3), satisfying λ1+λ2+λ3=1 and λi≥0. The steps for initializing using Sobol sequences are as follows: set the population size N (e.g., 100-200, adjusted according to the complexity of the restoration area); generate N three-dimensional Sobol sequence points, mapping them to the feasible region (simplex space) of the weight vector; normalize the vector to ensure that the sum of the weights is 1.

[0160] The resulting weight vector can uniformly cover all possible target combinations such as "high ecology-high cost", "medium ecology-medium cost", and "low ecology-low cost", avoiding the algorithm from over-focusing on a certain type of solution (such as optimizing only cost while ignoring ecological benefits), and providing comprehensive sub-problem coverage for subsequent decomposition and optimization.

[0161] Design a two-layer LSTM network to predict ecological response trends and adaptively adjust the neighborhood size;

[0162] It should be noted that LSTM (Long Short-Term Memory) is good at processing time-series data. Its two-layer structure further enhances the feature extraction capability. Its design is specifically designed to address the "long time delay and nonlinearity" characteristics of wetland ecological response: Input layer: receives historical restoration data, including: restoration measure parameters (such as corridor width, vegetation type, soil improvement intensity); environmental variables (such as rainfall, temperature, initial ecological state); ecological response indicators (such as the time-series changes in vegetation cover, species number, and water quality parameters).

[0163] The first layer of LSTM extracts short-term ecological response features (such as changes in vegetation survival rate 1-3 months after restoration); the second layer of LSTM captures long-term trend features (such as biodiversity recovery curves over 1-2 years).

[0164] Output layer: Predicts the ecological response trend over the next 6-12 months (e.g., "If scheme X is adopted, vegetation coverage can reach 75%±5% after 6 months").

[0165] Adaptive adjustment mechanism for neighborhood size

[0166] Based on the prediction results of LSTM, the neighborhood size T of each sub-problem is dynamically adjusted: the deviation δ between the predicted value and the target value is calculated (e.g., if the target is 80% vegetation coverage and the prediction is 60%, then δ=20); if δ≧th threshold (e.g., 15%): it indicates that the current solution is far from the ecological goal, and the neighborhood needs to be expanded (T=Tmax, e.g., 20-30 neighboring sub-problems) to introduce more diverse information to explore new solutions; if δ=5: it indicates that the solution is close to the goal, and a medium neighborhood (T=Tmid, e.g., 10-20) is adopted to balance exploration and utilization; if δ<5: it indicates that the solution is close to the optimal solution, and the neighborhood is reduced (T=Tmin, e.g., 5-10) to focus on local fine-tuning optimization.

[0167] This mechanism enables the algorithm to dynamically adjust its search strategy based on the actual progress of ecological restoration, avoiding the "one-size-fits-all" problem of traditional fixed neighborhoods, and improving the matching degree between the solution and ecological goals while ensuring optimization efficiency.

[0168] A landscape shape index is introduced as a diversity maintenance operator to prevent premature convergence of the solution set.

[0169] It should be noted that the Landscape Shape Index (LSI) is a core indicator in landscape ecology for measuring the complexity of patch shape. A higher LSI value indicates a more irregular patch shape (a stronger edge effect, which is more conducive to biodiversity). The calculation formula is:

[0170] Where P is the perimeter of the restoration path boundary and A is its area. For wetland ecological corridors: corridors with regular shapes (such as straight ones) have small LSI values ​​and weak edge effects, which are not conducive to species habitat; corridors with tortuous shapes (such as those along natural hydrological trends) have large LSI values ​​and high edge heterogeneity, which can provide richer habitats (such as the alternating distribution of shoals, marshes and meadows).

[0171] LSI is integrated into the selection operator of MOEA / D to prevent solution set convergence by "penalizing similar shapes and rewarding diverse shapes": For candidate schemes in each generation of the population, the LSI value of their repair path is calculated; the LSI similarity (such as cosine similarity) between schemes is calculated. If the LSI similarity of a scheme with an existing dominant scheme is ≥0.8 (highly similar in shape), its selection probability is reduced (multiplied by a penalty coefficient of 0.5); for schemes whose LSI value is in the "ecologically suitable range" (such as LSI∈[1.5,3.0] determined based on historical effective repair cases) and whose shape is greatly different from other schemes, their selection probability is increased (multiplied by a reward coefficient of 1.2); in the crossover and mutation operations, random perturbations are forcibly introduced into schemes with a single shape (such as randomly increasing the corridor tortuosity) to promote the generation of new shape schemes.

[0172] This mechanism ensures that the algorithm retains diverse corridor shape schemes during the iteration process, avoiding premature convergence and conforming to the ecological principle that "shape heterogeneity supports biodiversity" in wetland ecological corridors.

[0173] Secondly, the parameters are not fixed values, but can be dynamically adjusted according to wetland type, target species, etc. The specific steps to obtain them are as follows:

[0174] All weights and coefficients were obtained using a three-step method: theoretical setting → experimental calibration → field adjustment. The scope was initially set based on ecological theory; preliminary experiments were conducted in typical wetlands (such as the Sanjiang Plain wetlands and Poyang Lake wetlands) to optimize the parameters using a controlled variable method (e.g., fixing other parameters and adjusting the α value to observe changes in species throughput); when applied to specific remediation areas, fine-tuning was performed based on field survey data (such as the number of target species and soil type), with adjustments not exceeding 20% ​​of the initial values.

[0175] Supplement: "For coastal wetlands (where tidal influence is significant), the 'intertidal zone' resistance in the basic resistance assignment is set to 20 (between water and grassland), because the tidal flooding period provides migration channels for species, and the exposed period is a foraging area; the coefficient α is set to 0.5 for wading birds (such as egrets), because it depends on both hydrology (intertidal zone) and vegetation (reed beds), and this value is determined by analyzing the proportion of habitat markers (aquatic plants vs terrestrial plants) in the DNA of wading bird feces."

[0176] Example 2, Figure 4This invention presents a deep learning-based dynamic monitoring method for wetland ecological restoration, including:

[0177] The ecological sensing module is used to collect multi-source ecological data in real time and construct a spatiotemporal fusion data cube;

[0178] The status diagnosis module is used to generate dynamic topology maps of the ecological status and connectivity indices;

[0179] The hotspot localization module is used to identify degraded hotspot areas and generate a set of coordinates for priority repair areas;

[0180] The corridor simulation module is used to construct biological migration corridor models and generate initial repair path boundaries;

[0181] The solution generation module is used to perform multi-objective optimization and output the optimal repair solution.

[0182] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0183] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0184] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0185] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0186] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0187] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic monitoring method for wetland ecological restoration based on deep learning, characterized in that, Includes the following steps: Real-time acquisition of multi-source wetland ecological sensor data and remote sensing image streams, construction of spatiotemporal fusion data cube, and extraction of ecological feature tensors; Degradation pattern analysis is performed on the ecological feature tensor to generate a dynamic topological graph of the ecological state, and the ecological connectivity index is calculated. Based on the ecological connectivity index, restoration needs are identified, and degradation hotspots are located through a multimodal graph convolutional network to generate a set of coordinates for restoration priority areas. Ecological corridor modeling is performed on the coordinate set of the priority restoration area, and the initial restoration path boundary is generated using a cellular automata evolution algorithm. An ecological benefit optimization function is constructed, and an adaptive multi-objective evolutionary algorithm is used to collaboratively optimize the initial restoration path boundary, output the optimal restoration scheme, and dynamically adjust the restoration equipment. The degradation pattern analysis of the ecological feature tensor generates a dynamic topological graph of the ecological state, and the ecological connectivity index is calculated, specifically as follows: We analyze the three dimensions of hydrological fluctuations, vegetation cover, and soil moisture in the ecological characteristic tensor and construct a dynamic attribute map. The material exchange matrix between ecological units in the attribute map is calculated, and landscape ecology theory is introduced to extract patch structure features. By integrating the material exchange matrix and patch structure characteristics, a weighted dynamic topology map of ecological state is constructed. Calculate the spatial permeability gradient of the topology map during the continuous restoration phase, and use it as an ecological connectivity index; The process of modeling ecological corridors based on the coordinate set of priority restoration areas and generating initial restoration path boundaries using a cellular automata evolution algorithm is as follows: Analyze the ecological correlation of the coordinate set of the priority restoration area and construct a biological migration corridor model; Ant colony optimization algorithm is used to simulate species dispersal paths and generate candidate ecological corridors; Calculate the landscape pattern similarity between candidate ecological corridors and historical restoration achievements, and select paths that meet the similarity criteria as the initial restoration path boundaries.

2. The wetland ecological restoration dynamic monitoring method based on deep learning according to claim 1, characterized in that, The process of identifying restoration needs based on the ecological connectivity index, and locating degradation hotspots using a multimodal graph convolutional network to generate a set of coordinates for priority restoration areas, specifically involves: By mapping the ecological connectivity index to the Riemannian manifold space, an ecological degradation distance matrix is ​​constructed. An attention-weighted graph convolutional network is used to aggregate spatial features of the ecological degradation distance matrix; The latent space clustering of feature vectors is performed using a variational autoencoder to identify degradation hotspot regions that exceed the degradation threshold. Extract the coordinates of hydrological nodes, vegetation patch vectors, and soil pollution range in the degradation hotspot areas to generate a coordinate set of priority remediation areas.

3. The method for dynamic monitoring of wetland ecological restoration based on deep learning according to claim 2, characterized in that, The ecological correlation of the coordinate set of the priority restoration area is analyzed to construct a biological migration corridor model, specifically as follows: Each priority restoration zone coordinate point is treated as a cell node, and the node attributes include the ecological resistance coefficient. Based on species migration patterns and hydrological connectivity, the probability of ecological transition between nodes is calculated; The minimum cumulative resistance model is used to quantify the corridor construction cost between nodes; By integrating the probability of ecological leaps with the construction cost, a multi-scale ecological corridor network is generated.

4. The wetland ecological restoration dynamic monitoring method based on deep learning according to claim 3, characterized in that, The construction of the ecological benefit optimization function employs an adaptive multi-objective evolutionary algorithm to collaboratively optimize the initial restoration path boundary, outputting the optimal restoration scheme and dynamically adjusting the restoration equipment. Specifically: Construct an ecological benefit optimization function that includes biodiversity gains, restoration costs, and engineering sustainability; The initial repair path boundary is encoded as a gene population, and the improved MOEA / D algorithm is used for decomposition and optimization. A simulated annealing mechanism is introduced to perform dynamic neighborhood search and generate non-dominated solution sets; The optimal repair scheme is selected from the non-dominated solution set based on the fuzzy integral method, and the repair equipment is driven to execute it.

5. The method for dynamic monitoring of wetland ecological restoration based on deep learning according to claim 4, characterized in that, The initial repair path boundary is encoded into a gene population, and the improved MOEA / D algorithm is used for decomposition and optimization. Specifically: The weight vector is initialized using a Sobol sequence to ensure a uniform distribution of the solution set; Design a two-layer LSTM network to predict ecological response trends and adaptively adjust the neighborhood size; A landscape shape index is introduced as a diversity maintenance operator to prevent premature convergence of the solution set.

6. A deep learning-based dynamic monitoring system for wetland ecological restoration, used to implement the method described in any one of claims 1-5, characterized in that, include: The ecological sensing module is used to collect multi-source ecological data in real time and construct a spatiotemporal fusion data cube; The status diagnosis module is used to generate dynamic topology maps of the ecological status and connectivity indices; The hotspot localization module is used to identify degraded hotspot areas and generate a set of coordinates for priority repair areas; The corridor simulation module is used to construct biological migration corridor models and generate initial repair path boundaries; The solution generation module is used to perform multi-objective optimization and output the optimal repair solution.

Images