An estuary wetland degradation risk ai early warning method
By using the BiLSTM-Attention model and leveraging the MODIS NDVI time series for bidirectional time series modeling and attention mechanism, the problems of data dependence and insufficient early warning in estuarine wetland degradation monitoring are solved. This enables efficient and interpretable prediction of wetland degradation trends and critical early warning, and is applicable to multiple estuarine wetlands.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2026-05-09
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for monitoring estuarine wetland degradation rely on multi-source data, and the models lack two-way time-series modeling capabilities and the ability to focus on key time steps. They also lack critical early warning mechanisms, making it difficult to achieve rapid, large-scale, and forward-looking early warning of degradation.
A BiLSTM-Attention model is adopted, which uses MODIS NDVI time series for bidirectional time series modeling and combines attention mechanism to adaptively focus on key time features to achieve critical degradation early warning.
It achieves efficient and interpretable wetland degradation trend prediction and critical early warning using only the MODIS NDVI single indicator, supports near real-time monitoring and hierarchical early warning, and is applicable to multiple estuarine wetlands.
Smart Images
Figure CN122286227A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of ecological environment remote sensing monitoring and deep learning application technology. Specifically, it relates to a method that uses BiLSTM combined with attention mechanism based solely on the MODIS NDVI single index to predict the future degradation trend of estuarine wetlands (taking the Yangtze River Estuary as an example) and achieve critical degradation early warning. Background Technology
[0002] Estuarine wetlands are key zones for land-sea interaction and possess important ecosystem service functions. The Yangtze River Estuary wetlands are important coastal wetlands in my country, facing significant degradation risks due to multiple pressures including reduced sediment loads from upstream, sea-level rise, and human reclamation. Traditional wetland degradation monitoring relies on multi-source data (such as soil, hydrology, and vegetation data) and physical models, which are costly to acquire and involve complex parameters, making it difficult to achieve rapid, large-scale, and forward-looking degradation early warning.
[0003] Normalized Difference Vegetation Index (NDVI) is an effective indicator reflecting vegetation cover and growth status. MODIS NDVI has the advantages of high temporal resolution (16 days / month) and long-term continuity (2000 to present), and can independently reflect the dynamic changes of wetland vegetation ecosystems. However, existing studies mostly use NDVI for historical trend analysis or simple statistical prediction, lacking in-depth modeling of its nonlinear evolution. More importantly, traditional one-way LSTM can only capture forward dependencies from the past to the future, and cannot fully explore the bidirectional contextual information in the NDVI sequence; at the same time, LSTM treats all time steps equally, making it difficult to focus on the most critical "inflection points" or "abnormal fluctuation periods" for predicting degradation trends.
[0004] Therefore, developing a method that relies solely on the MODIS NDVI single index, utilizes bidirectional LSTM to capture complete temporal dependencies, and adaptively focuses on key temporal features using an attention mechanism, while simultaneously providing automatic early warning of critical degradation, has significant application value. Summary of the Invention
[0005] This invention addresses the problems of existing estuarine wetland degradation prediction methods, such as reliance on multi-source data, lack of bidirectional time series modeling capabilities and key time step focusing capabilities, and lack of critical early warning mechanisms. It provides a degradation trend prediction and critical early warning method based on a BiLSTM-Attention model that uses only MODIS NDVI time series data.
[0006] To achieve the above objectives, the present invention provides the following technical solution: 1. Data Acquisition and Preprocessing The MODIS NDVI product (MOD13Q1) of the Yangtze River Estuary wetland from 2000 to the present was obtained from Google Earth Engine, and the monthly maximum NDVI composite value of the wetland area was extracted.
[0007] Preprocessing: SG filtering (Savitzky-Golay) is used to smooth noise, linear interpolation is used to fill in missing values caused by cloud masking, and outliers outside the range of [-0.2, 1] are removed.
[0008] 2. Sample Construction NDVI time series Divide into a sliding window sequence. Set the historical time step. (Month), predicted step size (Month). Each training sample is... .
[0009] 3. Deep Learning Model Construction (BiLSTM + Attention) 3.1 BiLSTM Layer Input sequence The hidden state at each time step is obtained by passing the data through a forward LSTM and a backward LSTM. and spliced together The number of forward and backward hidden units is set to 32~64, and the total output dimension is 64~128.
[0010] 3.2 Attention Layer Hidden state sequence of BiLSTM output Calculate attention weights. Use additive attention:
[0011]
[0012]
[0013] Where c is the weighted context vector, representing the model's focus on the most critical time step in the historical sequence.
[0014] 3.3 Output Layer The context vector c, after passing through Dropout(0.2), is input to the fully connected layer and outputs... One future NDVI prediction value.
[0015] Loss function: Mean squared error (MSE); Optimizer: Adam, learning rate 0.001.
[0016] 4. Model Training Training set:Validation set:Test set = 7:1.5:1.5. Early stopping (patience=10) prevents overfitting.
[0017] 5. Future Trend Forecast Recursive multi-step prediction is adopted: the predicted first step is used as the input for the next step, and the NDVI is predicted in a rolling manner for the next 6 to 24 months.
[0018] 6. Degradation rate and critical warning Calculate the degradation rate of the predicted sequence Degradation acceleration .
[0019] Critical degradation threshold ,in This represents the average NDVI value for the same period in historically healthy years.
[0020] Early warning logic: like and and If the rate is less than the historical average, a Level 1 warning will be issued. If at the same time Then a level II warning will be issued; If at the same time Then, a Level 3 warning (critical state) is issued.
[0021] Beneficial effects 1. Minimal data dependency: It only uses MODIS NDVI, a public, free, long-term remote sensing indicator, without the need for on-site sampling or auxiliary environmental data, which greatly reduces the application threshold.
[0022] 2. Advantages of bidirectional time series modeling: BiLSTM utilizes both past and future information from the NDVI sequence (during the training phase), which can capture the dependencies between the annual rhythms of wetland vegetation more completely than unidirectional LSTM, especially its ability to represent mutation inflection points.
[0023] 3. Attention mechanism enhances interpretability and prediction accuracy: The attention layer automatically calculates the importance weight of each historical time step, enabling the model to focus on the key periods that contribute the most to the degradation trend (such as consecutive drought periods and abnormally warm winters), while providing ecological managers with a visual explanation of "which historical periods are most worthy of attention".
[0024] 4. For the first time, a critical degradation early warning mechanism is introduced: It not only predicts future trends, but also quantifies the degradation rate and acceleration, and issues graded early warnings based on historical statistical thresholds, especially identifying the critical state that is about to cross the inflection point of irreversible degradation.
[0025] 5. High portability: This method is not only applicable to the Yangtze River Estuary, but can also be applied to other estuarine wetlands such as the Yellow River Estuary and the Pearl River Estuary with slight adjustments.
[0026] 6. Supports near real-time monitoring: Combined with MODIS near real-time data stream, monthly or quarterly rolling early warning updates can be achieved. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the prediction results of the present invention; Figure 2 This is a schematic representation of the calculation results of the present invention; Figure 3 This is a schematic diagram of the early warning logic determination of the present invention; Detailed Implementation
[0028] The following section uses Chongming Dongtan, a typical wetland in the Yangtze River estuary, as an example to illustrate the specific implementation steps of this invention.
[0029] Example Step 1: Data Acquisition Data source: MODIS / Terra Vegetation Indices 16-Day L3 Global 250m (MOD13Q1v6) data; Time range: February 18, 2000 to February 18, 2025, a total of 25 complete years; Spatial scope: The core area of Chongming Dongtan Wetland, specifically a rectangular area bounded by 121.9°E to 122.1°E and 31.5°N to 31.7°N, covering approximately 400 square kilometers; Raw data bands: NDVI bands (band 1) were extracted, with a spatial resolution of 250m and a temporal resolution of 16 days.
[0030] Data volume: 25 years × 23 periods / year = 575 NDVI images.
[0031] Data acquisition method: Batch calls via the Google Earth Engine Python API.
[0032] Step 2: Preprocessing 2.1 Cloud Masking Using the pixel_reliability band built into the MODIS product, pixels with quality levels of 0 (good) and 1 (edge) are retained, while pixels with quality levels of 2 (snow / ice) and 3 (cloud) are removed.
[0033] Cells covered by clouds within each 16-day period are marked as missing values (NaN).
[0034] 2.2 Monthly Value Composition Since the 16-day cycle is not perfectly aligned with the calendar month, a time-weighted maximum value synthesis method is used: for each calendar month, the maximum value of all 16-day NDVI values within that month is selected as the NDVI for that month. Maximum value synthesis can further reduce the impact of atmospheric residues and represents the optimal vegetation growth state.
[0035] 2.3 Time Series Reconstruction and Smoothing Savitzky-Golay (SG) filtering was used for smoothing, with the following parameters: window size = 7 (months), polynomial order = 2. SG filtering effectively suppresses high-frequency noise while preserving the seasonal variation characteristics of NDVI. After filtering, the peak signal-to-noise ratio (PSNR) of the NDVI sequence improved by approximately 4.2 dB.
[0036] 2.4 Missing Value Imputation For consecutive missing values of no more than 3 months, cubic spline interpolation was used to fill the gaps. For consecutive missing values of more than 3 months (which did not occur in this dataset), the data for that year was discarded and skipped in subsequent sample construction. This resulted in 300 consecutive monthly NDVI values from January 2000 to December 2024, with a sequence range of 0.12–0.78 and no missing values.
[0037] 2.5 Outlier Removal The valid range for NDVI is set to [-0.2, 1.0], and all values in this sequence fall within this range. Outlier detection principle: Calculate the mean of the sequence. and standard deviation ,Exceed Points were considered outliers and replaced with the mean of the preceding and following months. Two outliers were detected in this sequence (July 2003 and February 2016), which have been corrected.
[0038] 2.6 Normalization To accelerate model convergence, the NDVI sequence is linearly normalized to the range [0,1].
[0039] in , Subsequent prediction outputs need to be denormalized back to the original scale.
[0040] Step 3: Sample Construction 3.1 Parameter Settings Historical time step (input window) (Months), meaning using NDVI data from the past two years. Prediction step size (output window) (Monthly), which is the monthly NDVI forecast for the next six months. The sliding step is 1 month.
[0041] 3.2 Sample Generation Algorithm Total time length Months. Total sample size: indivual.
[0042] The i-th sample (i starts from 0): enter (24 values); Label (6 values).
[0043] 3.3 Dataset Partitioning Divide the data chronologically, without using random partitioning, to maintain temporal integrity: Training set: Samples from January 2000 to December 2016 (i=0 to i=190), a total of 190 samples.
[0044] Validation set: Samples from January 2017 to June 2019 (i=191 to i=230), a total of 40 samples.
[0045] Test set: Samples from July 2019 to December 2024 (i=231 to i=271), a total of 41 samples.
[0046] The validation set is used for early stopping and hyperparameter tuning, while the test set is used for final evaluation.
[0047] 3.4 Data Shape The input shape for each sample is (24, 1), with 24 time steps and 1 feature (NDVI) per step.
[0048] Shapes used during batch training: ( , 24, 1).
[0049] Step 4: Deep Learning Model Construction (BiLSTM + Attention) 4.1 Overall Model Architecture Input layer: Input (shape=(24, 1)) BiLSTM layer: Bidirectional (LSTM(64, return_sequences=True)) - Forward LSTM: 64 units, output h_forward (24, 64) - Backward LSTM: 64 units, output h_backward (24, 64) - Concatenate: h = concatenate([h_forward, h_backward]) (24, 128) Attention layer: Attention (units=64) Dropout layer: Dropout (rate=0.2) Fully connected layer 1: Dense (32, activation='relu') Dropout layer: Dropout (rate=0.2) Output layer: Dense (6, activation='linear') 4.2 Detailed Configuration of BiLSTM Layer Use Keras's Bidirectional wrapper.
[0050] Basic LSTM unit: LSTM(64, return_sequences=True, activation='tanh', recurrent_activation='sigmoid').
[0051] There are 64 units each for the forward and backward directions, for a total output dimension of 128.
[0052] Total number of parameters: Number of parameters in one LSTM unit = 4 × (units × (input_dim + units + 1)). With an input dimension of 1 and units = 64, the number of parameters in a single LSTM is approximately 4 × (64 × 65) = 16,640. The number of parameters in a bidirectional LSTM is approximately 33,280.
[0053] 4.3 Detailed Design of Attention Layer (Additive Attention) Input: The hidden state sequence output by BiLSTM Each .
[0054] Learnable parameters: weight matrix bias ,vector .
[0055] Calculation process: (Obtain the scalar energy value) (Attention weights are obtained by Softmax normalization) (Weighted summation yields the context vector) Output: Context vector .
[0056] 4.4 Output Layer The context vector is processed by Dropout(0.2), Dense(32, relu), Dropout(0.2), and Dense(6, linear), outputting 6 values, which correspond to the NDVI prediction (normalized value) for the next 6 months.
[0057] 4.5 Model Compilation Loss function: Mean Squared Error (MSE).
[0058] Optimizer: Adam, initial learning rate = 0.001, beta_1 = 0.9, beta_2 = 0.999.
[0059] Evaluation metrics: Mean absolute error (MAE), root mean square error (RMSE).
[0060] Step 5: Model Training 5.1 Training Configuration Environment: Python 3.9, TensorFlow 2.12, CUDA 11.8, single NVIDIA Tesla T4 GPU.
[0061] Batch size: 32.
[0062] Maximum number of training rounds: 200.
[0063] Early Stopping: Monitors the validation set loss, patience=10, meaning that training stops if the validation loss does not decrease for 10 consecutive rounds.
[0064] Learning rate decay: ReduceLROnPlateau, patience=5, factor=0.5, min_lr=1e-6.
[0065] 5.2 Training Process Recording Round 1: Training loss = 0.0876, validation loss = 0.0652.
[0066] Round 10: Training loss = 0.0321, validation loss = 0.0289.
[0067] Round 25: Training loss = 0.0185, validation loss = 0.0193.
[0068] Round 40: Training loss = 0.0122, validation loss = 0.0167.
[0069] Round 52: Verification loss reached the minimum value of 0.0158.
[0070] Round 62: Validation loss = 0.0179 (no improvement after 10 consecutive rounds), early stop triggered, training terminated.
[0071] The final model uses the weights saved from round 52.
[0072] 5.3 Training Results Training time: Approximately 18 minutes (62 rounds).
[0073] Final training loss: 0.0115, training MAE: 0.0213.
[0074] Final validation loss: 0.0158, validation MAE: 0.0256.
[0075] 5.4 Test Set Evaluation The evaluation was conducted on 41 test samples, and the results are as follows: MAE = 0.028 (normalized value), corresponding to an absolute NDVI error of approximately 0.028 × (0.78-0.12) = 0.0185 based on the original NDVI scale. RMSE = 0.041. Relative error = 4.8% (based on an average NDVI of 0.55).
[0076] Comparative experiment (same data, same parameter settings): One-way LSTM (without Attention): MAE=0.035, RMSE=0.052.
[0077] BiLSTM (without Attention): MAE=0.031, RMSE=0.047.
[0078] BiLSTM + Attention (this invention): MAE=0.028, RMSE=0.041.
[0079] The results show that BiLSTM improves performance by about 11.4% compared to unidirectional LSTM, and further improves it by about 9.7% after adding Attention.
[0080] 5.5 Visual Analysis of Attention Weights Extract attention weights from all samples in the test set. And take the average according to the time position.
[0081] The results showed that the model assigned the highest average weight (0.12-0.15) to the 13th to 18th time steps in the historical window (i.e. the latter half of the input window, about 6 to 12 months before the prediction start point), while the weight of the most recent 1 to 2 months (steps 23 to 24) was relatively low (0.02-0.04).
[0082] Ecological explanation: Wetland degradation does not occur suddenly, but rather has a "latent degradation period" of about six months to a year—when vegetation vitality begins to decline but NDVI has not yet decreased significantly. The BiLSTM-Attention model successfully captures this early warning signal, which is something traditional methods cannot do.
[0083] Step 6: Future Trend Prediction (Application Phase) 6.1 Input Data Preparation The measured MODIS NDVI values for the most recent 24 months (January 2023 to December 2024) were used as input. The original NDVI values are: [0.61, 0.58, 0.63, 0.68, 0.71, 0.72, 0.69, 0.66, 0.62, 0.57, 0.52, 0.48, 0.59, 0.56, 0.62, 0.66, 0.69, 0.70, 0.67, 0.64, 0.60, 0.55, 0.50, 0.46]. These values were then normalized and input into the model.
[0084] 6.2 Recursive Multistep Prediction Method: Recursive Strategy, which involves predicting the next month's forecast each time, and then using the predicted value as part of the historical input to continue predicting the next month.
[0085] The forecast for January to December 2025 (12 months) will be conducted in two phases: first, forecast January to June, and then use the forecast values for January to June (along with actual historical data) to forecast July to December.
[0086] Prediction results (normalized and then denormalized back to the original scale):
[0087] 6.3 Trend Analysis During the spring (March-May), NDVI remained between 0.58 and 0.64, still within the normal range.
[0088] Starting in July 2025, NDVI began to decline significantly, dropping to 0.45 in September and 0.42 in December.
[0089] Compared with the same period in history (the average value of 0.69 during the healthy period of 7-12 months from 2005 to 2010), the decrease is 35% to 39%.
[0090] Step 7: Critical Warning Judgment 7.1 Degradation rate and acceleration calculation Degradation rate (Monthly changes); Degradation acceleration
[0091] Calculation results (July-December 2025):
[0092] The maximum degradation rate occurred in July and August: -0.04 / month.
[0093] 7.2 Setting the Critical Degradation Threshold Historical health reference period: 2005-2010 (the recognized healthy and stable period for the Yangtze River Estuary wetlands).
[0094] Average NDVI during the same period (July to December) Standard deviation Critical degradation threshold Validation of alternative thresholds: If mutation point detection (Pettitt test) is used, September 2013 was identified as the first significant mutation point in the historical sequence. The mean NDVI value for the six months prior to this point was 0.44, which is highly consistent with 0.45.
[0095] 7.3 Statistical Analysis of Historical Degradation Rates Calculate the historical degradation rate (monthly variation) for all months from 2000 to 2020, and obtain: mean / month, showing a slight downward trend overall, standard deviation / month. Threshold is (Note that this is a comparison of absolute rates.)
[0096] 7.4 Early Warning Logic Judgment
[0097] 7.5 Warning Level Output July 2025: NDVI did not fall below the threshold, so no warning was triggered.
[0098] August 2025: NDVI (0.48) is still not below the threshold (0.45), and although the degradation rate is fast, it does not trigger.
[0099] September 2025: NDVI (0.45) equals the threshold, degradation rate -0.01 / month, but acceleration is positive (deceleration), and the rate does not exceed [the threshold value]. Threshold → Triggering Level 1 Warning (Note): NDVI has reached the threshold within the next 3 months, but the rate of decline is slowing down. It is recommended to strengthen monitoring.
[0100] October 2025 and beyond (assuming the forecast extends to December): NDVI will remain below 0.45, but the degradation rate will drop to -0.01 / month, with zero acceleration. The Level 1 warning will be maintained, and no upgrade will be implemented.
[0101] Step 8: System Deployment and Automated Operation 8.1 System Architecture Data layer: Google Earth Engine monthly NDVI extraction script (Python API).
[0102] Model layer: The trained BiLSTM-Attention model (SavedModel format).
[0103] Application layer: Flask Web service + alert push module.
[0104] Database: SQLite stores historical NDVI data, prediction results, and alert records.
[0105] 8.2 Automated Processes Automatically triggered on the 5th of each month (after the previous month's data is fully available): Obtain the latest month's NDVI from GEE and append it to the local time series.
[0106] Using data from the most recent 24 months as input, the model is called to predict the next 12 months.
[0107] Calculate degradation indicators and determine the warning level.
[0108] If the warning level is level 1 or higher, a warning notification will be sent via WeChat / email, along with an attention weight heatmap and a prediction curve.
[0109] Model retraining: The model is retrained and the weights are updated every quarter (March, June, September, and December) using all available data.
Claims
1. An AI-based early warning method for the risk of estuarine wetland degradation, characterized in that, Includes the following steps: Step 1: Acquire long-term MODIS NDVI remote sensing data of the target estuarine wetland area (taking the Yangtze River estuary as an example) as the sole data source; Step 2: Preprocess the acquired MODIS NDVI data, including cloud masking, time series reconstruction, and outlier removal, to construct a continuous NDVI time series dataset; Step 3: Based on the preprocessed NDVI time series, construct a training sample set. Each sample contains a continuous historical NDVI sequence as input and the corresponding future NDVI sequence as the label. Step 4: Construct a deep learning prediction model, which is a bidirectional long short-term memory network (BiLSTM). An attention mechanism is introduced after the BiLSTM layer to adaptively focus on the key time step features in the NDVI time series and learn the nonlinear evolution law. Step 5: Train the deep learning model using the training sample set, with the goal of minimizing the error between the predicted value and the true value; Step 6: Use the trained model for multi-step prediction of future NDVI change trends to obtain the prediction curve of wetland vegetation cover in the future time period. Step 7: Based on the predicted NDVI sequence, calculate the degradation rate and acceleration index. When the predicted NDVI value continues to decrease and falls below the preset critical degradation threshold, issue a critical warning signal.
2. The method according to claim 1, characterized in that, In step two, the time series reconstruction uses SG filtering or linear interpolation to fill in the missing values caused by cloud cover.
3. The method according to claim 1, characterized in that, In step four, the BiLSTM model structure includes: an input layer, a bidirectional LSTM layer (each containing 32-128 hidden units in the forward and backward directions), an attention layer, a dropout layer, a fully connected output layer, and a time step set to 12-36 months.
4. The method according to claim 3, characterized in that, The attention mechanism is additive attention or scaled dot product attention, and its weights are normalized by the Softmax function and used to perform a weighted summation of the hidden states at each time step of the BiLSTM output.
5. The method according to claim 1, characterized in that, The multi-step prediction in step six adopts a recursive prediction or direct multi-step prediction strategy, and outputs the monthly NDVI values for the next 6 to 24 months.
6. The method according to claim 1, characterized in that, The critical degradation threshold in step seven is set to 60% to 70% of the average NDVI of historical normal years, or the degradation inflection point in the historical sequence can be automatically identified as a reference threshold through a mutation point detection algorithm (such as BEAST or Pettitt test).
7. The method according to claim 1, characterized in that, The critical warning signal is divided into three levels: Level 1 Warning (Attention): The predicted NDVI value is below the threshold but the rate of decline is slow within the next 3 months; Level 2 Warning: The predicted value is below the threshold and the degradation acceleration is negative within the next 6 months; Level 3 warning (critical): The predicted value remains below the threshold for the next 12 months, and the rate of degradation exceeds twice the standard deviation of the historical average rate.
8. A system for predicting and providing critical early warning of estuarine wetland degradation trends based on single-index deep learning, characterized in that, include: Data acquisition module: used to acquire MODIS NDVI time series data of the Yangtze River Estuary wetlands; Preprocessing module: used for quality control and sequence reconstruction of NDVI data; Model training module: Used to train the BiLSTM-Attention deep learning model; Trend prediction module: Used to output predictions of future NDVI changes; Degradation detection and early warning module: used to calculate degradation indicators and determine whether a critical early warning is triggered.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 7.