Method and system for steady-state conversion of early warning signal of coastal wetland ecosystem

Abstract

The invention discloses a coastal wetland ecosystem steady-state conversion early warning signal method and system, and belongs to the technical field of coastal wetland ecosystem steady-state analysis, and the method comprises the steps: S100, determining key state variables and parameters in a coastal wetland ecosystem; s200, a system dynamic model is constructed based on a Bayesian equation, the model comprises a change point detection equation in the time sequence data, and a change point detection result is obtained; s300, on the basis of the change point detection result, critical moderation signal indexes of the state variables are calculated through a critical moderation signal quantification model, and the critical moderation signal indexes comprise variance, an autocorrelation function and a recovery rate; s400, constructing an EHI ecosystem health index prediction model based on PCA dimension reduction and an LSTM neural network, and outputting a time sequence prediction result; and S500, through Bayesian dynamic threshold discrimination, in combination with a critical moderated signal index and a prediction result, outputting a graded early warning signal, and the method can effectively predict sudden ecological transition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coastal wetland ecosystem steady state analysis, more particularly, to a method and system for coastal wetland ecosystem steady state transition early warning signal. BACKGROUND

[0002] Coastal wetland ecosystems are located in the transition zone between terrestrial and marine ecosystems, containing various topographies such as estuaries, shallow seas, and beaches, and have important ecological service functions. However, due to human activities such as reclamation and pollution discharge, coastal wetland ecosystems are under great pressure and may undergo steady state transition, i.e., sudden, drastic, and persistent changes in ecosystem structure and function to another stable state, such as from a healthy wetland state to a degraded state.

[0003] Bayesian equations play a key role in analyzing coastal wetland ecosystem steady state transition early warning signals. Bayesian equations are equations involving multiple variables, including independent variables, unknown functions, and their derivatives or variational equations, which can describe the state changes of things in time and space. In the study of coastal wetland ecosystems, many phenomena in wetlands can be represented by Bayesian equations. For example, the diffusion of biological populations in wetlands, the transport of resource materials such as water and nutrients, and the interaction between organisms and the environment are all dynamic processes related to space and time, which can be described by corresponding Bayesian equations. For early warning signal analysis, we are actually looking for abnormal changes in certain variables when the system is about to undergo steady state transition. These variables may be biomass, water quality indicators, soil parameters, and other key factors related to wetland ecological health. By establishing Bayesian equation models for these factors, we can track the changes in key variables (such as the derivatives of certain variables, threshold values, etc.) over time and space. When these variables change in a special way (such as exceeding the normal fluctuation range or approaching the critical value), it can be judged as an early warning signal for possible steady state transition. Bayesian equations can describe the dynamic changes of multiple variables in an ecosystem over time and space, thus capturing the key characteristics of ecosystem steady state transition.

[0004] Many ecosystems may exhibit sudden transitions between different stable states. Due to their important ecological and economic consequences, however, how to design early warning signals to predict such sudden ecological transitions, in particular, existing theoretical studies have shown that changes in the spatial characteristics of the system can provide early warning of an impending transition. However, empirical verification of these indicators lags behind theoretical development.

[0005] Therefore, how to provide a method and system for coastal wetland ecosystem steady state transition early warning signal is a problem that those skilled in the art need to solve. SUMMARY

[0006] Therefore, the present application provides a coastal wetland ecosystem steady state conversion early warning signal method and system to solve the above technical problems in the prior art.

[0007] To achieve the above object, the present application provides the following technical solutions. A coastal wetland ecosystem steady state conversion early warning signal method, comprising the following steps: S100: determining key state variables and parameters in the coastal wetland ecosystem; S200: constructing a system dynamic model based on a Bayesian equation, the model including a variable point detection equation in time series data, to obtain a variable point detection result; S300: based on the variable point detection result, calculating critical slowing down signal indicators of the state variables through a critical slowing down signal quantification model, including variance, autocorrelation function and recovery rate; S400: constructing an EHI ecosystem health index prediction model based on PCA dimension reduction and LSTM neural network, to output a time series prediction result; S500: outputting a graded early warning signal through Bayesian dynamic threshold discrimination, in combination with the critical slowing down signal indicators and the prediction result.

[0008] Further, the state variables include at least one of vegetation coverage, soil moisture, biomass, water quality indicators and soil parameters, and the parameters include at least one of vegetation growth rate and water diffusion coefficient.

[0009] Further, the S200: constructing a system dynamic model based on a Bayesian equation, the model including a variable point detection equation in time series data, to obtain a variable point detection result, comprises: S210: setting model assumptions based on Bayesian statistical theory, defining distribution parameter differences of the state variables before and after the variable point; S220: setting prior distribution for the variable point position and parameters; S230: constructing a segmented independent likelihood function, satisfying different data distribution parameters before and after the variable point; S240: calculating the marginal posterior distribution of the variable point position through the joint posterior distribution, and determining the variable point position based on the maximum posterior probability; S250: obtaining the variable point detection equation; S260: obtaining the variable point detection result according to the variable point detection equation.

[0010] Further, the expression of the variable point detection equation is:

[0011] In the formula, tFor point-in-time / data indexing, tau To change the position of the point, T Total time duration / sample size y t These are time series observations. The average of the time series observations. mu The mean parameter, sigma 2 For variance parameter, N For quantity parameters, y t Let t be the t-th observation data point, which follows a mean of μ1 and a variance of . sigma 2 It follows a normal distribution.

[0012] Furthermore, S300: Based on the change point detection results, the critical slowing signal index of the state variable is calculated through the critical slowing signal quantization model, including variance, autocorrelation function, and recovery rate, including: S310: Calculate the variance. An increase in variance is considered a warning signal. The formula is: ; In the formula, The window variance over time t is used to detect data volatility; The range of data used when calculating variance using the sliding window length; The original data points for observations at time k; In-window mean The average value of the data up to time t: ; S320: Calculate the autocorrelation function, and determine the warning signal based on the autocorrelation time extension. The formula is: ; In the formula, AC 1( t The first-order autocorrelation coefficient measures the correlation between adjacent data points; X k+1 It is time k +1 observation value, and Xk The next adjacent data point; S330: Using a first-order autoregressive model The fitting and recovery rate are: ; In the formula, The mean within the window. For recovery rate, These are the autoregressive coefficients. This is the random error term.

[0013] Further, the S400: based on PCA dimension reduction and LSTM neural network to construct EHI ecosystem health index prediction model, output EHI time series prediction results, including: S410: principal component analysis PCA dimension reduction: input normalized multi-source data matrix , calculate the covariance matrix , extract the first Principal component, contribution rate must meet ; In the formula, X is the dimension of the standardized data matrix ; C is the covariance matrix, which describes the linear relationship between variables; n is the total number of sample points or observation values; m is the characteristic dimension of multi-source data; The first Eigenvalue, eigenvalue of covariance matrix C, reflects the variance contribution of principal component; The number of retained principal components meets the minimum integer of cumulative contribution rate ; S420: calculate the coastal wetland ecosystem health index EHI: ; Where, Health index of time t, quantitative index of coastal wetland ecosystem health degree, The weight of the first Principal component, The value of the first Principal component at time t is the dimension-reduced feature vector The weight of the first Principal component is usually taken as the eigenvalue proportion: ; ; S430: construct EHI ecosystem health index prediction model, including: input layer, hidden layer and output layer, S440: take the historical EHI sequence As the input of EHI ecosystem health index prediction model, output the prediction value of the next Months .

[0014] Further, in the S500, the dynamic threshold updating formula of Bayesian dynamic threshold discrimination is: ; Where, Dynamic threshold mean value at time t, threshold reference updated with new data; Smooth parameter , control the threshold update speed.

[0015] Further, the hierarchical early warning signal triggering condition of S500 is: the predicted value and the deviation ; Among them, the primary early warning: if the variance , and , the confidence ; High-level early warning: if the EHI ecosystem health index prediction model predicts exceeds the dynamic threshold interval, combined with the grade division: ; In the formula, sigma 历史 is the historical variance reference; AC 1 历史 is the historical autocorrelation reference | autocorrelation typical value in stable state of the system; SE is the standard error; p is the confidence.

[0016] On the other hand, the present application provides a coastal wetland ecosystem steady state conversion early warning signal system, comprising: An initialization module determines the key state variables and parameters in the coastal wetland ecosystem; A detection module constructs a system dynamic model based on the Bayesian equation, wherein the model includes a change point detection equation in time series data to obtain a change point detection result; A calculation module calculates the critical slowing down signal indicators of the state variables, including variance, autocorrelation function and recovery rate, based on the change point detection result through a critical slowing down signal quantification model; A prediction module constructs an EHI ecosystem health index prediction model based on PCA dimension reduction and LSTM neural network to output a time series prediction result; An output module outputs a hierarchical early warning signal through a Bayesian dynamic threshold judgment module combined with the critical slowing down signal indicators and the prediction result.

[0017] According to the above technical solution, compared with the prior art, the present application provides a coastal wetland ecosystem steady state conversion early warning signal method and system, which realizes quantitative early warning 3-6 months before the mutation through the mathematical model of critical slowing down signal (IV, AC1, The Bayesian threshold model is updated in real time, adapts to environmental changes, and reduces the false alarm rate The high-precision prediction: the prediction error (RMSE) of the LSTM model on the EHI sequence . BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0019] Figure 1 The method flowchart of the present application; Figure 2 The prior and posterior distribution process diagram provided for embodiment 2 of the present application; Figure 3 The key month early warning signal change point confirmation process diagram provided for embodiment 2 of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0021] Embodiment 1

[0022] Referring to Figure 1 The embodiment of the present application discloses a method for early warning signal of coastal wetland ecosystem steady state conversion, comprising the following steps: S100: determining the key state variables and parameters in the coastal wetland ecosystem; S200: constructing a system dynamic model based on the Bayesian equation, wherein the model comprises a change point detection equation in time series data, and a change point detection result is obtained; S300: based on the change point detection result, calculating the critical slowing down signal indicators of the state variables through a critical slowing down signal quantification model, including variance, autocorrelation function and recovery rate; S400: constructing an EHI ecosystem health index prediction model based on PCA dimension reduction and LSTM neural network, and outputting a time series prediction result; S500: outputting a graded early warning signal by Bayesian dynamic threshold discrimination, combining the critical slowing down signal indicators and the prediction result.

[0023] Specifically, the key variables and parameters in the coastal wetland system are identified.

[0024] In one embodiment, if the influence of vegetation coverage, soil moisture and other factors on the transition of ecosystem steady state is studied, vegetation coverage, soil moisture and other factors are variables, and the growth rate of vegetation, the diffusion coefficient of water and other factors are parameters. These variables and parameters need to be determined according to the specific research system.

[0025] The collection and preprocessing of the above multi-source data specifically include: Data sources: coastal wetland ground monitoring data: hydrology (water level , salinity , value), soil (such as organic matter content , nitrogen and phosphorus concentration ), biology (vegetation coverage , benthic organism abundance ). Remote sensing data: landscape index such as NDVI, FVC, NDWI, land surface temperature , water area change rate . Meteorological data: rainfall , wind speed .

[0026] Preprocessing: Temporal and spatial alignment: resample the remote sensing data to the same temporal and spatial resolution (such as , monthly) as the ground monitoring data.

[0027] Data interpolation: Kriging interpolation (Kriging) is used to fill in missing values, and the formula is: ; Wherein is the weight coefficient, which is determined by the semi-variogram model.

[0028] Standardization: Z-score standardization is performed on each index : ; In one embodiment, a Bayesian variable point detection equation is constructed, including: According to the principles of physics, chemistry or biology, a Bayesian equation model containing the determined variables and parameters is constructed. This process requires a deep understanding of the internal mechanism of the studied system.

[0029] The core of the Bayesian variable point detection equation constructed by the application is to infer the unknown variable point position and its statistical characteristics in the time series through a probability model. The following is a step-by-step framework and key equation (based on Bayesian statistical theory): (1) Model assumption Observation data: ; Variable point position: ; Parameters: time series observation data, where represents the observation value at time t, T is the total time length (total number of samples). is the change point position (the time when the mutation occurs), which is the core target to be detected.

[0030] Before the change point: (subject to distribution ); After the change point: (subject to distribution ).

[0031] : data distribution parameters (such as mean, variance, etc.) before the change point, subject to prior distribution p ( theta 1). : data distribution parameters after the change point, subject to prior distribution theta 2).

[0032] (2) Prior distribution

[0033] Change point position: uniform prior: ; Parameter independent prior: ; (3) Likelihood function Data segmentation is independent, and it is assumed that the distributions before and after the change point are the same but the parameters are different: ; represents that the data before the change point is subject to the distribution with parameter . represents that the data after the change point is subject to the distribution with parameter .

[0034] (4) Joint posterior distribution ; Marginal posterior distribution (integration on parameters) To obtain the marginal posterior distribution of the change point, the parameters θ1, θ2 need to be integrated: ; (5) Normal distribution mean change point (known variance) Assume that the data mean mutates at the change point, and the variance is known.

[0035] Prior: ; The former edge likelihood (analytically computable under conjugate prior) : ; wherein, ; In the formula, in the formula, t is a time point / data index, tau is a change point position, T is a total time length / sample size, y t is a time series observation value, is a time series observation value mean, mu is a mean parameter, sigma 2 is a variance parameter, N is a quantity parameter, indicating the size of the population or sample, used to calculate the statistic, y t is the tth observation data point (the result of the tth measurement), a random variable, subject to a normal distribution with a mean of μ1 and a variance of sigma 2 Specifically, μ1 represents the estimated value or true value of the variable. As a parameter to be estimated, subject to a normal distribution prior with a mean of and a variance of . is the mean parameter of the prior distribution, reflecting the initial expectation of In a normal distribution, determines the center position of the distribution. is the standard deviation parameter of the prior distribution, measuring the uncertainty of . The larger the value is, the more dispersed the prior distribution is, and the lower the sensitivity to data is. For details, see Table 1.

[0036] Table 1 Symbol meaning and explanation

[0037] Specifically, the latter edge likelihood is the same.

[0038] (6) Change point detection result

[0039] By maximizing or calculating its distribution, the most likely change point position is determined.

[0040] Early warning signal indicator equation: Variance: used to measure the fluctuation degree of the state variable. For the state variable, the calculation formula of its variance is: ; Where is the 1st observation, is the mean, and is the number of observations. Increasing variance can be a signal that the system is approaching a steady-state transition.

[0041] Autocorrelation at lag-1 (AC1): Calculate the autocorrelation coefficient at lag-1: At lag time , the formula is: ; Specifically, the autocorrelation time extension proposed in the present application, i.e. maintaining a high value at a larger value, can be an early warning signal that the system is about to undergo a steady-state transition.

[0042] Specifically, S300: based on the change point detection result, calculate the critical slowing down signal indicators of the state variable through a critical slowing down signal quantification model, including variance, autocorrelation function and recovery rate, including: Critical slowing down signal quantification model Increasing variance (Ⅳ): calculate the time series variance in a sliding window, the window length is , the step length can be set to 1 month or 1 year: ; Where, The window variance at time t is used to detect data volatility (increasing variance as an early warning signal); The data range used when calculating variance in a sliding window (such as the last w time points); The observation value at time k is the original data point (such as an ecological system indicator); The mean value in the window The average value of the data at time t: .

[0043] Autocorrelation at lag-1 (AC1): Calculate the autocorrelation coefficient at lag-1: ; In the formula, AC 1( t ) is the first-order autocorrelation coefficient, which measures the correlation between adjacent data points (autocorrelation extension as an early warning signal); X k+1 is the observation value at time k +1, and X k is the next adjacent data point, and the rest of the symbols are the same as the variance formula.

[0044] State recovery rate decreases: fit through a first-order autoregressive model , and the recovery rate .

[0045] In the formula, For window length, The mean within the window. Recovery rate reflects how quickly the system recovers from a disturbance. For autoregressive coefficients, in a first-order autoregressive model... The slope parameter, For the random error term, the model residuals (white noise process).

[0046] Specifically, S400: Based on PCA dimensionality reduction and LSTM neural network, an EHI ecosystem health index prediction model is constructed, outputting time series prediction results.

[0047] Principal Component Analysis (PCA) Dimensionality Reduction: Input standardized multi-source data matrix Calculate the covariance matrix Before extraction Each principal component must have a contribution rate that meets the following requirements. .

[0048] In the formula, X represents the dimension of the standardized data matrix. C is the covariance matrix, describing the linear relationship between variables; n is the number of samples, time points, or the total number of observations; m is the number of variables and the feature dimension of the multi-source data. For the first The eigenvalues ​​of the covariance matrix C reflect the variance contributions of the principal components. To ensure the number of principal components retained meets the cumulative contribution rate requirement The smallest integer; Coastal Wetland Ecosystem Health Index (EHI): ; in, The health index over time t is a quantitative indicator of the health status of coastal wetland ecosystems. For the first The weights of each principal component (determined by the proportion of eigenvalues). For the first The principal components are in The eigenvectors after dimensionality reduction at each time step No. The weights of each principal component are typically taken as the proportion of eigenvalues: .

[0049] Constructing an EHI ecosystem health index prediction model (dynamic threshold early warning model), including: LSTM prediction model: The network structure includes an input layer and hidden layers (number of neurons). The forward propagation formula for the output layer and the forward layer is: ; Input: historical EHI sequence , output: predicted value of future months .

[0050] The above formula, : input value at time t (i.e. ). : hidden state at previous time (dimension determined by number of neurons ). : state at previous time. : weight matrix (corresponding to forget gate, input gate, candidate memory and output gate respectively); : bias vector. : sigmoid activation function (output value between 0 and 1). : hyperbolic tangent activation function (output value between -1 and 1). : element-wise multiplication (Hadamard product); : vector concatenation operation; : hidden state at current time (also the output of LSTM unit).

[0051] The dynamic threshold updating formula of Bayesian dynamic threshold discrimination is: ; Among them , dynamic threshold mean at time t, threshold reference updated with new data; smooth parameter (learning rate), , control the threshold update speed (the smaller the value, the smoother).

[0052] Specifically, the output rules of early warning signals are: Primary warning: if the variance , and , the confidence ; Advanced warning: if the LSTM prediction exceeds the dynamic threshold, combined with risk level division: ; In the formula, sigma 历史 is the historical variance reference, the typical value of variance in the stable state of the system; AC 1 历史 , historical autocorrelation reference | typical value of autocorrelation in stable state of the system; SE is the standard error, used to calculate the confidence interval (1.96.SE corresponding to a 95% confidence interval; p a confidence level (p-value) of 0.05 indicates that the difference is statistically significant. p a confidence level (p-value) of 0.05 indicates that the difference is statistically significant. a confidence level (p-value) of 0.05 indicates that the difference is statistically significant.

[0053] Specifically, the present application is suitable for the steady state conversion early warning signal identification of the Yellow River Delta salt marsh, the Liaohe Estuary salt marsh of Suaeda salsa and the Yellow River Delta light beach.

[0054] Specifically, the analysis results of the early warning signal indicators are realized through a statistical toolbox, which includes a change point detection algorithm, a variance analysis module and a autocorrelation function calculation module.

[0055] Specifically, the change point detection results are displayed through a visualization tool, including a probability distribution graph of the change point position and a spatiotemporal dynamic change graph of the early warning signal indicators.

[0056] Specifically, the present application further includes spatial heterogeneity analysis based on the early warning signal, which identifies the potential risk level of the steady state conversion in different regions of the coastal wetland ecosystem.

[0057] On the other hand, the present application also discloses a system for steady state conversion early warning signal of coastal wetland ecosystem, comprising: an initialization module for determining the key state variables and parameters in the coastal wetland ecosystem; a detection module for constructing a system dynamic model based on the Bayesian equation, wherein the model includes a change point detection equation in the time series data, and the change point detection result is obtained; a calculation module for calculating the critical slowing down signal indicators of the state variables, including variance, autocorrelation function and recovery rate, based on the change point detection result through the critical slowing down signal quantification model; a prediction module for constructing an EHI ecosystem health index prediction model based on PCA dimension reduction and LSTM neural network, and outputting the time series prediction result; an output module for outputting the graded early warning signal through the Bayesian dynamic threshold judgment module in combination with the critical slowing down signal indicators and the prediction result.

[0058] Specifically, in addition to providing a step-by-step method for applying these signals to spatial data sets, a statistical tool is also proposed, which can be used to help detect approaching transitions in various spatial data, thereby promoting the application and testing of spatial early warning signals on real spatial data.

[0059] Example 2: Based on the method and system for steady state conversion early warning signal of coastal wetland ecosystem provided by the present application, the steady state conversion early warning signal of the Yellow River Delta salt marsh is identified, and the results are shown in Figures 2-3 , which specifically includes: First step: First, Landsat data from 1985 to 2023 (from Landsat-1 to Landsat-9 satellites) are prepared, and NDVI index data are calculated. At the same time, humidity, temperature, and salinity data are collected for calculation. See Table 2 for details.

[0060] NDVI index calculation formula: NDVI = (NIR - R) / (NIR + R), where NIR represents the reflectivity of the near-infrared band, and R represents the reflectivity of the red band. If the near-infrared band reflectivity of a pixel is 0.8, and the red band reflectivity is 0.4, the NDVI value of the pixel is calculated.

[0061] Table 2 Variable parameter table

[0062] Second step: Bayesian change point detection (S200) (convert the calculation formula of this step into python code): import pymc3 as pm import arviz as az # Seasonal adjustment from statsmodels.tsa.seasonal import STL stl = STL(ndvi, period=12, seasonal=13) res = stl.fit() ndvi_deseason = ndvi - res.seasonal # MCMC sampling with pm.Model() as cp_model: # Prior distribution tau = pm.DiscreteUniform("tau", lower=24, upper=456) # Exclude the first and last 2 years mu1 = pm.Normal("mu1", mu=0.6, sigma=0.1) sigma1 = pm.HalfNormal("sigma1", sigma=0.05) mu2 = pm.Normal("mu2", mu=0.5, sigma=0.1) sigma2 = pm.HalfNormal("sigma2", sigma=0.05) Likelihood function mu = pm.math.switch(tau > np.arange(months), mu1, mu2) sigma = pm.math.switch(tau > np.arange(months), sigma1, sigma2) y_obs = pm.Normal("y_obs", mu=mu, sigma=sigma, observed=ndvi_deseason) # Sampling trace = pm.sample(5000, tune=2000, chains=4, target_accept=0.95) # Change point analysis tau_samples = trace.get_values("tau", combine=True) tau_mode = np.argmax(np.bincount(tau_samples)) change_date = time_index[tau_mode] The calculation results are shown in Table 3: Table 3 Main change point and secondary change point calculation results

[0063] Calculation description: Use STL decomposition to remove seasonal effects; build a piecewise normal distribution model; MCMC sampling 5000 times (annealing 2000 times); the posterior distribution of the change point position τ shows two significant peaks; the change point in 1996 corresponds to the intensification of the Yellow River cutoff, and the change point in 2007 corresponds to the ecological restoration project.

[0064] Step 3: Quantification of critical slow-down signals (S300) (convert the calculation formula of this step into python code)

[0065] w = 60 # 5-year sliding window

[0066] variances = np.full(months, np.nan)

[0067] ac1_values = np.full(months, np.nan)

[0068] kappa_values = np.full(months, np.nan)

[0069] for t in range(w, months):

[0070] window = ndvi_deseason[t-w:t]

[0071] mean_val = np.mean(window)

[0072] # Variance calculation

[0073] variances[t] = np.mean((window - mean_val) 2)

[0074] # Autocorrelation calculation

[0075] diff = window - mean_val

[0076] numerator = np.sum(diff[:-1] diff[1:])

[0077] denominator = np.sum(diff 2)

[0078] ac1_values[t] = numerator / denominator if denominator > 1e-6 else 0

[0079] # Recovery rate calculation

[0080] X = window[:-1]

[0081] y = window[1:]

[0082] a = np.linalg.lstsq(X.reshape(-1,1), y, rcond=None)[0][0]

[0083] kappa_values[t] = -np.log(abs(a)) if 0 < abs(a) < 1 else np.nan

[0084] # Historical benchmark calculation (1985-1996 stable period)

[0085] hist_var = np.nanmean(variances[84:180]) # 1985-1996

[0086] hist_ac1 = np.nanmean(ac1_values[84:180])

[0087] hist_kappa = np.nanmean(kappa_values[84:180])

[0088] Typical example results in December 2022 are shown in Table 4: Table 4 Calculation results of each index in December 2022

[0089] Calculation note: Use a 5-year sliding window to calculate the index, and the historical benchmark is the mean value of the stable period from 1985 to 1996. The variance in 2022 increased by 412%, more than twice the historical benchmark (0.0050), the autocorrelation increased by 95.4%, more than the historical benchmark +1.96SE (0.392), and the recovery rate decreased by 54.8%, indicating that the system's recovery ability has weakened.

[0090] Step 4: EHI health index prediction (S400) (convert the calculation formula of this step into python code)

[0091] PCA dimensionality reduction calculation: from sklearn.decomposition import PCA from sklearn.preprocessing import StandardScaler # Create a multi-source data matrix X = np.column_stack([ndvi, precip, temp, salinity]) X_scaled = StandardScaler().fit_transform(X) # PCA dimensionality reduction pca = PCA(n_components=0.85) X_pca = pca.fit_transform(X_scaled) print(f"Main components retained: {pca.n_components_}, Cumulative variance: {sum(pca.explained_variance_ratio_):.2%}") # EHI Calculation ehi = np.zeros(months) for i in range(pca.n_components_): ehi += pca.explained_variance_ratio_[i] X_pca[:, i] The PCA results are shown in Table 5: Table 5. PCA Results (Illustrated)

[0092] LSTM prediction model: from tensorflow.keras.models import Sequential from tensorflow.keras.layers import LSTM,Dense # Prepare training data lookback = 36 # 3-year history forecast = 6 # 6-month forecast X_train, y_train = [], [] for i in range(lookback, months-forecast): X_train.append(ehi[i-lookback:i]) y_train.append(ehi[i:i+forecast]) X_train = np.array(X_train)[..., np.newaxis] y_train = np.array(y_train) # Model Building model = Sequential([ LSTM(64, input_shape=(lookback, 1)), Dense (forecast) ]) model.compile(optimizer='adam', loss='mse') history = model.fit(X_train, y_train, epochs=100, batch_size=32, verbose=0) # 2023 Predictions X_pred = ehi[-lookback:].reshape(1, lookback, 1) ehi_pred = model.predict(X_pred).flatten() The forecast results are shown in Table 6 (taking 2023 as an example): Table 6. Schematic diagram of prediction results

[0093] Model evaluation: Training loss: 0.0012 (MSE); Test set R²: 0.89; Prediction uncertainty: ±0.02 (95% confidence interval)

[0094] Step 5: Dynamic Threshold Early Warning (S500) (Convert the calculation formula for this step into Python code)

[0095] eta = 0.05 # Smoothing coefficient

[0096] mu = np.zeros(months)

[0097] sigma = np.zeros(months)

[0098] mu[0] = ehi[0]

[0099] sigma[0] = 0.01

[0100] for t in range(1, months):

[0101] # Mean Update

[0102] mu[t] = mu[t-1] + eta (ehi[t] - mu[t-1])

[0103] # Standard Deviation Update

[0104] sigma[t] = np.sqrt((1-eta) sigma[t-1] 2 + eta (ehi[t]-mu[t-1]) 2)

[0105] # Current threshold parameters (2022-12)

[0106] current_mu = mu[-1] # 0.612

[0107] current_sigma = sigma[-1] # 0.042

[0108] threshold_low = current_mu - 2 current_sigma # 0.528

[0109] threshold_high = current_mu + 2 current_sigma # 0.696

[0110] Early warning signal analysis: Primary early warning condition: Variance 0.0128 > 2x0.0025 = 0.0050 → Satisfied; Autocorrelation 0.635 > 0.325 + 1.96x0.034 = 0.392 → Satisfied; P-value: <0.001 (t-test) < 0.05 → Satisfied, → Trigger primary early warning.

[0111] Forecast value analysis see Table 7: Table 7 Forecast value analysis table

[0112] Advanced early warning level: Maximum deviation 33.6% ≥ 30%, exceed threshold months: 5 / 6 months → Trigger advanced early warning (high level).

[0113] Through the complete calculation process, the steady state conversion early warning of Yellow River Delta salt marsh wetland is realized, and the turning point detection accurately identifies two ecological turning points in 1996 and 2007. The turning point in 1996 corresponds to the intensification of Yellow River cutoff (NDVI decreased by 20%), and the turning point in 2007 corresponds to ecological restoration project (NDVI rebounded by 11.5%).

[0114] Critical slowing down signal 2022 variance increased by 412% (exceeded threshold 157%), autocorrelation increased by 95.4% (exceeded threshold 62%), recovery rate decreased by 54.8%, and early warning of unstable state of ecological system for 24 months; Health prediction, PCA-LSTM model effectively integrates multi-source data (R²=0.89), predicts that EHI will decrease by 33.6% in 2023, the maximum deviation occurs in June 2023, triggering high-level warning (red alert), the main driving factor is soil salinity (r=-0.82), and the recommended priority measures are salinity control (BCR=0.126); The early warning of the ecosystem risk for 6-24 months is realized, which provides scientific decision support for the protection of coastal wetlands.

[0115] Embodiment 3: Based on the method and system for providing a coastal wetland ecosystem steady-state conversion early warning signal provided by the present application, the Yellow River Delta tidal wetland is early warned and identified, and through the steps of embodiment 1, data collection is obtained: ground monitoring: salinity The monthly average value increases from 1.5 ‰ to 3.2 ‰, and the variance increases from 0.4 to 1.1. Remote sensing data: NDVI decreases from 0.6 to 0.4, and the ground temperature rises by 2.5℃.

[0116] Model calculation: extract principal components (PCA) ), build EHI index, and the historical average value is EHI=0.8. LSTM predicts that EHI will decrease to 0.5 in the next 6 months, which exceeds the lower limit of the dynamic threshold . Early warning output: trigger high-level warning (risk level "high"), prompt salinization risk, and EHI rebounds to 0.7 after the ecological water supplement implemented by the management department.

[0117] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, it is described more simply, and the related parts can be referred to the method part.

[0118] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for early warning signals of steady-state transitions in coastal wetland ecosystems, characterized in that, Includes the following steps: S100: Identify key state variables and parameters in coastal wetland ecosystems; S200: Construct a system dynamic model based on Bayesian equations, the model including a change point detection equation in time series data, and obtain change point detection results; S300: Based on the change point detection results, calculate the critical slowing signal index of the state variable through the critical slowing signal quantization model, including variance, autocorrelation function and recovery rate; S400: Construct an EHI ecosystem health index prediction model based on PCA dimensionality reduction and LSTM neural network, and output time series prediction results; S500: By using Bayesian dynamic threshold discrimination, combined with the critical slowdown signal index and prediction results, a graded early warning signal is output.

2. The method for early warning signal of steady-state transition in coastal wetland ecosystems according to claim 1, characterized in that, The state variables include at least one of vegetation cover, soil moisture, biomass, water quality indicators, and soil parameters, and the parameters include at least one of vegetation growth rate and water diffusion coefficient.

3. The method for early warning signal of steady-state transition in coastal wetland ecosystems according to claim 1, characterized in that, S200: Constructing a system dynamic model based on Bayesian equations, the model includes a change point detection equation in time series data, obtaining change point detection results, including: S210: Based on Bayesian statistical theory, set model assumptions and define the differences in distribution parameters of state variables before and after the change point; S220: Set the prior distribution for the variable point location and parameters; S230: Construct piecewise independent likelihood functions to ensure that the data distribution parameters are different before and after the change point; S240: Calculate the marginal posterior distribution of the change point location through the joint posterior distribution, and determine the change point location based on maximizing the posterior probability; S250: Obtain the change point detection equation; S260: Obtain the change point detection result based on the change point detection equation.

4. The method for early warning signal of steady-state transition in coastal wetland ecosystems according to claim 3, characterized in that, The expression for the change point detection equation is: ； In the formula, t For point-in-time / data indexing, τ To change the position of the point, T Total time duration / sample size y t These are time series observations. The average of the time series observations. μ The mean parameter, σ 2 For variance parameter, N For quantity parameters, y t Let t be the t-th observation data point, which follows a mean of μ1 and a variance of . σ 2 It follows a normal distribution.

5. The method for early warning signal of steady-state transition in coastal wetland ecosystems according to claim 1, characterized in that, S300: Based on the change point detection results, the critical slowing signal index of the state variable is calculated through the critical slowing signal quantization model, including variance, autocorrelation function, and recovery rate, including: S310: Calculate the variance. An increase in variance is considered a warning signal. The formula is: ； In the formula, The window variance over time t is used to detect data volatility; The range of data used when calculating variance using the sliding window length; The original data points for observations at time k; In-window mean The average value of the data up to time t: ; S320: Calculate the autocorrelation function, and determine the warning signal based on the autocorrelation time extension. The formula is: ； In the formula, AC 1( t The first-order autocorrelation coefficient measures the correlation between adjacent data points; X k+1 It is time k +1 observation value, and Xk The next adjacent data point; S330: Using a first-order autoregressive model The fitting and recovery rate are: ； In the formula, The mean within the window. For recovery rate, These are the autoregressive coefficients. This is the random error term.

6. The method for early warning signal of steady-state transition in coastal wetland ecosystems according to claim 5, characterized in that, The S400: Based on PCA dimensionality reduction and LSTM neural network, an EHI ecosystem health index prediction model is constructed, outputting EHI time-series prediction results, including: S410: Perform Principal Component Analysis (PCA) for dimensionality reduction: Input the standardized multi-source data matrix Calculate the covariance matrix Before extraction Each principal component must have a contribution rate that meets the following requirements. ; In the formula, X represents the dimension of the standardized data matrix. C is the covariance matrix, describing the linear relationship between variables; n is the number of samples, time points, or the total number of observations; m is the number of variables and the feature dimension of the multi-source data. For the first The eigenvalues ​​of the covariance matrix C reflect the variance contribution of the principal components. To ensure the number of principal components retained meets the cumulative contribution rate requirement The smallest integer; S420: Calculating the Ecosystem Health Index (EHI) for Coastal Wetlands ； in, The health index over time t is a quantitative indicator of the health status of coastal wetland ecosystems. For the first The weights of each principal component, No. The principal components are in The eigenvectors after dimensionality reduction at each time step No. The weights of each principal component are typically taken as the proportion of the eigenvalues: ; S430: Construct an EHI ecosystem health index prediction model, including: input layer, hidden layer, and output layer. S440: Transfer historical EHI sequence As input to the EHI ecosystem health index prediction model, the output is the future. Monthly forecast .

7. The method for early warning signal of steady-state transition in coastal wetland ecosystems according to claim 6, characterized in that, In S500, the dynamic threshold update formula for Bayesian dynamic threshold discrimination is: ； in, The dynamic threshold mean at time t, and the threshold benchmark updated with new data; Smoothing parameters, Control the threshold update speed.

8. The method for early warning signal of steady-state transition in coastal wetland ecosystems according to claim 7, characterized in that, The triggering condition for the graded early warning signal of the S500 is: predicted value. And deviation ; Among them, the primary warning: if the variance is satisfied at the same time ,and Confidence level ; Advanced warning: If the EHI ecosystem health index prediction model predicts... Values ​​exceeding dynamic thresholds, combined with grade classification: ； In the formula, σ 历史 Use historical variance as the benchmark; AC 1 历史 Historical autocorrelation benchmark | Typical autocorrelation value under steady-state conditions of the system; SE This is a standard error; p , where is the confidence level.

9. A system for early warning signals of steady-state transitions in coastal wetland ecosystems, characterized in that, include: Initialization module to determine key state variables and parameters in coastal wetland ecosystems; The detection module constructs a system dynamic model based on Bayesian equations. The model includes a change point detection equation in time series data to obtain change point detection results. The calculation module, based on the change point detection results, calculates the critical slowing signal index of the state variable through the critical slowing signal quantization model, including variance, autocorrelation function and recovery rate; The prediction module constructs an EHI ecosystem health index prediction model based on PCA dimensionality reduction and LSTM neural network, and outputs time series prediction results. The output module, through the Bayesian dynamic threshold discrimination module, combines the critical slowing signal index with the prediction results to output a graded early warning signal.

Images