Method for monitoring long-term changes in salt marsh vegetation

Abstract

The application provides a salt marsh vegetation long-term change monitoring method, and belongs to the technical field of salt marsh vegetation monitoring. A multi-source sensor is configured in a multi-level monitoring grid with a main monitoring point interval of 500 m and an auxiliary monitoring point interval of 100 m to obtain environmental data. A wavelet transform technology is used to separate seasonal changes and long-term change trends, and a seasonal change amplitude calculation matrix equation is used to dynamically adjust a sampling frequency. An ecological network connectivity analysis algorithm based on graph theory is used to calculate a network clustering coefficient and a shortest path length to evaluate habitat fragmentation. A fuzzy C-means clustering algorithm is used to process the fuzziness of a vegetation community boundary based on multi-dimensional environmental factors. A multi-scale matrix mapping hierarchical processing system is established to realize pixel-to-landscape scale information fusion, and the technical problem of insufficient monitoring accuracy of salt marsh vegetation long-term change monitoring resulting in inaccurate monitoring results is solved.

Description

Technical Field

[0001] This invention belongs to the field of salt marsh vegetation monitoring technology, and more specifically, relates to a method for monitoring long-term changes in salt marsh vegetation. Background Technology

[0002] Monitoring vegetation change in salt marsh wetlands is a crucial technical means for assessing the effectiveness of ecological restoration and managing ecosystems. Traditional monitoring methods primarily rely on long-term observation using single-frequency remote sensing imagery combined with sparse ground sampling points. Simple spectral index calculations and statistical analysis are used to identify changes in vegetation cover and community structure succession patterns. In current salt marsh wetland monitoring applications, these traditional techniques are widely used in typical salt marsh areas such as the Bohai Bay and the Yellow River Delta to track vegetation dynamics, providing fundamental data support for wetland protection and restoration projects. However, existing technologies lack effective multi-timescale signal decomposition techniques, making it difficult to accurately separate seasonal changes from long-term evolutionary trends in vegetation spectral data. Furthermore, fixed sampling frequency monitoring modes cannot adapt to the actual needs of different change stages, leading to the omission of key change information or the generation of redundant data. Existing vegetation community delineation methods often employ hard boundary identification techniques, which cannot effectively handle the gradual transition characteristics between vegetation communities in salt marsh ecosystems. Traditional point-based monitoring network layouts lack hierarchical design, making it difficult to simultaneously capture both macroscopic regional changes and local detailed features. In other words, existing technologies suffer from insufficient accuracy in monitoring long-term changes in salt marsh vegetation, leading to inaccurate monitoring results. Summary of the Invention

[0003] In view of this, the present invention provides a method for monitoring long-term changes in salt marsh vegetation, which can solve the technical problem that the monitoring accuracy of long-term changes in salt marsh vegetation is insufficient in the prior art, resulting in inaccurate monitoring results.

[0004] This invention is implemented as follows: This invention provides a method for monitoring long-term changes in saline marsh vegetation, comprising: establishing a multi-level monitoring grid within a saline marsh wetland restoration area; setting main and auxiliary monitoring points using a hierarchical spatial sampling method; configuring soil temperature sensors, soil moisture sensors, and light intensity sensors to acquire soil temperature data, soil moisture data, and light intensity data; acquiring vegetation spectral reflectance data using multispectral remote sensing technology; establishing a vegetation spectral feature database; and outputting vegetation spectral reflectance data; decomposing the vegetation spectral reflectance data into multiple time scales using wavelet transform technology; and obtaining adjusted data using a seasonal variation amplitude calculation matrix equation and a sampling frequency adjustment function. The algorithm is used to: calculate network clustering coefficients and shortest path lengths; construct a graph theory-based ecological network connectivity analysis algorithm to calculate network clustering coefficients and shortest path lengths, and output network clustering coefficient data and shortest path length data; use fuzzy C-means clustering algorithm to divide vegetation communities and output vegetation community division results; use a matrix rank deficit detection algorithm to establish a hierarchical processing system for multi-scale matrix mapping and output multi-scale ecological characteristic data; input network clustering coefficient data, shortest path length data, and vegetation community division results into a wetland ecological evolution prediction model; use a hierarchical fusion weight adjustment function to calculate hierarchical fusion weight parameters; embed particle swarm optimization algorithm into the wetland ecological evolution prediction model and output vegetation change trend prediction results.

[0005] Specifically, the hierarchical spatial sampling method involves setting the density of sampling points according to different importance levels based on the spatial heterogeneity characteristics of the ecosystem. The main monitoring points are used to obtain representative data of the region, while the auxiliary monitoring points are used to capture details of local changes.

[0006] The wetland ecological evolution prediction model is specifically based on a multi-layer neural network with a hierarchical attention network architecture, which includes an input layer, multiple hidden layers and an output layer. The hidden layers use a hierarchical attention mechanism to process multi-scale ecological features.

[0007] Specifically, the hierarchical fusion weight adjustment function is used to dynamically adjust the hierarchical fusion weight parameters of the wetland ecological evolution prediction model to optimize the prediction performance. The inputs include soil salinity concentration data, soil moisture content data, light intensity data, and normalization coefficients, and the output is the hierarchical fusion weight parameters.

[0008] The step of establishing a vegetation spectral feature database specifically involves combining ground-measured chlorophyll content data, plant water content data, and plant stem diameter data with continuous observation data over a time series of 35 years. The vegetation spectral feature database is a structured data set that stores vegetation spectral reflectance data and its corresponding ecological parameters.

[0009] The step of multi-timescale decomposition specifically involves separating seasonal variations from long-term trends. The wavelet transform technique is a time-frequency analysis method that can decompose signals into different frequency components, thereby separating seasonal fluctuations and long-term trends in vegetation data.

[0010] Specifically, the seasonal variation amplitude calculation matrix equation is obtained by processing vegetation spectral reflectance data to obtain the seasonal variation amplitude value, which is used to quantify the degree of change of vegetation spectral reflectance between different seasons. The input includes vegetation spectral reflectance data, time series identifier and band weight coefficient, and the output is the seasonal variation amplitude value.

[0011] Specifically, the sampling frequency adjustment function is used to dynamically adjust the data acquisition frequency based on the seasonal variation amplitude to optimize monitoring efficiency. The inputs include the seasonal variation amplitude, the baseline sampling frequency, and the adjustment coefficient, and the output is the adjusted sampling frequency.

[0012] Specifically, when the seasonal variation amplitude value is within the first seasonal variation threshold range, the data sampling frequency is reduced to the preset reduction rate of the original frequency; when the seasonal variation amplitude value is within the second seasonal variation threshold range, the current data sampling frequency is maintained.

[0013] The first seasonal change threshold range is [15%, 25%], the second seasonal change threshold range is [5%, 15%], and the preset reduction rate is 60%.

[0014] Optionally, after the vegetation change trend prediction results are output, the system also includes establishing a long-term monitoring and assessment system for salt marsh vegetation based on the vegetation change trend prediction results and multi-scale ecological characteristic data. The system uses an ecological risk assessment function to classify the vegetation change trend prediction results into risk levels. When the risk level is high, the system activates an emergency monitoring mode to increase the monitoring frequency. When the risk level is medium, the system maintains the current monitoring frequency. When the risk level is low, the system reduces the monitoring frequency.

[0015] This invention effectively improves the accuracy of vegetation change information acquisition and time-series analysis by establishing a multi-level monitoring grid and employing wavelet transform multi-time-scale decomposition technology, solving the information gap problem caused by traditional single monitoring modes. This invention utilizes a graph-based ecological network connectivity analysis algorithm and fuzzy C-means clustering technology to accurately quantify habitat fragmentation and the gradual boundary characteristics of vegetation communities, overcoming the limitations of existing hard boundary identification methods in complex ecological structure analysis and significantly improving the accuracy of community delineation. This invention achieves accurate information fusion from pixel-scale to landscape-scale and reliable prediction of long-term change trends by constructing a multi-scale matrix mapping processing system and embedding a neural network prediction model with particle swarm optimization algorithm. Furthermore, it establishes an adaptive sampling frequency adjustment mechanism to optimize monitoring strategies based on the dynamic characteristics of vegetation change. In summary, this invention solves the technical problem mentioned in the background art where insufficient monitoring accuracy of long-term changes in salt marsh vegetation leads to inaccurate monitoring results. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a diagram showing the connectivity analysis results of a salt marsh vegetation network in a certain sea area in Example 2.

[0018] Figure 3 This is a distribution map of the ecological risk assessment levels for each vegetation type in Example 2. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0020] like Figure 1 The diagram shown is a flowchart of a method for monitoring long-term changes in salt marsh vegetation provided by this invention. This method includes the following steps:

[0021] S01. Establish a multi-level monitoring grid within the salt marsh wetland restoration area. Use a hierarchical spatial sampling method to set up main monitoring points and auxiliary monitoring points. The spacing between main monitoring points is 500m, and the spacing between auxiliary monitoring points is 100m. Simultaneously, configure soil temperature sensors, soil moisture sensors, and light intensity sensors to acquire soil temperature data, soil moisture data, and light intensity data.

[0022] S02. Use multispectral remote sensing technology to acquire vegetation spectral reflectance data, combine it with ground-measured chlorophyll content data, plant water content data and plant stem diameter data, establish a vegetation spectral characteristic database, record continuous observation data with a time series of 35 years, and output vegetation spectral reflectance data.

[0023] S03. Wavelet transform technology is used to decompose vegetation spectral reflectance data into multiple time scales to separate seasonal changes and long-term trends. The seasonal change amplitude calculation matrix equation is used to process the vegetation spectral reflectance data to obtain the seasonal change amplitude value. The sampling frequency adjustment function is used to calculate the adjusted sampling frequency based on the seasonal change amplitude value. When the seasonal change amplitude value is within the first seasonal change threshold range, the data sampling frequency is reduced to the preset reduction rate of the original frequency. When the seasonal change amplitude value is within the second seasonal change threshold range, the current data sampling frequency is maintained.

[0024] S04. Construct an ecological network connectivity analysis algorithm based on graph theory, taking salt marsh wetland vegetation patches as network nodes and ecological corridors as connecting edges, using vegetation spectral reflectance data and vegetation spectral feature database to calculate network clustering coefficients and shortest path lengths, assess the degree of habitat fragmentation, and output network clustering coefficient data and shortest path length data.

[0025] S05. The fuzzy C-means clustering algorithm is used to divide the vegetation community. A multi-dimensional feature vector is established based on environmental factors such as soil salinity data, soil pH data, relative humidity data and precipitation data. The fuzziness of community boundaries is handled by the membership function, and the vegetation community division results are output.

[0026] S06. When a pattern of change in the coverage of Suaeda salsa is detected and the length of the change is within the range of 3 to 5 years, the autoregressive moving average prediction algorithm is used to process the vegetation spectral reflectance data; when an expansion pattern of Spartina alterniflora in salt marshes is detected, the support vector machine prediction algorithm is used to process the vegetation community division results; when a vegetation degradation pattern is detected, the neural network prediction algorithm is used to process the network clustering coefficient data and the shortest path length data.

[0027] S07. A matrix rank deficiency detection algorithm is adopted. When the ecological feature matrix is ​​found to have a rank deficiency, the environmental constraint vector is automatically added to restore the full rank characteristic of the matrix. A hierarchical processing system for multi-scale matrix mapping is established. The hierarchical mapping from pixel scale to landscape scale realizes multi-level extraction of information and outputs multi-scale ecological feature data.

[0028] S08. Input the network clustering coefficient data, shortest path length data, and vegetation community division results into the wetland ecological evolution prediction model to predict long-term change trends. Use the hierarchical fusion weight adjustment function to calculate the hierarchical fusion weight parameters based on soil salinity concentration data, soil moisture content data, and light intensity data. Adjust the hierarchical fusion weight of the wetland ecological evolution prediction model through the hierarchical fusion weight parameters. Embed the particle swarm optimization algorithm into the forward propagation process of the wetland ecological evolution prediction model so that the parameter update of each layer depends not only on gradient information but also on the global search results of the particle swarm optimization algorithm. Output the vegetation change trend prediction results.

[0029] S09. Optionally, it also includes establishing a long-term monitoring and assessment system for salt marsh vegetation based on vegetation change trend prediction results and multi-scale ecological characteristic data, using an ecological risk assessment function to classify the risk level of vegetation change trend prediction results, and when the risk level is high, activating an emergency monitoring mode to increase the monitoring frequency to 150% of the original frequency, maintaining the current monitoring frequency when the risk level is medium, and reducing the monitoring frequency to 75% of the original frequency when the risk level is low, and outputting a long-term monitoring and assessment report for salt marsh vegetation.

[0030] Hierarchical spatial sampling is a method that sets sampling point densities according to different importance levels based on the spatial heterogeneity of the ecosystem. Primary monitoring points are used to acquire representative regional data, while auxiliary monitoring points are used to capture details of local changes. Multispectral remote sensing technology extracts physiological and biochemical parameters of vegetation by simultaneously acquiring reflectance information from multiple bands, including visible light, near-infrared, and shortwave infrared. A vegetation spectral feature database is a structured dataset storing vegetation spectral reflectance data and its corresponding ecological parameters. Wavelet transform is a time-frequency analysis method that decomposes signals into different frequency components, thereby separating seasonal fluctuations and long-term trends in vegetation data. The seasonal variation amplitude calculation matrix equation quantifies the degree of variation in vegetation spectral reflectance between different seasons. Inputs include vegetation spectral reflectance data, time series identifiers, and band weight coefficients; output is the seasonal variation amplitude value. The sampling frequency adjustment function dynamically adjusts the data acquisition frequency based on the seasonal variation amplitude value to optimize monitoring efficiency. Inputs include the seasonal variation amplitude value, the baseline sampling frequency, and the adjustment coefficient; output is the adjusted sampling frequency. The first seasonal change threshold interval is determined based on the variance analysis results of vegetation spectral reflectance data using a threshold calculation equation. The threshold calculation equation uses the ratio of the inter-seasonal variance to the inter-annual variance of the vegetation spectral reflectance data as a basic parameter, and combines this with vegetation type correction coefficients and environmental impact coefficients to calculate the upper and lower bounds of the first seasonal change threshold interval. The second seasonal change threshold interval is determined based on the stability analysis results of vegetation spectral reflectance data using a threshold calculation equation. The preset reduction rate is determined based on the monitoring cost-benefit analysis results using a reduction rate calculation equation. The reduction rate calculation equation uses the ratio of monitoring cost savings to data quality loss as the optimization objective function, and obtains the preset reduction rate by solving an optimization problem. Further, the first seasonal change threshold interval is [15%, 25%], the second seasonal change threshold interval is [5%, 15%], and the preset reduction rate is 60%.

[0031] Graph theory-based ecological network connectivity analysis algorithms treat ecosystems as networks composed of nodes and edges, quantifying ecological connectivity by calculating network topology indicators. Fuzzy C-means clustering allows data points to partially belong to multiple clusters, describing the gradual characteristics of vegetation communities through membership functions. Autoregressive moving average prediction algorithms combine the autocorrelation and random perturbations of historical data for time series prediction. Support vector machine prediction algorithms achieve nonlinear prediction by constructing optimal separating hyperplanes. Neural network prediction algorithms simulate the connection patterns of biological neurons for complex pattern recognition and prediction. Matrix rank deficiency detection algorithms are used to identify linearly correlated rows or columns in a data matrix; when the matrix is ​​not full-rank, it affects the stability of data analysis. Multi-scale matrix mapping is a technique for achieving unified analysis of data at different spatial scales through mathematical transformations. Multi-scale ecological feature data is a comprehensive dataset containing information from pixel scale to landscape scale, obtained through multi-scale matrix mapping processing.

[0032] The specific structure of the wetland ecological evolution prediction model is a multi-layer neural network based on a hierarchical attention network architecture, which includes an input layer, multiple hidden layers and an output layer. The hidden layers use a hierarchical attention mechanism to process multi-scale ecological features. Each hidden layer contains 256 neurons, the activation function is a modified linear unit function, and the hierarchical fusion weights are used to control the importance of information at different abstract levels.

[0033] The hierarchical fusion weight adjustment function is used to dynamically adjust the hierarchical fusion weight parameters of the wetland ecological evolution prediction model to optimize prediction performance. The inputs include soil salinity concentration data, soil moisture content data, light intensity data, and normalization coefficients. The output is the hierarchical fusion weight parameters. The calculation process involves weighted summation of the three input data and then mapping them to the [0, 1] interval using a sigmoid activation function to obtain the hierarchical fusion weight parameters. When the hierarchical fusion weight parameter w∈[0, 0.3), a lower hierarchical fusion weight of 0.2 is used to adjust the hierarchical fusion weight of the wetland ecological evolution prediction model; when the hierarchical fusion weight parameter w∈[0.3, 0.7], a medium hierarchical fusion weight of 0.5 is used to adjust the hierarchical fusion weight of the wetland ecological evolution prediction model; and when the hierarchical fusion weight parameter w∈(0.7, 1], a higher hierarchical fusion weight of 0.8 is used to adjust the hierarchical fusion weight of the wetland ecological evolution prediction model.

[0034] Particle Swarm Optimization (PSO) is a global optimization algorithm based on swarm intelligence. It finds the optimal solution by simulating bird flock foraging behavior. Each particle in the algorithm represents a potential solution, and the algorithm searches for the optimal parameter combination in the solution space using velocity and position update formulas. The vegetation change trend prediction result is a quantitative prediction of the future direction and extent of vegetation change in salt marshes, output by a wetland ecological evolution prediction model. The ecological risk assessment function is used to assess the ecosystem risk level based on the vegetation change trend prediction result. Inputs include the vegetation change trend prediction result, ecological threshold parameters, and risk weight coefficients; the output is the risk level. The long-term monitoring and assessment report of salt marsh vegetation is a comprehensive assessment document based on all monitoring data and prediction results, including vegetation change trends, ecological risk assessment, and management recommendations.

[0035] The specific implementation methods of the above steps are described in detail below.

[0036] The specific implementation of step S01 involves first establishing a multi-level monitoring grid system within the restoration area using a hierarchical spatial sampling method, based on the spatial heterogeneity and ecosystem complexity of the salt marsh wetland. The aim of this step is to ensure the representativeness and comprehensiveness of the monitoring data, simultaneously capturing both overall regional changes and local detailed changes through the configuration of sampling points at different densities. Specifically, the boundaries and internal heterogeneous distribution characteristics of the study area are first determined using geographic information systems and remote sensing image analysis. Then, main monitoring points are set at 500m intervals, primarily distributed in typical representative areas of the vegetation community, to obtain regional-scale ecosystem status information. Auxiliary monitoring points are set around the main monitoring points at 100m intervals, focusing on ecological transition zones and areas with significant environmental gradient changes, to capture fine-scale environmental change information. Soil temperature sensors, soil moisture sensors, and light intensity sensors are installed at each monitoring point, with installation depths of 10cm underground, 15cm underground, and at the surface, respectively, achieving measurement accuracies of ±0.1℃, ±2%, and ±5%, respectively. The data collection frequency is set to once per hour, and the data is transmitted in real time to the central database for storage and preliminary processing via a wireless transmission system.

[0037] The specific implementation of step S02 involves using multispectral remote sensing technology to acquire spectral reflectance data of vegetation within the study area. The purpose of this step is to establish a quantitative relationship between vegetation physiological and biochemical parameters and spectral characteristics. The process begins by selecting a suitable multispectral sensor, including those with spectral information acquisition capabilities in the visible light band (400–700 nm), the near-infrared band (700–1300 nm), and the short-wave infrared band (1300–2500 nm). Combined with ground-measured key physiological indicators such as chlorophyll content, plant water content, and plant stem diameter, spectral reflectance data at the canopy and leaf scales are obtained through on-site measurements using a spectrometer. The process of establishing a vegetation spectral characteristic database includes data quality control, spectral preprocessing, feature parameter extraction, and database structure design. Spectral preprocessing employs baseline correction, smoothing filtering, and normalization to remove the effects of atmospheric scattering and instrument noise. Feature parameter extraction focuses on calculating key vegetation indices such as the normalized difference vegetation index, enhanced vegetation index, and chlorophyll absorption reflectance index. The database is constructed using a relational data structure, recording 35 years of continuous observation data, including time stamps, spatial locations, spectral data, and corresponding ground-measured parameters, with a data volume of approximately 2TB.

[0038] The specific implementation of step S03 involves using wavelet transform technology to decompose the vegetation spectral reflectance data across multiple time scales. The purpose of this step is to separate the seasonal variations and long-term trends in the vegetation spectral data. Wavelet transform, as a time-frequency analysis method, can simultaneously provide information in both the time and frequency domains. By selecting an appropriate mother wavelet function, such as the Morlet wavelet or Daubechies wavelet, the spectral time series is decomposed. The decomposition process generates wavelet coefficients at different scales; low-frequency coefficients reflect long-term trends, while high-frequency coefficients reflect seasonality and short-term fluctuations. The calculation of the seasonal variation amplitude is achieved by constructing a matrix equation. The input parameters include the decomposed high-frequency wavelet coefficients, the time series identifier, and the weighting coefficients of each waveband. The output is a quantified value of the seasonal variation amplitude. The sampling frequency adjustment mechanism dynamically adjusts the data collection frequency based on the calculated seasonal variation amplitude value. When the seasonal variation amplitude value is in the range of 15% to 25%, it indicates that the vegetation is in a period of rapid change. At this time, the data sampling frequency is reduced to 60% of the original frequency to save monitoring costs. When the seasonal variation amplitude value is in the range of 5% to 15%, it indicates that the vegetation change is relatively stable. The current data sampling frequency is maintained to ensure monitoring continuity.

[0039] The specific implementation of step S04 involves constructing a graph theory-based ecological network connectivity analysis algorithm. The purpose of this step is to quantify the ecological connectivity between vegetation patches and the degree of habitat fragmentation within the salt marsh wetland. Graph theory analysis abstracts the salt marsh wetland ecosystem into a network structure composed of nodes and edges, where vegetation patches serve as network nodes and ecological corridors as connecting edges. Node attributes include ecological parameters such as patch area, vegetation cover, and biomass. Edge weights are determined through a comprehensive evaluation of vegetation spectral similarity and spatial distance. The calculation of the network clustering coefficient reflects the tightness of connections within the node's neighborhood. The calculation process involves the number of neighboring nodes for each node and the actual number of connections between neighboring nodes, quantifying the clustering characteristics of the local network through a ratio. The shortest path length is calculated using either the Dijkstra algorithm or the Floyd algorithm to determine the shortest connected path between any two nodes, reflecting the efficiency of information transmission and material flow within the ecosystem. The threshold settings for connectivity indicators are based on relevant ecological studies: a clustering coefficient greater than 0.6 indicates good connectivity, less than 0.3 indicates poor connectivity, and a shortest path length exceeding 10 node distances indicates impaired connectivity.

[0040] The specific implementation of step S05 involves using a fuzzy C-means clustering algorithm to divide the vegetation community. The purpose of this step is to identify and classify different types of vegetation communities, resolving the ambiguity of community boundaries. The fuzzy clustering algorithm allows each sample point to belong to multiple cluster centers with different membership degrees, better reflecting the gradual changes in vegetation communities within a natural ecosystem. Input parameters include multi-dimensional feature vectors constructed based on environmental factors such as soil salinity, soil pH, relative humidity, and precipitation. Each feature vector contains 15–20 environmental variables. The clustering process first randomly initializes cluster centers, then iteratively optimizes the objective function to achieve convergence of the cluster centers. The membership function uses an exponential decay form to describe the degree to which a sample point belongs to different cluster centers. The fuzziness parameter is set between 1.5 and 2.5, controlling the degree of fuzziness in the clustering results; a larger value indicates stronger fuzziness. The convergence criterion is set as a change in the objective function of less than 0.001 or a change in membership degree of less than 0.01 during continuous iterations, with a maximum iteration count limited to 200 times. The number of clusters is determined by the silhouette coefficient and clustering effectiveness index, and is usually set to 5 to 8 to reflect the main types of vegetation communities in salt marsh wetlands.

[0041] The specific implementation of step S06 involves selecting appropriate prediction algorithms based on the detected different vegetation change patterns. The purpose of this step is to improve prediction accuracy and the applicability of the algorithms. The pattern recognition process first detects the changing characteristics of vegetation cover through time series analysis, including key parameters such as the direction, magnitude, and duration of change. When the coverage of the Suaeda salsa salt marsh exhibits periodic fluctuations with a cycle length within the range of 3–5 years, an autoregressive moving average model is used to process the vegetation spectral reflectance data. This model achieves time series prediction by combining the autocorrelation and random disturbance terms of historical data. The model parameters include an autoregressive order p set to 2–4 and a moving average order q set to 1–3. Parameter optimization is achieved using the Akaike information criterion. When the Spartina alterniflora salt marsh exhibits an expansion pattern, i.e., a continuous increase in coverage area with a growth rate exceeding 5%, a support vector machine algorithm is used to process the vegetation community division results. Nonlinear prediction is achieved by constructing an optimal separating hyperplane. The kernel function is a radial basis function, and the regularization parameter C is set within the range of 1–100. The optimal parameter combination is determined through cross-validation. When a vegetation degradation pattern is detected, i.e., a decrease in vegetation cover of more than 10% for more than 2 years, a neural network algorithm is used to process the network clustering coefficient and shortest path length data. The network structure includes an input layer, 2 to 3 hidden layers and an output layer. The number of neurons in the hidden layer is set to 64 to 128, and the activation function is the modified linear unit function.

[0042] The specific implementation of step S07 involves using a matrix rank deficiency detection algorithm to identify linear correlation problems in the data matrix. This step aims to ensure the stability and reliability of the data analysis. Matrix rank deficiency detection is achieved through singular value decomposition. When the rank of the ecological feature matrix is ​​detected to be less than the minimum dimension of the matrix, it indicates the existence of linearly correlated rows or columns. In this case, environmental constraint vectors are automatically added to restore the full-rank characteristic of the matrix. The construction of environmental constraint vectors is based on the physicochemical constraints of the ecosystem, including ecological principles such as energy balance, material cycling, and interspecific competition. Multi-scale matrix mapping establishes a hierarchical processing system from a pixel-scale resolution of 30m to a landscape-scale resolution of 1km. A scale transformation function is used to achieve unified analysis of data at different spatial scales. Pixel-scale data reflects the fine structural information of vegetation, community-scale data reflects the composition and structural characteristics of vegetation communities, and landscape-scale data reflects the spatial pattern and connectivity characteristics of the ecosystem. The information extraction process uses a combination of wavelet transform and principal component analysis. Through multi-level information fusion, it achieves the transformation from low-level visual features to high-level semantic features, outputting a comprehensive ecological feature dataset containing information at multiple scales.

[0043] The specific implementation of step S08 involves inputting network clustering coefficient data, shortest path length data, and vegetation community division results into the wetland ecological evolution prediction model for long-term trend prediction. The purpose of this step is to achieve quantitative prediction of future changes in salt marsh vegetation. The hierarchical fusion weight adjustment function calculates weight parameters based on soil salinity concentration, soil moisture content, and light intensity data. The calculation process includes three steps: data standardization, weighted summation, and sigmoid function mapping. The output weight parameters are used to adjust the importance of information at different abstract levels in the prediction model. When the weight parameter is in the range of 0–0.3, a lower-level fusion weight of 0.2 is used, focusing on bottom-level feature information. When the weight parameter is in the range of 0.3–0.7, a medium-level fusion weight of 0.5 is used to balance bottom-level and high-level feature information. When the weight parameter is in the range of 0.7–1, a higher-level fusion weight of 0.8 is used, focusing on high-level semantic information. In the forward propagation process of the Particle Swarm Optimization (PSO) algorithm embedded in the model, each particle represents a set of neural network parameters. The optimal solution is searched in the parameter space through velocity and position update formulas. The particle swarm size is set to 50-100 particles, the maximum number of iterations is set to 200-500, the inertia weight is linearly decayed from 0.9 to 0.4, and the acceleration coefficient is set to 2.0. The global search capability is used to improve the optimization effect of model parameters.

[0044] Step S09 is optional. Its specific implementation involves establishing a long-term monitoring and assessment system for salt marsh vegetation based on vegetation change trend prediction results and multi-scale ecological characteristic data. The purpose of this step is to achieve dynamic assessment and early warning of ecosystem health. The input parameters of the ecological risk assessment function include vegetation change trend prediction results, ecological threshold parameters, and risk weight coefficients. The risk level of the ecosystem is calculated through a comprehensive assessment function. A three-level standard is adopted for risk classification: when the comprehensive risk index is greater than 0.7, it is considered high risk, and an emergency monitoring mode is activated, increasing the monitoring frequency to 150% of the original frequency, strengthening the observation of key areas and sensitive indicators; when the risk index is between 0.3 and 0.7, it is considered medium risk, maintaining the current monitoring frequency to ensure continuous observation; when the risk index is less than 0.3, it is considered low risk, reducing the monitoring frequency to 75% of the original frequency to optimize resource allocation. The assessment report is generated using an automated template-filling method, including modules such as vegetation change trend analysis, ecological risk assessment results, key indicator change charts, and management recommendations. Charts and maps are generated through data visualization technology to provide intuitive information support for decision-makers.

[0045] It should be noted that the wetland ecological evolution prediction model adopts a deep neural network structure based on a hierarchical attention network architecture. The overall framework includes an input layer, multiple hidden layers, and an output layer in a hierarchical organization. The input layer receives multi-dimensional ecological feature data, including 127 variables such as vegetation spectral reflectance, soil physicochemical properties, meteorological elements, and ecological connectivity indicators. Feature normalization and dimensionality reduction convert the input data into standardized feature vectors. The hidden layers employ a hierarchical attention mechanism, containing five hidden layers, each with 256 neurons. An attention weight mechanism adaptively focuses on different scales and types of ecological feature information. The first hidden layer focuses on local features at the pixel scale, the second on structural features at the community scale, the third on spatial pattern features at the landscape scale, the fourth on multi-scale feature fusion, and the fifth on high-level semantic feature extraction. The activation function of each hidden layer uses a modified linear unit function, maintaining the effectiveness of gradient propagation through linear rectification while avoiding the gradient vanishing problem. The hierarchical fusion weights are implemented using learnable parameters, adaptively adjusting the importance of information at different abstract levels during training. Weight updates employ a hybrid optimization strategy combining backpropagation and particle swarm optimization. The output layer contains three output nodes, corresponding to vegetation cover change trends, community structure succession directions, and ecosystem stability assessments, respectively. The output results are predicted in probability distribution form using a softmax activation function.

[0046] The establishment of the training dataset for the wetland ecological evolution prediction model includes key steps such as data collection, quality control, preprocessing, and data segmentation. During the data collection phase, 18 representative saline wetland restoration areas in the Bohai Bay were selected as data sources. These areas cover different restoration periods, methods, and environmental conditions, ensuring data diversity and representativeness. Historical monitoring data spans from 1986 to 2021, totaling 35 years of long-term observation records. Data types include 127 variables across four main categories: vegetation cover change, soil physicochemical properties, meteorological elements, and ecological connectivity indicators. Vegetation cover change data was obtained through a combination of remote sensing image interpretation and ground surveys, including indicators such as coverage area, biomass, and health status of different vegetation types. Soil physicochemical property data, including key parameters such as soil salinity, pH, organic matter content, moisture content, and temperature, was obtained through regular sampling and laboratory analysis. Meteorological element data, including environmental factors such as air temperature, precipitation, relative humidity, wind speed, and light intensity, was obtained through meteorological station observations and satellite remote sensing data. Ecological connectivity index data are calculated using landscape ecology analysis methods, including parameters such as landscape connectivity index, habitat fragmentation degree, and ecological corridor quality.

[0047] The data quality control process employs a multi-level verification mechanism. First, outliers are identified and removed using statistical analysis methods, with three times the standard deviation used as the criterion for outlier judgment. Data points exceeding the normal range undergo manual review and verification. Missing value handling utilizes multiple imputation methods. Variables with a missing value rate below 10% are imputed using linear or regression interpolation, while samples with a missing value rate above 20% are directly removed to ensure data quality. Data standardization employs Z-score standardization, converting variables with different dimensions and numerical ranges into standardized data with a mean of 0 and a standard deviation of 1, eliminating the impact of dimensional differences on model training. The feature engineering process includes feature selection, feature transformation, and feature combination. Key features with high contribution to the prediction target are screened using methods such as correlation analysis, principal component analysis, and information gain analysis, ultimately identifying 87 core feature variables. The dataset was split using time-series cross-validation, with data from 1986 to 2016 used as the training set (25,000 samples), data from 2017 to 2019 used as the validation set (8,000 samples), and data from 2020 to 2021 used as the test set (5,000 samples) to ensure the model's time generalization ability.

[0048] The model training process employs a phased training strategy. The first phase involves pre-training, using a large amount of unsupervised data to learn latent representations of ecological features. The second phase involves fine-tuning, using labeled data to optimize the model's predictive performance. The optimizer chosen is the Adam optimizer with an adaptive learning rate. The initial learning rate is set to 0.001, and a learning rate decay strategy is used to dynamically adjust the learning rate during training, with a decay rate set to 0.95, decreasing every 50 epochs. The batch size is set to 64 to balance computational resources and training efficiency. The total number of training epochs is set to 500. An early stopping strategy is used to prevent overfitting; training automatically stops when the validation set loss does not decrease for 10 consecutive epochs. Model performance is evaluated using a multi-metric comprehensive evaluation method, including mean absolute error, root mean square error, correlation coefficient, and prediction accuracy. The final model achieves a prediction accuracy of 92.3% on the test set, with the root mean square error controlled within 0.08 and a correlation coefficient of 0.94, indicating that the model has good predictive ability and generalization performance.

[0049] It should be noted that the first key technical concept of this invention is the integrated application of a multi-scale hierarchical monitoring grid and an adaptive sampling frequency adjustment mechanism. Traditional ecological monitoring methods typically use a regular grid with fixed spacing to deploy monitoring points, maintaining a constant monitoring frequency. This makes it difficult to balance regional representativeness and local accuracy requirements, while also resulting in resource waste or information loss. This invention establishes a two-layer monitoring network of main and auxiliary monitoring points through a hierarchical spatial sampling method. Combined with wavelet transform technology, it performs time-frequency decomposition of vegetation spectral data, achieving accurate separation of seasonal changes and long-term trends. The sampling frequency is then dynamically adjusted based on the magnitude of seasonal changes. The core advantage of this technical approach lies in its ability to adaptively optimize monitoring resource allocation according to the actual characteristics of ecosystem changes. During periods of rapid vegetation change, the sampling frequency is reduced to save costs, while during stable periods, an appropriate frequency is maintained to ensure monitoring continuity, achieving a balance between monitoring efficiency and economy.

[0050] The second key technical approach is the synergistic application of graph theory-based ecological network connectivity analysis and fuzzy clustering algorithms. Traditional vegetation community segmentation methods are mostly based on hard clustering, strictly assigning each sample point to a single category, which is insufficient to accurately describe the gradual characteristics and transitional properties of vegetation communities in natural ecosystems. This invention constructs a graph theory-based ecological network model, abstracting vegetation patches as network nodes and ecological corridors as connecting edges. Ecological connectivity is quantified using topological indicators such as network clustering coefficients and shortest path lengths. Simultaneously, a fuzzy C-means clustering algorithm is used to handle vegetation community segmentation, allowing sample points to belong to multiple community types with different membership degrees. The core advantage of this technical approach lies in its ability to more realistically reflect the complexity and continuity of ecosystems. Network connectivity analysis reveals the structural characteristics and functional relationships of ecosystems, while fuzzy clustering addresses the uncertainty of community boundaries, providing a more precise scientific basis for ecosystem management.

[0051] The third key technical approach is a predictive modeling strategy that employs adaptive selection of multiple algorithms and hybrid optimization. Traditional ecological prediction methods typically use a single prediction algorithm, which struggles to adapt to the complexity and diversity of different vegetation change patterns, limiting prediction accuracy and generalization ability. This invention adaptively selects the most suitable prediction algorithm based on the different characteristics of vegetation change patterns: an autoregressive moving average model for periodic changes, a support vector machine algorithm for expansive changes, and a neural network algorithm for degenerative changes. Simultaneously, a particle swarm optimization algorithm is embedded in the parameter optimization process of the deep neural network, achieving an organic combination of gradient information and global search. The core advantage of this approach lies in its ability to fully leverage the strengths of different algorithms in different scenarios, improving prediction accuracy through algorithmic complementarity, and avoiding local optima through a hybrid optimization strategy, thus achieving intelligent and adaptive prediction models.

[0052] The fourth key technical approach is a data processing framework based on matrix rank deficiency detection and multi-scale information fusion. Traditional multi-source data fusion methods often neglect the linear correlation problem of data matrices, leading to instability and unreliability of analysis results, and lack an effective multi-scale information integration mechanism. This invention identifies linear correlation problems in data through a matrix rank deficiency detection algorithm, automatically supplements environmental constraint vectors to restore the full-rank characteristic of the matrix, and establishes a multi-level information extraction system from pixel scale to landscape scale. Through hierarchical processing, it achieves unified analysis and comprehensive utilization of information at different spatial scales. The core advantage of this technical approach lies in ensuring the mathematical stability and ecological rationality of data analysis, achieving comprehensive understanding from micro to macro through multi-scale information fusion, and providing a reliable data foundation for the comprehensive analysis of complex ecosystems.

[0053] The synergistic effect of these four key technological approaches demonstrates significant technical effectiveness and advantages over existing technologies. First, the multi-level monitoring grid and adaptive sampling mechanism provide high-quality, multi-scale foundational data for subsequent data analysis, ensuring data representativeness and cost-effectiveness. Second, the combination of graph theory network analysis and fuzzy clustering offers a new technical approach for the quantitative description of ecosystem structure and function, improving the accuracy and completeness of ecological feature identification. Third, the multi-algorithm adaptive selection and hybrid optimization strategy fully leverage the complementary advantages of different algorithms, significantly enhancing the accuracy and generalization ability of the prediction model. Finally, matrix rank deficiency detection and multi-scale information fusion ensure the mathematical rigor and ecological rationality of data processing, laying a solid foundation for the reliability of the entire monitoring and assessment system. The synergistic effect of these technological approaches constitutes a complete ecological monitoring and assessment technology system, achieving end-to-end technological innovation from data acquisition, feature extraction, pattern recognition to trend prediction. Compared to traditional single-technology methods, it possesses comprehensive technical advantages such as higher accuracy, better efficiency, stronger adaptability, and better reliability.

[0054] It should be noted that in long-term monitoring of salt marsh vegetation, vegetation spectral data simultaneously contains seasonal periodic variations and interannual long-term trend variations. Traditional analysis methods cannot effectively separate these two different time-scale variation patterns, leading to interference from seasonal fluctuations when identifying long-term evolutionary trends and affecting the accuracy of trend judgment. This invention employs wavelet transform technology to decompose vegetation spectral reflectance data across multiple time scales. Through time-frequency domain analysis, the composite signal is decomposed into different frequency components, achieving precise separation of seasonal variations and long-term trends. This technology utilizes the localization characteristics of wavelet basis functions to simultaneously provide time-domain and frequency-domain information, accurately identifying variation patterns at different time scales and providing a clear data foundation for subsequent trend analysis and prediction. Traditional habitat fragmentation assessments mainly rely on area statistics and simple spatial distance calculations, failing to accurately reflect the functional connectivity between ecological patches and the transmission efficiency of ecological flows, making it difficult to provide a scientific basis for ecological restoration and protection strategy formulation. This invention constructs a graph-based algorithm for ecological network connectivity analysis. It treats saline marsh wetland vegetation patches as network nodes and ecological corridors as connecting edges. Local connectivity density is quantified by calculating network clustering coefficients, and global connectivity efficiency is assessed using shortest path lengths. This method transforms complex ecosystem structures into quantifiable network topology parameters, accurately identifying key nodes and weak links in the ecological network. It provides a reliable technical means for the quantitative assessment of habitat connectivity, significantly improving the scientific rigor and accuracy of ecosystem structure analysis.

[0055] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of inaccurate monitoring results due to insufficient accuracy in long-term monitoring of vegetation changes in salt marshes, mainly based on the synergistic mechanism of multi-level accuracy improvement technologies. At the data acquisition level, the hierarchical spatial sampling method, by setting main and auxiliary monitoring points of different densities, achieves simultaneous capture of regional representative features and local change details, laying a solid data foundation for subsequent high-precision analysis. The powerful decomposition capability of wavelet transform technology in the time-frequency domain enables precise separation of multi-timescale information in vegetation spectral reflectance data, effectively distinguishing seasonal fluctuations and long-term evolution trends, avoiding the information confusion problem of traditional single-timescale analysis methods. At the spatial analysis level, the graph theory-based ecological network connectivity analysis algorithm transforms the complex ecosystem structure into precisely quantifiable network topology parameters, achieving accurate assessment of habitat connectivity through indicators such as network clustering coefficients and shortest path lengths. The fuzzy C-means clustering algorithm effectively handles the continuous change characteristics of vegetation community boundaries through a membership function mechanism, avoiding misjudgments caused by hard boundary divisions and improving the accuracy of community identification. At the predictive analysis level, multi-scale matrix mapping technology establishes a hierarchical information transformation mechanism from micro to macro, ensuring the complete preservation and accurate fusion of features at different spatial scales. The neural network prediction model embedded with particle swarm optimization significantly enhances the accuracy of parameter optimization and the reliability of prediction results through a combination of global search and local gradient optimization. Meanwhile, the adaptive sampling frequency adjustment mechanism optimizes the data acquisition strategy based on the real-time dynamic characteristics of vegetation change, ensuring the complete acquisition of key change information.

[0056] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0057] In this embodiment, the specific implementation methods of steps S01 to S02 are the same as those described above, and will not be repeated in detail here.

[0058] The specific implementation of step S03 involves using wavelet transform technology to decompose the vegetation spectral reflectance data into multiple time scales. The specific representation of the matrix equation for calculating the seasonal variation amplitude is as follows:

[0059] ;

[0060] In the formula, This represents the magnitude of seasonal variation. This represents the length of the time series, ranging from 12 to 48 months. This represents the number of spectral bands, ranging from 6 to 12 bands. For the first The weighting coefficients for each band are determined through principal component analysis. ; For the first The first time point Spectral reflectance data for each band; This represents the time-averaged spectral reflectance.

[0061] The parameter acquisition method is as follows: The study employed multispectral remote sensing technology, including the following steps: Step 1: acquiring spectral images of the study area using a multispectral sensor under clear, cloudless weather conditions; Step 2: performing atmospheric and geometric correction processing; and Step 3: extracting the spectral reflectance values ​​corresponding to each monitoring point. The weighting coefficients are obtained by calculation, through principal component analysis of historical spectral data, and by selecting the eigenvectors corresponding to principal components with a cumulative contribution rate of over 85%.

[0062] The sampling frequency adjustment function is expressed as follows:

[0063] ;

[0064] In the formula, To adjust the post-sampling frequency; The baseline sampling frequency is set to once per day; This is an adjustment coefficient, with a value ranging from 0.4 to 0.6; This represents the minimum seasonal variation. This represents the maximum seasonal variation.

[0065] The first seasonal variation threshold range is determined by a threshold calculation equation, as follows:

[0066] ;

[0067] ;

[0068] In the formula, and These are the lower and upper bounds of the first seasonal variation threshold interval, respectively, and are dimensionless. and These are the basic parameter coefficients, taking values ​​of 0.12 and 0.22 respectively, and are dimensionless; represents the inter-seasonal variance of vegetation spectral reflectance data, which is dimensionless; represents the interannual variance of vegetation spectral reflectance data, which is dimensionless; The value is a vegetation type correction factor, with 0.03 for Suaeda salsa community and 0.02 for Spartina alterniflora community, and is dimensionless. The environmental impact coefficient ranges from 0.01 to 0.05 and is dimensionless.

[0069] The parameter acquisition method is as follows: The data was obtained through calculation, specifically by calculating the variance of spectral reflectance data from different seasons within the same year. The variance of the spectral reflectance data from the same season over many years was calculated using a computational method. The study employed an experimental approach, including step 1: measuring environmental factors such as soil moisture content, salinity, and temperature in the study area; step 2: determining the degree of influence of environmental factors on vegetation spectral changes through correlation analysis; and step 3: calculating the environmental impact coefficient based on the degree of influence.

[0070] The second seasonal variation threshold range is determined by a threshold calculation equation, as follows:

[0071] ;

[0072] ;

[0073] In the formula, and These are the lower and upper bounds of the second seasonal variation threshold interval, respectively, and are dimensionless. and These are the basic parameter coefficients, taking values ​​of 0.03 and 0.13 respectively, and are dimensionless; The stability coefficient of vegetation spectral reflectance data is calculated using the coefficient of variation and is dimensionless.

[0074] in, The method of obtaining the data is as follows: Step 1: Calculate the standard deviation of the vegetation spectral reflectance data over time; Step 2: Calculate the corresponding mean over time; Step 3: Obtain the coefficient of variation as the stability coefficient by the ratio of the standard deviation to the mean.

[0075] The preset reduction rate is determined by the reduction rate calculation equation, as shown below:

[0076] ;

[0077] In the formula, The preset reduction rate is dimensionless. To monitor the degree of cost savings, the function represents the reduction rate. The relative cost savings at that time are dimensionless. Let be a function representing the degree of data quality loss, indicating the reduction rate. The relative data quality loss at that time, dimensionless; This represents the parameter that minimizes the objective function.

[0078] in, The calculation method is used, including step 1: calculating the equipment operating cost and labor cost at the original monitoring frequency; step 2: calculating the reduction rate. The corresponding cost at that time; Step 3: The degree of cost saving is obtained by comparing the cost difference with the original cost. The data is obtained through calculation, including step 1: analyzing the data integrity at different sampling frequencies using historical data; step 2: evaluating data quality indicators such as signal-to-noise ratio and temporal resolution loss; and step 3: comprehensively calculating the degree of quality loss.

[0079] The specific implementation of step S04 involves constructing a graph theory-based ecological network connectivity analysis algorithm. The calculation formula for the network clustering coefficient is as follows:

[0080] ;

[0081] In the formula, For nodes The clustering coefficient is dimensionless. For nodes The actual number of connections between neighboring nodes, dimensionless; For nodes The degree of the node, i.e., the degree of the node. The number of directly connected nodes, dimensionless.

[0082] The parameter acquisition method is as follows: Obtained through computation by traversing nodes. For all neighboring nodes, count the actual number of connections between neighboring nodes; Obtained through calculation, using statistics and nodes. The total number of directly connected nodes.

[0083] The shortest path length is calculated using Dijkstra's algorithm, and the update formula for the distance matrix is ​​as follows:

[0084] ;

[0085] In the formula, For the first After the next iteration, the node To the node The shortest distance, in km; This is the distance value from the previous iteration, in km; For nodes in the previous iteration To the intermediate node The distance, in km; intermediate nodes in the previous iteration To the node The distance, in km; This is an index for intermediate nodes, dimensionless; note the significance of this. This serves as the identifier for intermediate nodes during the iteration, and is related to the node degree in the aforementioned clustering coefficient formula. They have different meanings.

[0086] The parameter acquisition method is as follows: The initial values ​​were obtained through field measurements, including step 1: measuring the center coordinates of vegetation patches using a GPS device; step 2: calculating the Euclidean distance between patches as the initial distance value; and step 3: adjusting the distance weights according to the connectivity of the ecological corridor. and It is obtained through an iterative algorithm process.

[0087] The specific implementation of step S05 involves using the fuzzy C-means clustering algorithm to divide the vegetation community. The objective function is expressed as follows:

[0088] ;

[0089] In the formula, Let be the objective function for fuzzy clustering, with the dimension being the square of the distance; The number of clusters, ranging from 5 to 8, is dimensionless. The sample size is dimensionless. For the sample Cluster centers The membership degree, ranging from 0 to 1, is dimensionless; This is a fuzzification parameter, with a value range of 1.5 to 2.5, and is dimensionless. For the first The feature vector of each sample contains 15 environmental factors, including soil salinity, soil pH, relative humidity and precipitation, and each component has corresponding physical dimensions. For the first Cluster centers, dimensions and same; It represents Euclidean distance.

[0090] The parameter acquisition method is as follows: The data were obtained through experiments and observations, including: Step 1: collecting soil samples at each sampling point and measuring salt concentration and pH value; Step 2: using a meteorological station to observe relative humidity and precipitation data; Step 3: standardizing all environmental factors to eliminate the influence of dimensions. The optimal number of clusters is determined by using a clustering effectiveness index and by calculating the silhouette coefficient.

[0091] The membership update formula is expressed as follows:

[0092] ;

[0093] The formula for updating cluster centers is expressed as follows:

[0094] .

[0095] The specific implementation method of step S06 is the same as described above, and will not be repeated in detail here.

[0096] The specific implementation of step S07 is to use the matrix rank deficit detection algorithm, and the singular value decomposition is represented as follows:

[0097] ;

[0098] In the formula, This is an ecological feature matrix with dimensions of 1. Each element has a corresponding ecological parameter dimension; The row number of the ecological feature matrix represents the number of monitoring points, which is dimensionless. The column number of the ecological feature matrix represents the number of ecological feature variables and is dimensionless. It is a left singular vector matrix with dimension . Dimensionless; It is a singular value diagonal matrix with dimension . Dimensions and Related; It is a right singular vector matrix with dimension . , dimensionless.

[0099] The parameter acquisition method is as follows: The data was acquired through an integration method, including step 1: collecting vegetation spectral reflectance data, soil physicochemical property data, and ecological connectivity index data; step 2: organizing different types of data into a matrix according to their temporal and spatial correspondences; and step 3: standardizing the matrix elements to eliminate dimensional differences. The number of monitoring points is determined by the monitoring grid design and is equal to the total number of main monitoring points and auxiliary monitoring points. The number of spectral bands, soil physicochemical parameters, and ecological indicators are determined through feature selection.

[0100] The criterion for determining the rank of a matrix is:

[0101] ;

[0102] In the formula, For matrix The rank of is dimensionless; For the first A singular value, dimensionless Related; This is an indicator function that returns 1 if the condition is true, and 0 otherwise; it is dimensionless. This is the tolerance threshold, with a value of [value]. Its dimensions are the same as those of the singular values.

[0103] The parameter acquisition method is as follows: Calculated using the singular value decomposition algorithm; The accuracy is set according to the numerical calculation requirements, and is usually the square root of the machine's accuracy.

[0104] The transformation formula for multi-scale matrix mapping is expressed as follows:

[0105] ;

[0106] In the formula, This is a landscape-scale feature vector, the dimensions of which depend on specific ecological parameters; Here is the scaling transformation matrix, with dimension 1. The dimension is the ratio of the dimensions of the output to those of the input parameters; It is a pixel-scale feature vector, the dimensions of which depend on the specific ecological parameters; For the bias vector, the dimensions are the same as... same; and These are the feature dimensions at the landscape scale and pixel scale, respectively, and are dimensionless.

[0107] The parameter acquisition method is as follows: An optimization method was used to obtain the transformation relationship between different scales, and the least squares method was used to fit the transformation relationship between different scales. The method employs remote sensing data extraction, including step 1: acquiring high-resolution remote sensing images; step 2: extracting pixel-level spectral and texture features; and step 3: calculating vegetation indices and ecological parameters.

[0108] The specific implementation of step S08 is to use a hierarchical fusion weight adjustment function, which is expressed as follows:

[0109] ;

[0110] In the formula, The weight parameters for hierarchical fusion are dimensionless. It is a sigmoid activation function. Dimensionless; The normalization coefficients are determined through cross-validation, and the constraints are as follows: Dimensionless; The data represents standardized soil salinity concentrations and is dimensionless. The data represents standardized soil moisture content, which is dimensionless. The light intensity data is standardized and dimensionless. This is the bias term, with a value range of -0.5 to 0.5, and is dimensionless.

[0111] The parameter acquisition method is as follows: The data were obtained experimentally, including: Step 1: collecting surface soil samples at each monitoring point at a depth of 0–10 cm; Step 2: determining soil salinity using the conductivity method; and Step 3: processing the raw data using the Z-score normalization method. The soil moisture content is measured using a sensor, which monitors the soil moisture content in real time, and then the data is standardized. The solar radiation intensity is measured using a light intensity sensor and then standardized. An optimization method was used to obtain the optimal parameter combination through grid search and cross-validation.

[0112] The piecewise function for hierarchical fusion weight adjustment is expressed as follows:

[0113] ;

[0114] In the formula, The final hierarchical fusion weight is dimensionless.

[0115] The velocity and position update formulas for the particle swarm optimization algorithm are expressed as follows:

[0116] ;

[0117] ;

[0118] In the formula, and The first The particle in the first The speed and position of each iteration are measured in units that depend on the specific parameter type. The inertial weight decreases linearly from 0.9 to 0.4 and is dimensionless. and The acceleration factor is 2.0 and is dimensionless. and A dimensionless random number between 0 and 1; For the first The historical optimal position of a particle, its dimensions and same; For the group's historical optimal position, the dimensions and same; For the first The particle in the first The position of the next iteration represents the neural network parameter vector.

[0119] The parameter acquisition method is as follows: and The fitness function value is determined by dynamically updating the fitness function through algorithm iteration. and Generated by a random number generator; A linear decay strategy is adopted. ,in , , The current iteration number is dimensionless. This represents the maximum number of iterations, ranging from 200 to 500, and is dimensionless.

[0120] The specific implementation of step S09 is to establish a long-term monitoring and assessment system for salt marsh vegetation. The ecological risk assessment function is expressed as follows:

[0121] ;

[0122] In the formula, This is an ecological risk index, dimensionless. The number of predictive indicators, ranging from 3 to 5, is dimensionless; For the first The risk weight coefficients for each indicator were determined using the analytic hierarchy process, with the following constraints: Dimensionless; For the first The predicted value of each indicator has dimensions that depend on the specific indicator type. For the first Ecological threshold parameters for each indicator, with units of... same; For the Heaviside step function, when hour Otherwise, it is 0, dimensionless.

[0123] The parameter acquisition method is as follows: The predicted values ​​of the corresponding indicators are obtained by using model prediction methods and output through the wetland ecological evolution prediction model. The data was obtained through literature review and expert consultation, including step 1: collecting threshold data from relevant ecological studies; step 2: inviting ecological experts to conduct threshold assessment; and step 3: comprehensively determining the ecological thresholds for each indicator. The analytic hierarchy process (AHP) is used to obtain the following steps: Step 1: Constructing the indicator importance judgment matrix; Step 2: Calculating the eigenvectors and eigenvalues; Step 3: Performing consistency checks and determining the weight coefficients.

[0124] It should be explained that the principle of the seasonal variation amplitude calculation matrix equation is based on the analysis of variance theory in statistics. It quantifies the degree of seasonal variation in vegetation by calculating the standard deviation of weighted spectral reflectance over time. This equation introduces multi-band weighting coefficients, which can comprehensively consider the differences in sensitivity of different spectral bands to vegetation changes. Compared with traditional single-band analysis methods, it significantly improves the accuracy and reliability of seasonal variation detection, and provides a precise data foundation for subsequent adaptive sampling frequency adjustment.

[0125] The sampling frequency adjustment function employs the concept of linear interpolation, mapping the seasonal variation amplitude to a sampling frequency adjustment coefficient to achieve dynamic optimization of the data acquisition frequency. Its core advantage lies in its ability to adaptively adjust monitoring resource input based on the actual rate of vegetation change. During periods of rapid vegetation change, the sampling frequency can be appropriately reduced to save costs, while maintaining the necessary monitoring density during periods of slower change. Compared to traditional fixed-frequency monitoring methods, this achieves an optimal balance between monitoring efficiency and economy.

[0126] The threshold calculation equation is based on the statistical characteristics of ecosystem variability. It reflects the relative strength of seasonal and interannual variations through the variance ratio, and incorporates correction terms for vegetation type and environmental factors to achieve adaptive threshold determination. The first threshold interval calculation formula is as follows: Second threshold interval calculation formula Compared to empirical threshold setting methods, it is more scientific and adaptable, and can automatically adjust the judgment criteria according to the ecological characteristics of different research areas, which significantly improves the accuracy of change pattern recognition.

[0127] The preset reduction rate calculation equation adopts the objective function minimization method in optimization theory, through... The goal is to find the optimal balance between cost savings and data quality loss in monitoring. This optimization equation, compared to a fixed reduction rate method, can dynamically determine the optimal parameters based on the cost-effectiveness requirements of specific application scenarios, thus achieving scientific and refined management of monitoring strategies.

[0128] The principle behind the network clustering coefficient calculation formula originates from complex network theory. It assesses the local clustering characteristics of a network by quantifying the connection density within the neighborhood of a node. In ecological applications, this formula effectively reflects the degree of ecological connectivity between vegetation patches. Compared to traditional distance-based connectivity analysis methods, it more accurately characterizes the structural features and functional relationships of ecosystems, providing a more scientific quantitative indicator for habitat fragmentation assessment.

[0129] The shortest path length is calculated using the dynamic programming approach of Dijkstra's algorithm, which iteratively updates the distance matrix. This algorithm finds the shortest path between any two nodes in a network. Compared to the traditional Euclidean distance calculation method, it more accurately reflects the actual connectivity in the ecosystem, providing a scientific basis for ecological corridor planning and ecological network optimization.

[0130] The objective function of fuzzy C-means clustering is based on fuzzy set theory. By introducing the concept of membership degree, it allows sample points to partially belong to multiple clusters, thus better handling the fuzziness and gradual changes in community boundaries in natural ecosystems. Membership degree update formula. Cluster center update formula Compared to traditional hard clustering algorithms, it can more realistically reflect the continuous distribution characteristics of vegetation communities, significantly improve the accuracy and ecological rationality of community classification, and provide more accurate community information for ecosystem management.

[0131] The principle of the singular value decomposition (SVD) formula is based on linear algebra theory. By decomposing the complex ecological feature matrix into the product of three simple matrices, it achieves in-depth analysis of the data structure and dimensionality reduction. (Matrix rank determination criteria) The method of determining the linearity of a matrix by using the magnitude of singular values ​​has a core advantage in ecological data processing: it can automatically identify and handle linear correlation problems in the data. Compared with traditional data processing methods, it significantly improves the stability and reliability of the analysis results, laying a solid mathematical foundation for the fusion analysis of multi-source ecological data.

[0132] Multi-scale matrix mapping transformation formula This method employs the mathematical principle of linear transformation to map and integrate information across different spatial scales through transformation matrices and bias vectors. Compared to simple scale averaging or interpolation methods, this transformation method better preserves ecological relationships across different scales, providing a mathematical tool for multi-scale ecological analysis.

[0133] The principle of the hierarchical fusion weight adjustment function is based on the attention mechanism in neural networks, using the sigmoid activation function. By mapping a linear combination of multiple environmental factors to a range of 0 to 1, adaptive adjustment of the weight parameters is achieved. Piecewise weight adjustment function. By implementing piecewise linear mapping, the weights are discretized and adjusted. Compared with the traditional method with fixed weights, this function can automatically adjust the importance of information at different levels according to the dynamic changes in environmental conditions, which significantly improves the adaptability and prediction accuracy of the prediction model and provides an effective technical means for accurate prediction of complex ecosystems.

[0134] The principle behind the velocity and position update formulas of the Particle Swarm Optimization (PSO) algorithm simulates the collective intelligent behavior of bird flocks foraging, achieving a global search of the parameter space through the dual guidance of individual and collective experience. Velocity update formula. The position update formula comprises three components: inertia, individual cognition, and social learning. This algorithm enables particle movement in the solution space. Its core advantage in neural network parameter optimization is that it can effectively avoid local optima. Compared with traditional gradient descent methods, it has stronger global search capabilities and convergence stability, significantly improving the training effect and generalization performance of prediction models.

[0135] The principle of the ecological risk assessment function is based on risk assessment theory. It identifies risk states by calculating the relative deviation between predicted results and ecological thresholds and introducing a step function, thus achieving a quantitative assessment of ecosystem health. (Heaviside step function) Used to identify threshold states, this function is more objective and operable than traditional qualitative risk assessment methods, and can provide clear numerical basis for ecosystem management decisions, significantly improving the pertinence and effectiveness of ecological protection measures.

[0136] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team conducted long-term monitoring of vegetation changes in a salt marsh wetland in a certain sea area, mainly covering typical vegetation types such as cogongrass, reeds, Spartina alterniflora salt marshes, and Suaeda salsa salt marshes. Following the monitoring method proposed in this invention, the technical team established a complete monitoring system and obtained continuous observation data for 35 years, from 1986 to 2021.

[0137] In step S01, the technical team established a multi-level monitoring grid within a saline wetland restoration area in a certain sea area. Nine main monitoring points were set up using a hierarchical spatial sampling method, with an interval of approximately 500m between each main monitoring point. Simultaneously, four auxiliary monitoring points were deployed around each main monitoring point, with an interval of 100m between the auxiliary monitoring points, forming a spatial network of a total of 45 monitoring points. Each monitoring point was equipped with a soil temperature sensor, a soil moisture sensor, and a light intensity sensor. Environmental data obtained during the monitoring period showed that the annual average soil temperature was 15.2℃, the annual average soil moisture content was 68.5%, and the annual average light intensity was 420. .

[0138] In step S02, the technical team used satellite multispectral remote sensing technology to acquire vegetation spectral reflectance data. Combined with ground-based measured data, including chlorophyll content data, plant water content data, and plant stem diameter data, a vegetation spectral characteristic database was established. This database records 35 years of continuous observation data from 1986 to 2021, including reflectance information in the visible light band, near-infrared band, and short-wave infrared band. Monitoring results showed that the average near-infrared reflectance of thatch grass was 0.65, that of reeds was 0.58, that of Spartina alterniflora salt marsh was 0.72, and that of Suaeda salsa salt marsh was 0.61.

[0139] In step S03, the technical team used wavelet transform technology to decompose the vegetation spectral reflectance data into multiple time scales. The seasonal variation amplitude was calculated by processing the vegetation spectral reflectance data using a matrix equation. The seasonal variation amplitude values ​​were 18.7% for cogongrass, 22.3% for reeds, 8.9% for Spartina alterniflora salt marshland, and 27.2% for Suaeda salsa salt marshland. According to the sampling frequency adjustment function, when the seasonal variation amplitude value is within the first seasonal variation threshold range [15%, 25%], the data sampling frequency is reduced to 60% of the original frequency; when the seasonal variation amplitude value is within the second seasonal variation threshold range [5%, 15%], the current data sampling frequency is maintained.

[0140] In step S04, the technical team constructed an ecological network connectivity analysis algorithm based on graph theory. Using vegetation patches in a certain sea area as network nodes and ecological corridors as connecting edges, the network clustering coefficient and shortest path length were calculated. For example... Figure 2 As shown, the network clustering coefficient of the monitored area is 0.73, and the average shortest path length is 4.2 node units, indicating that the area has good ecological connectivity. The habitat fragmentation assessment results show that the fragmentation index is 0.31 in the cogongrass distribution area, 0.28 in the reed distribution area, 0.45 in the Spartina alterniflora salt marsh distribution area, and 0.52 in the Suaeda salsa salt marsh distribution area.

[0141] In step S05, the technical team used a fuzzy C-means clustering algorithm to classify the vegetation communities. A multidimensional feature vector was established based on soil salinity concentration data, soil pH data, relative humidity data, and precipitation data. As shown in Table 1, the environmental factor data of the monitoring area reflect the habitat characteristics of different vegetation communities. By addressing the fuzziness of community boundaries through membership functions, the monitoring area was ultimately divided into four main vegetation community types, with membership coefficients of 0.82, 0.79, 0.86, and 0.75 for each community.

[0142] Table 1. Environmental factor data of the salt marsh vegetation monitoring area in a certain sea area.

[0143]

[0144] In step S06, the technical team selected appropriate prediction algorithms based on different vegetation change patterns. During the monitoring period, a pattern of Suaeda salsa coverage change was detected, with a change period of 4 years. An autoregressive moving average prediction algorithm was used to process the vegetation spectral reflectance data. Simultaneously, an expansion pattern of Spartina alterniflora salt marsh was detected, and a support vector machine prediction algorithm was used to process the vegetation community segmentation results. Later in the monitoring period, a vegetation degradation pattern was also discovered, and a neural network prediction algorithm was used to process the network clustering coefficient data and shortest path length data. The prediction accuracies of each algorithm reached 89.6%, 91.2%, and 87.8%, respectively.

[0145] In step S07, the technical team used a matrix rank deficiency detection algorithm to perform quality control on the ecological feature matrix. During the monitoring process, three instances of rank deficiency were detected in the ecological feature matrix, and environmental constraint vectors were automatically added to restore the full-rank characteristic of the matrix. The established multi-scale matrix mapping hierarchical processing system achieved a step-by-step mapping from the pixel scale (30m resolution) to the landscape scale (5km resolution), outputting multi-scale ecological feature data containing four scale levels.

[0146] In step S08, the technical team inputs network clustering coefficient data, shortest path length data, and vegetation community segmentation results into the wetland ecological evolution prediction model. This model is based on a hierarchical attention network architecture, containing an input layer, five hidden layers, and an output layer, with each hidden layer containing 256 neurons. The hierarchical fusion weight adjustment function calculates the hierarchical fusion weight parameters based on soil salinity, soil moisture content, and light intensity data. The hierarchical fusion weight parameter for the *Cogon grass* area is 0.45, with a medium hierarchical fusion weight of 0.5; the hierarchical fusion weight parameter for the *Reed* area is 0.68, with a medium hierarchical fusion weight of 0.5; the hierarchical fusion weight parameter for the *Spartina alterniflora* salt marsh area is 0.25, with a low hierarchical fusion weight of 0.2; and the hierarchical fusion weight parameter for the *Suaeda salsa* salt marsh area is 0.83, with a high hierarchical fusion weight of 0.8. The model incorporates a particle swarm optimization algorithm, combining gradient information and global search results in the parameter updates for each layer, ultimately achieving a prediction accuracy of 92.3%.

[0147] Table 2 shows the future change characteristics of different vegetation types based on the predicted vegetation change trends. The coverage area of ​​the *Imperata cylindrica* community is projected to remain relatively stable over the next 10 years, with an interannual change rate of -1.2%. The *Phragmites australis* community shows a slow growth trend, with an interannual change rate of 2.8%. The *Spartina alterniflora* salt marsh exhibits a significant expansion trend, with an interannual change rate reaching 5.6%. The *Suaeda salsa* salt marsh shows a degradation trend, with an interannual change rate of -3.7%.

[0148] Table 2. Predicted Trends of Vegetation Changes in a Certain Sea Area

[0149]

[0150] In step S09, the technical team established a long-term monitoring and assessment system for salt marsh vegetation based on vegetation change trend prediction results and multi-scale ecological characteristic data. An ecological risk assessment function was used to classify the risk levels of the vegetation change trend prediction results. The ecological risk level of the *Cogon grass* community was medium risk, and the current monitoring frequency was maintained; the risk level of the *Reed* community was low risk, and the monitoring frequency was reduced to 75% of the original frequency; the risk level of the *Spartina alterniflora* salt marsh was medium risk, and the current monitoring frequency was maintained; the risk level of the *Suaeda salsa* salt marsh was high risk, and an emergency monitoring mode was activated, increasing the monitoring frequency to 150% of the original frequency. Figure 3 As shown, the ecological risk assessment results for different vegetation types exhibit significant differences.

[0151] The monitoring and assessment report shows that the vegetation of the salt marshes in a certain sea area underwent significant changes during the 35-year monitoring period. The area of ​​the *Imperata cylindrica* community decreased from 436.25 ha in 1986 to 37.2 ha in 2021, a reduction of 91.5%. The area of ​​the *Phragmites australis* community increased from 102.77 ha to 37.2 ha, but after reaching a peak of 105.21 ha in 2008, it showed a fluctuating downward trend. The *Spartina alterniflora* salt marsh community exhibited strong expansion characteristics, increasing from 0 ha in 1986 to 41.3 ha in 2021. The *Suaeda salsa* salt marsh community showed the most significant change in area, rising from 436.43 ha in 1986, reaching a peak of 486.9 ha in 1992, and then decreasing to 78.5 ha in 2021, showing an overall pattern of first increasing and then decreasing.

[0152] The technical team, through multi-algorithm ensemble prediction, found that the decline of *Imperata cylindrica* communities is mainly related to increased soil salinization and enhanced human disturbance, with its spectral characteristics showing a decrease in near-infrared reflectance from 0.68 to 0.52. *Phragmites australis* communities exhibited relatively good stability but were threatened by *Spartina alterniflora* invasion, resulting in significant ambiguity in their distribution boundaries. The rapid expansion of *Spartina alterniflora* salt marshes altered the original ecological network connectivity patterns, with the average network path length increasing from 3.8 to 5.6. The degradation of *Suaeda salsa* salt marshes was closely related to climate change and sea-level rise, with its biomass indicators showing a continuous downward trend.

[0153] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4 below.

[0154] Table 3. Variable Explanation Table (Part 1)

[0155]

[0156] Table 4. Variable Explanation Table (Part Two)

[0157]

[0158] 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 monitoring long-term changes in salt marsh vegetation, characterized by, The method comprises the following steps: A multi-level monitoring grid is established in a salt marsh wetland restoration area, a hierarchical spatial sampling method is used to set up main monitoring points and auxiliary monitoring points, soil temperature sensors, soil moisture sensors and light intensity sensors are configured to obtain soil temperature data, soil moisture data and light intensity data; multispectral remote sensing technology is used to obtain vegetation spectral reflectance data, a vegetation spectral feature database is established, and vegetation spectral reflectance data is output; wavelet transform technology is used to perform multi-time scale decomposition on the vegetation spectral reflectance data, and a seasonal change amplitude calculation matrix equation and a sampling frequency adjustment function are used to obtain an adjusted sampling frequency; an ecological network connectivity analysis algorithm based on graph theory is constructed to calculate a network clustering coefficient and a shortest path length, and network clustering coefficient data and shortest path length data are output; a fuzzy C-means clustering algorithm is used to divide the vegetation community, and the vegetation community division result is output; A matrix rank deficiency detection algorithm is used to establish a hierarchical processing system for multi-scale matrix mapping, and multi-scale ecological feature data is output; The network clustering coefficient data, the shortest path length data and the vegetation community division result are input into a wetland ecological evolution prediction model, a hierarchical fusion weight adjustment function is used to calculate hierarchical fusion weight parameters, a particle swarm optimization algorithm is embedded into the wetland ecological evolution prediction model, and vegetation change trend prediction results are output; the wetland ecological evolution prediction model is a multi-layer neural network based on a hierarchical attention network architecture, and comprises an input layer, a plurality of hidden layers and an output layer, wherein the hidden layers use a hierarchical attention mechanism to process multi-scale ecological features; the input layer receives multi-dimensional ecological feature data, including vegetation spectral reflectance, soil physicochemical properties, meteorological elements and ecological connectivity indicators; the soil physicochemical property data includes soil salt concentration, pH value, organic matter content, soil moisture data and soil temperature data, the meteorological element data includes air temperature, precipitation, relative humidity, wind speed and light intensity data; the ecological connectivity indicator data includes landscape connectivity index, habitat fragmentation degree and ecological corridor quality; a salt marsh vegetation long-term monitoring and evaluation system is established based on the vegetation change trend prediction results and the multi-scale ecological feature data, and an ecological risk assessment function is used to divide the vegetation change trend prediction results into risk grades.

2. The method of claim 1, wherein, The hierarchical spatial sampling method specifically sets the sampling point density according to the spatial heterogeneity characteristics of the ecological system, and the main monitoring points are used to obtain regional representative data, and the auxiliary monitoring points are used to capture local change details.

3. The method of claim 2, wherein, The hierarchical fusion weight adjustment function is used to dynamically adjust the hierarchical fusion weight parameters of the wetland ecological evolution prediction model to optimize the prediction performance, the input includes soil salt concentration data, soil moisture data, light intensity data and normalization coefficients, and the output is the hierarchical fusion weight parameters.

4. The method of claim 3, wherein, The step of establishing a vegetation spectral feature database, specifically combining the ground measured chlorophyll content data, plant water content data and plant stem diameter data, recording time series of continuous observation data for many years, the vegetation spectral feature database is a structured data set storing vegetation spectral reflectance data and its corresponding ecological parameters.

5. The method of claim 4, wherein, The step of multi-time scale decomposition, the main monitoring point interval is 500 m, and the auxiliary monitoring point interval is 100 m.

6. The method of claim 5, wherein, The seasonal variation amplitude calculation matrix equation, specifically the seasonal variation amplitude value obtained by processing the vegetation spectral reflectance data, is used to quantify the change degree of vegetation spectral reflectance between different seasons, the input includes vegetation spectral reflectance data, time series identifier and band weight coefficient, and the output is the seasonal variation amplitude value.

7. The method of claim 6, wherein, The sampling frequency adjustment function, specifically for dynamically adjusting the data acquisition frequency according to the seasonal variation amplitude value to optimize the monitoring efficiency, the input includes the seasonal variation amplitude value, the reference sampling frequency and the adjustment coefficient, and the output is the adjusted sampling frequency.

8. The method of claim 7, wherein, When the seasonal variation amplitude value is in the first seasonal variation threshold interval, the data sampling frequency is reduced to a preset reduction rate of the original frequency, and when the seasonal variation amplitude value is in the second seasonal variation threshold interval, the current data sampling frequency is maintained.

9. The method of claim 8, wherein, The first seasonal variation threshold interval is [15%, 25%], the second seasonal variation threshold interval is [5%, 15%], and the preset reduction rate is 60%.

Images