Method for intelligently monitoring coastal water eutrophication

Abstract

The invention belongs to the technical field of marine environment monitoring and artificial intelligence, and discloses a method for intelligently monitoring coastal water eutrophication. The method comprises the following steps: acquiring and matching spectrum, space and time parameters of a remote sensing image and actually measured data; constructing a multi-modal feature data set; a hybrid prediction model formed by cascading a Transform network and a support vector regression model is constructed and trained, and the Transform network performs unified modeling and deeply fuses multi-source features such as spectrum, space and time through a self-attention mechanism of the Transform network, so that deep intelligent fusion of the multi-source features is realized; performing water quality index prediction and eutrophication state evaluation by using the trained model; and finally, carrying out uncertainty quantization on a prediction result by adopting a Monte Carlo random inactivation technology. According to the invention, the problems of low precision and weak generalization caused by spatial heterogeneity and complex optical characteristics in coastal water monitoring are effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of marine environmental monitoring and artificial intelligence technology, and discloses a method for intelligent monitoring of eutrophication in coastal waters. More specifically, it relates to an intelligent monitoring method that integrates multi-source remote sensing data and deep learning for assessing the eutrophication status and quantifying the uncertainty of coastal waters. Background Technology

[0002] Eutrophication in nearshore waters is a major global environmental problem, leading to frequent harmful algal blooms (red tides), oxygen depletion in water bodies, and other ecological disasters that seriously threaten the health and sustainable development of marine ecosystems. Therefore, achieving large-scale, high-precision, and dynamic monitoring of eutrophication in nearshore waters is of paramount importance for environmental protection, disaster early warning, and policy formulation.

[0003] Traditional water quality monitoring methods rely on on-site sampling and laboratory analysis. While highly accurate, these methods suffer from inherent drawbacks such as high cost, poor timeliness, and discontinuous spatial coverage, making them unsuitable for macroscopic dynamic monitoring. Remote sensing technology, with its advantages of wide-area, rapid, and periodic observation, has become an important tool for eutrophication monitoring. However, water quality inversion based on remote sensing, especially in nearshore waters, still faces significant challenges:

[0004] 1. Complex Optical Components and Spatial Heterogeneity: Nearshore waters are influenced by various factors, including terrestrial input, human activities (shipping, aquaculture), and hydrodynamics. These factors result in high concentrations of optically active substances such as suspended solids, chlorophyll, and colored soluble organic matter (CDOM), exhibiting significant spatiotemporal variability. This leads to a highly nonlinear and spatially heterogeneous relationship between water quality parameters and remote sensing spectral signals. Specifically, the same eutrophication level may correspond to completely different spectral characteristics in different regions or at different times, greatly increasing the difficulty of remote sensing inversion.

[0005] 2. Limitations of traditional models: Existing remote sensing inversion models are mostly based on empirical or semi-empirical formulas, estimating water quality parameters by constructing simple band combinations (such as ratios and differences). These models have clear physical meanings, but they are poorly adaptable to water bodies with complex and variable optical characteristics, have weak generalization ability, and often make misjudgments of "different spectra for the same object" or "same spectra for different objects," resulting in poor monitoring accuracy.

[0006] 3. Bottlenecks in Deep Learning Applications: Although recent studies have attempted to use deep learning models (such as CNNs and RNNs) for water quality inversion, most have simply concatenated multi-source features (spectral, temporal, and spatial) before inputting them into the network, failing to effectively handle the inherent correlations and heterogeneity between different modalities. In particular, conventional deep learning models (such as the standard Transformer), as general feature extractors, have not undergone adaptive structural design to address the core challenges of "spatial heterogeneity" and "multi-factor coupling" in water quality remote sensing, resulting in their potential not being fully realized. Furthermore, existing remote sensing monitoring methods typically only provide a single predicted value for water quality indicators, lacking assessment of the reliability of the prediction results. This may introduce uncertainty risks in environmental decision-making, representing a major limitation of current technology.

[0007] Therefore, there is an urgent need for an innovative method that can intelligently integrate multi-source information, effectively overcome the effects of spatial heterogeneity, and provide reliability assessments to improve the accuracy and practicality of remote sensing monitoring of eutrophication in nearshore waters and provide more comprehensive information for decision-making. Summary of the Invention

[0008] To address the aforementioned problems, this invention discloses a method for intelligent monitoring of eutrophication in coastal waters, aiming to solve the issues of low accuracy, weak model generalization ability, and inability to assess the reliability of prediction results in complex nearshore water bodies using remote sensing for eutrophication monitoring.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for intelligent monitoring of eutrophication in coastal waters includes the following steps:

[0011] S1. Data Acquisition and Preprocessing: Based on remote sensing images and measured water quality data, extract and match the spectral reflectance, spatial parameters and temporal parameters of the sampling points;

[0012] S2. Construct a multidimensional feature dataset: Transform the spectral reflectance obtained in step S1 to generate diverse spectral features, and combine spatial and temporal parameters to construct a water quality inversion dataset containing three modes: spectral, spatial, and temporal.

[0013] S3. Model Construction and Training: Construct and train a hybrid prediction model consisting of a Transformer network and a Support Vector Regression (SVR) model cascaded together. The Transformer network is used to perform deep feature fusion on the dataset constructed in step S2, and the SVR model is used to predict water quality indicators based on the fused feature vector output by the Transformer network.

[0014] S4. Eutrophication status monitoring: The trained hybrid prediction model is used to predict the water quality indicators of the target sea area, and the eutrophication index EI is calculated based on the prediction results to realize the monitoring of the eutrophication status of the sea area.

[0015] S5. Uncertainty Quantification Assessment: When using the hybrid prediction model for prediction, the uncertainty of the prediction results is quantitatively assessed to generate an uncertainty measure of the predicted water quality index value.

[0016] Furthermore, in the above method, in step S1, the spatial parameters include longitude, latitude, and water depth; the time parameters include the number of days in a year, and the number of days in a year is subjected to periodic sine and cosine transformation processing.

[0017] Furthermore, in the above method, the Transformer network in step S3 includes:

[0018] The Transformer is primarily used to extract high-dimensional feature representations from input data. Its structure includes input embedding, positional encoding, multi-head attention (MHA) layers, residual connections and layer normalization (Add & Norm), and a feedforward network (FFN). Furthermore, in the above method, the Transformer encoder, through its internal MHA layers, performs global interaction and dynamic weighting of the spectral, spatial, and temporal features, achieving deep intelligent fusion of multi-source features.

[0019] Furthermore, in the above method, in step S4, the eutrophication index EI is determined based on the predicted values ​​of chemical oxygen demand (COD), dissolved inorganic nitrogen (DIN), and soluble reactive phosphorus (SRP) using the formula... Calculated.

[0020] Furthermore, in the above method, the uncertainty quantification assessment in step S5 adopts the Monte Carlo random deactivation method, including: in the prediction stage, the random deactivation layer in the Transformer network is activated multiple times for forward propagation to obtain a set of prediction results output by the SVR model; by analyzing the statistical distribution of the set of prediction results, the uncertainty measure corresponding to each prediction value is calculated.

[0021] This invention also discloses a device for intelligent monitoring of eutrophication in coastal waters, comprising:

[0022] The data preprocessing module is used to perform data acquisition, preprocessing, and construction of multidimensional feature datasets in steps S1 and S2 above.

[0023] A hybrid prediction model, which is a model trained according to any of the methods described above, consisting of a cascaded Transformer network and a support vector regression model.

[0024] The monitoring and evaluation module is used to load the hybrid prediction model, perform the eutrophication status monitoring and uncertainty quantification evaluation in steps S4 and S5 above, and output the monitoring results and corresponding uncertainty information.

[0025] Furthermore, the aforementioned device also includes:

[0026] The visualization module is used to generate a spatial distribution map of the eutrophication index and a corresponding spatial distribution map of uncertainty from the monitoring results and uncertainty information output by the monitoring and evaluation module.

[0027] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps described above in the intelligent monitoring method for eutrophication of coastal waters.

[0028] The beneficial effects of this invention are:

[0029] 1. Significantly Improved Monitoring Accuracy and Robustness: Through the designed Transformer network, this invention upgrades from simple feature assembly and processing to deep understanding and intelligent fusion of multi-source features. Comparative experiments demonstrate that the prediction accuracy of the method in this invention is significantly superior to existing technologies. For example, in the prediction of Chemical Oxygen Demand (COD), the root mean square error (RMSE) of the method in this invention is reduced by approximately 30% and 15% compared to traditional empirical models and general deep learning models, respectively, demonstrating superior accuracy and generalization ability.

[0030] 2. The model design has clear physical interpretability: Each module of the Transformer network is designed to solve a specific physical or real-world problem, which makes the complex deep learning model no longer a complete "black box", enhancing the model's interpretability, reliability, and credibility in real-world business applications.

[0031] 3. Providing decision-level reliability assessment and creating new application value: This invention introduces uncertainty quantification assessment into the eutrophication remote sensing monitoring process for the first time. This not only provides predicted values ​​for eutrophication levels but also gives the "credibility" of the prediction results, highlighting areas where the model is uncertain. This has extremely high application value and can directly empower environmental management decisions. For example, environmental management departments can use the generated uncertainty map to prioritize the deployment of on-site sampling resources for the next quarter in areas with high uncertainty, thereby achieving dynamic optimization of the monitoring plan and greatly improving monitoring efficiency and cost-effectiveness within a limited budget. Attached Figure Description

[0032] Figure 1 A flowchart of a method for intelligent monitoring of eutrophication in coastal waters provided in an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of the structure of the Transformer network, a core component of an embodiment of the present invention.

[0034] Figure 3 This is the loss curve of the Transformer-SVR hybrid model in an embodiment of the present invention;

[0035] Figure 4 This is a visual comparison chart of the performance of the three models in the embodiments of the present invention;

[0036] Figure 5 This is a visualization diagram of the prediction results in an embodiment of the present invention. Detailed Implementation

[0037] A method for intelligent monitoring of eutrophication in coastal waters includes the following steps:

[0038] S1. Data Acquisition and Preprocessing: Based on remote sensing images such as MODIS and historical water quality measurement data, the spectral reflectance, spatial parameters (longitude, latitude, water depth) and temporal parameters (number of days and months in the year) of the sampling points are extracted through spatiotemporal matching.

[0039] S2. Constructing a multidimensional feature dataset: Perform single-band transformation and multi-band combination operations on the extracted spectral reflectance to generate diverse spectral features. Combine these spectral features with spatial parameters and periodically encoded temporal parameters to construct a multidimensional water quality inversion dataset containing three independent modes: spectral, spatial, and temporal.

[0040] S3. Model Building and Training: Build an innovative hybrid prediction model.

[0041] The model consists of a Transformer network and a Support Vector Regression (SVR) model cascaded together. The Transformer network is responsible for deep feature learning and fusion of the multidimensional dataset, while the SVR model uses the high-dimensional fused features output by the Transformer network to perform the final regression prediction of water quality indicators (such as COD, DIN, and SRP).

[0042] S4. Eutrophication Status Monitoring: The trained hybrid prediction model is applied to remote sensing images of the target sea area to generate a spatial distribution map of water quality indicators. Based on the national seawater quality standards, the eutrophication level of the sea area is calculated and evaluated using the eutrophication index (EI) formula.

[0043] S5. Uncertainty Quantification Assessment and Decision Support: To assess the reliability of the prediction results, this invention also includes an uncertainty quantification step. Using Monte Carlo Dropout technology, the model's prediction results are randomly sampled multiple times. By analyzing the statistical distribution (such as standard deviation) of the sampled results, a confidence level or uncertainty map is generated for each predicted pixel, thereby elevating the monitoring results from a single numerical value to decision support information containing reliability information.

[0044] Further, step S1 includes the following sub-steps:

[0045] S11. Use the quartile method to remove outliers from the measured values ​​of the sampling points;

[0046] S12. Retrieve MODOCGA V6 product data corresponding to the date and location of each water quality sample;

[0047] S13. The extracted information includes reflectance values ​​of bands 8 to 16 of the MODOCGA V6 product and water depth at each sample location;

[0048] S14. A periodic time parameter was introduced to more accurately capture the periodic characteristics of water quality changes over time.

[0049] Furthermore, outliers removed in step S11 include measurement errors, data entry errors, or actual extreme cases.

[0050] Furthermore, in step S12, the remote sensing image data corresponding to the date and location of each water quality sample was retrieved using a "same-day matching" strategy, which paired each water quality sampling date with the corresponding MODIS satellite acquisition date; in terms of spatial matching, the nearest neighbor pixel matching method was used to match within a 1-kilometer resolution of the sampling point.

[0051] Furthermore, in step S13, quality control (QC) variables from the MODOCGA product were used to exclude satellite imagery affected by cloud contamination. Based on sample coordinates, multidimensional information was extracted from the MODOCGA V6 image, which had been processed to a uniform resolution, and the GEBCO bathymetry data.

[0052] Furthermore, the time parameters in step S14 include the periodic month (cMOY) in a year, the periodic day (cDOW) in a week, and the periodic day (cDOY) in a year.

[0053] Furthermore, step S2 includes the following sub-steps:

[0054] S21. Based on the band reflectance of the sampling points extracted from the acquired remote sensing images, a series of band combinations and spectral indices were constructed.

[0055] S22. Based on the obtained band combinations, spatial parameters, and temporal parameters of the sampling points, as well as the corresponding periodic transformations, construct a long-term series, multi-dimensional water quality inversion dataset.

[0056] Furthermore, in step S21, the band combination and spectral index include single-band transformations (such as reciprocal, square), two-band combinations (such as summation, ratio, difference), and three-band interaction terms.

[0057] Furthermore, in step S22, the spatial parameters include water depth, longitude, and latitude, and the time parameters include the number of days in the year (DOY), the number of months (MOY), the number of days of the week (DOW), and the corresponding periodic transformations (cDOY, cMOY, and cDOW). The periodic time parameters are converted using the following formula:

[0058] ;

[0059] Where T represents cMOY, cDOW, and cDOY respectively, t represents MOY, DOW, and DOY respectively, and N_T represents the complete cycle length, i.e. 12 (months in a year), 7 (days in a week), or 365 (days in a year).

[0060] Furthermore, step S3 includes the following sub-steps:

[0061] S31. Extract features from the original input data using a Transformer network;

[0062] S32. Input the feature vectors extracted by the Transformer into the SVR model, and use the robustness and generalization ability of SVR to complete the prediction of the target variable (water quality index).

[0063] Furthermore, the Transformer encoder in step S31 employs a lightweight configuration, specifically stacking two encoder blocks, each containing four attention heads with an embedding dimension of 60. A dropout rate of 0.1 is added after each ReLU activation layer to suppress overfitting. Model training uses the Adam optimizer with an initial learning rate of 0.001, combined with cosine annealing scheduling and early stopping (with a tolerance of 20 epochs).

[0064] Furthermore, the high-dimensional features extracted by the Transformer in step S32 are used as the input vector of the SVR, and the Radial Basis Function (RBF) is used as the kernel function. The key hyperparameters are set as follows: penalty coefficient C = 100, kernel width parameter γ = 0.1, and ε-insensitive loss threshold ε = 0.01.

[0065] Furthermore, step S4 includes the following sub-steps:

[0066] S41. Calculate the eutrophication index (EI) of the water body using the formula;

[0067] S42. Quantitatively evaluate the nutrient status of nearshore marine waters using the eutrophication index (EI).

[0068] Furthermore, in step S41, the construction of the EI (Environmental Information) refers to the "Specifications for Environmental Monitoring of Nearshore Marine Areas" (Ministry of Ecology and Environment of the People's Republic of China, https: / / www.mee.gov.cn / ), and a comprehensive weighted calculation is performed on the main nutrients in seawater—chemical oxygen demand (COD), dissolved inorganic nitrogen (DIN), and soluble reactive phosphorus (SRP). The expression is as follows:

[0069] ;

[0070] Where COD, DIN, and SRP are all in mg / L, 4500 is a standardization constant, and 10 6 This is an exponential amplification factor used to unify the dimensions and result range.

[0071] Furthermore, in step S42, eutrophication is quantitatively assessed using the Eutrophication Index (EI). When EI ≥ 1, the sea area is determined to be in a eutrophic state, and the higher the EI value, the more severe the eutrophication.

[0072] Furthermore, step S5 includes the following sub-steps:

[0073] S51. In the prediction phase, enable the random deactivation layer of the Transformer network in the hybrid prediction model and perform Monte Carlo random deactivation.

[0074] S52. Perform multiple random forward propagation samplings on each pixel to be predicted in the target sea area to obtain a set of predicted values;

[0075] S53. Based on the set of predicted values ​​obtained from sampling, calculate the final predicted value of each pixel and its corresponding uncertainty measure;

[0076] S54. Based on the uncertainty measure of all pixels, generate an uncertainty spatial distribution map covering the target sea area for decision support.

[0077] Furthermore, when random deactivation is enabled in step S51, the deactivation rate set during the network training phase remains unchanged, and the random deactivation rate after each feedforward network layer in the Transformer network is 0.1.

[0078] Furthermore, the number N of multiple random forward propagation samplings in step S52 is set to 50 times, which is determined based on a comprehensive trade-off between the stability of uncertainty estimation and computational efficiency.

[0079] Furthermore, the final predicted value in step S53 is the arithmetic mean of the set of predicted values, and the uncertainty measure is the standard deviation of the set of predicted values.

[0080] Furthermore, the uncertainty spatial distribution map generated in step S54 has pixel values ​​corresponding to the standard deviation calculated in step S53, and the spatial differences in prediction reliability are intuitively represented by color gradients.

[0081] Furthermore, the Transformer network, which is the core of this invention, has the following structure:

[0082] 1. A unified self-attention feature fusion framework: For three heterogeneous but related features—spectral, spatial, and temporal—this invention employs a unified framework based on the Transformer encoder for modeling. The design motivation stems from a deep understanding of the physical mechanisms of remote sensing inversion.

[0083] 1) Spectral Self-Attention Modeling: The multi-head self-attention mechanism in the framework takes all spectral features (original bands and their combinations) as part of the input sequence, automatically mining and quantifying the complex global dependencies between different bands. It can adaptively enhance the contribution of key spectral features that are sensitive to target indicators such as COD, DIN, and SRP, and achieve intelligent analysis of the optical response of water components.

[0084] 2) Spatial self-attention modeling: Spatial parameters such as latitude and longitude and water depth are input into the model as sequential elements on par with spectral features. The self-attention mechanism dynamically establishes the correlation between spatial location and all spectral features in global computation, thereby implicitly learning the unique "water quality-spectrum" mapping patterns of different geographical units (such as estuaries, bays, and open waters), fundamentally addressing the spatial heterogeneity of nearshore water bodies.

[0085] 3) Temporal self-attention modeling: Periodically encoded temporal features (such as annual days and months) are also incorporated into the input sequence. Through attention weight calculation, the model can naturally capture the rhythmic changes and long-term trends of water quality parameters with seasons and months, encode temporal dynamics into the feature representation, and understand the evolution of eutrophication.

[0086] 2. Intrinsic Cross-Modal Fusion Mechanism: The Transformer's self-attention computation itself constitutes a powerful and intrinsic fusion mechanism. In each layer of the encoder, each feature (whether spectral, spatial, or temporal) is reweighted and represented based on its correlation with all other features in the sequence. This is equivalent to dynamically performing intelligent fusion of "how to interpret and weight spectral signals in a specific spatiotemporal context (defined by latitude, longitude, water depth, and season)." This mechanism enables the model to adaptively adjust the interpretation of spectral features based on geographical location and time, effectively alleviating the classic problems of "different spectra for the same object" and "same spectra for different objects."

[0087] Example:

[0088] Reference Figure 1 The present invention provides an intelligent monitoring method for eutrophication in coastal water quality research, comprising the following steps:

[0089] S1. Data Acquisition and Preprocessing

[0090] This step is similar to existing technologies and mainly involves data collection and preparation.

[0091] (1) Data source: Remote sensing data was obtained from MODIS MODOCGA V6, which contains nine spectral bands in the range of 405-877nm. Measured water quality data were obtained from historical monitoring station water quality reports, including the latitude and longitude of the sampling point, sampling time, and concentration values ​​of indicators such as chemical oxygen demand (COD), dissolved inorganic nitrogen (DIN), and soluble reactive phosphorus (SRP). Water depth data were obtained from publicly available water depth products such as GEBCO.

[0092] (2) Spatiotemporal matching: The “same-day matching” strategy is adopted to pair the water quality sampling date with the MODIS image acquired on the same day. Spatially, the nearest neighbor pixel method is used to extract the spectral reflectance value and water depth value of the pixel where the geographic coordinates of the sampling point are located.

[0093] (3) Quality control: Remove low-quality remote sensing pixels contaminated by clouds, solar flares, etc., and use the quartile method to remove water quality samples with missing or obvious abnormalities in the data records.

[0094] S2. Constructing a multidimensional feature dataset

[0095] (1) Spectral Feature Engineering: To enhance the representational ability of spectral information, various transformations were performed on the reflectance of the nine original bands extracted from S1 to construct a rich candidate spectral feature library. These included single-band transformation, two-band combination, and three-band interaction terms. The calculation formulas are shown in Table 1.

[0096]

[0097] (2) Spatiotemporal characteristic engineering: Spatial parameters directly use longitude, latitude, and water depth. For time parameters, such as the number of days in a year (DOY) and the number of months (MOY), in order to avoid the problem of numerical jumps at the end of the year and the beginning of the year, periodic encoding conversion is required. For example, sine and cosine functions are used for transformation: cDOY=sin(2π ·DOY / 365), cDOY=cos(2π ·DOY / 365).

[0098] (3) Dataset construction: The generated spectral features, spatial parameters and encoded temporal parameters are integrated to form a multidimensional feature matrix. Each row represents a spatiotemporally matched sample point, each column represents a feature, and the corresponding measured water quality index values ​​(COD, DIN, SRP) are associated with it to form the final dataset for model training and validation.

[0099] S3, Model Building and Training

[0100] This step is the core of the invention.

[0101] Reference Figure 2 We construct a hybrid model consisting of a Transformer network and an SVR cascade.

[0102] (1) Transformer network structure:

[0103] 1. Input Layer: Input data first passes through an embedding layer, mapping the low-dimensional raw input to a high-dimensional feature space, injecting temporal information through positional encoding. 2. Transformer Encoder Structure: A lightweight configuration is adopted, specifically stacking two encoder blocks, each containing four attention heads with an embedding dimension of 60. A dropout rate of 0.1 is added after each ReLU activation layer to suppress overfitting. Model training uses the Adam optimizer with an initial learning rate of 0.001, combined with cosine annealing and early stopping (with a tolerance of 20 epochs). Training automatically terminates when the validation set loss does not significantly decrease for 20 consecutive epochs, saving the optimal weights and learning the feature dependencies within each modality.

[0104] 3. Cross-modal fusion module: The data then enters a feedforward neural network (FFN) for non-linear mapping. The FFN consists of two fully connected layers and a ReLU activation function, which improves the model's feature transformation capability. 4. Output layer: The fused vector then passes through another feedforward neural network (FFN) and layer normalization (LayerNorm), ultimately outputting a high-dimensional, deeply fused feature vector.

[0105] (2) SVR Model: An SVR model is trained by using the feature vector output by the Transformer network as input. The high-dimensional features extracted by the Transformer are used as the input vector of the SVR, and the Radial Basis Function (RBF) is used as the kernel function. The key hyperparameters are set as follows: penalty coefficient C = 100, kernel width parameter γ = 0.1, and ε-insensitive loss threshold ε = 0.01. Compared with the traditional single SVR model, combining it with the Transformer can make full use of its deep feature extraction capabilities, thereby improving prediction accuracy and generalization ability. Three independent hybrid models can be trained for the three indicators COD, DIN, and SRP.

[0106] (3) Model Training and Comparative Validation: The dataset is divided into training, validation, and test sets. The entire Transformer-SVR model is trained end-to-end using the training set. Specifically, the hybrid prediction model adopts a joint training strategy: the Transformer network and SVR are connected into a unified computational graph, and the mean squared error (MSE) between the predicted and actual water quality indicators output by the model is used as the loss function. The Adam optimizer is used to update all trainable parameters in the Transformer network and SVR at once through the backpropagation algorithm. To verify the superiority of this invention, the following comparative experiments are set up:

[0107] 1. Raw Data: The target inversion area is the nearshore waters of Qingdao (119.51°-121.56°E, 35.08°-36.63°N), which is a typical semi-enclosed bay system on the western coast of the Yellow Sea. The study area centers on Jiaozhou Bay, bordering Dingzi Bay to the north and Langya Bay to the south, with a total coastline of 1041 kilometers and a sea area of ​​2.2 × 10⁻⁶. 4 km 2 Influenced by the continental shelf topography of the Jiaodong Peninsula, a complex geomorphological system of "bay-cape-tidal channel" has formed. The tidal channel between Xuejiadao and Tuandao (maximum depth 64m, (Wang Wenhai et al., 1993)) constitutes a natural barrier to water exchange, limiting the diffusion capacity of pollutants such as COD and DIN. Seasonal rivers along the coast, such as the Dagu River, and industrial point sources form a dual-source input system, transporting 7 × 10⁻⁶ DIN annually. 3 Tons, SRP 1.6×10 2 The area contains tons of eutrophication, with industrial point sources contributing 33.87% of DIN and 23.22% of SRP (Song Jinming et al., 2020). This coupling effect of high-intensity terrestrial input and weak diffusion capacity makes this sea area an ideal region for studying the spatiotemporal evolution of eutrophication. The water quality data for this study were obtained from the National Marine Environmental Monitoring Center (NMEMC; https: / / www.nmemc.org.cn / ), including concentrations of Chemical Oxygen Demand (COD), Dissolved Inorganic Nitrogen (DIN), and Soluble Reactive Phosphorus (SRP) from 2017 to 2022. Due to the monitoring station's closure during winter, the data covers the period from April to November each year, spatially spanning 117.67°E-124.50°E and 35.00°N-40.84°N, with a total of 8129 raw data entries. Table 2 presents the statistical indicators for COD, DIN, and SRP parameters, including their maximum (Max), minimum (Min), mean (Mean), standard deviation (Std), and coefficient of variation (CV).

[0108]

[0109] 2. Model 1 (This Invention): Transformer-SVR Hybrid Model. After... Figure 1 After the process shown, the loss curve of the Transformer-SVR hybrid model is as follows: Figure 3 As shown, epoch represents the number of training iterations.

[0110] 3. Model 2 (Machine Learning Model): Based on a random forest regression model, this model automatically captures the complex nonlinear relationships between water quality parameters and multi-band reflectance and derived features by integrating multiple decision trees. Specifically, it first constructs an initial spectral feature set based on the visible to near-infrared reflectance of the sensor, including the original bands, ratios, and normalized differences; then, these features are input into the random forest model for training. The model constructs multiple decision trees through bootstrap sampling. When each tree splits at a node, it selects the optimal feature from a random subset to maximize information gain or reduce variance. Finally, it outputs concentration estimates of COD, DIN, or SRP by integrating the prediction results of all trees (e.g., taking the average), and evaluates its performance using independent test sets, with metrics including RMSE and R².

[0111] 4. Model 3 (General Deep Learning Model): This model uses all spectral bands, spatial location, and temporal information of each pixel in the sea area as input features. These are simply concatenated to form a one-dimensional feature vector, which is then input into a one-dimensional convolutional neural network (1D-CNN) for regression prediction. First, the input features are normalized to eliminate dimensional differences and scale effects. Then, features are extracted through multiple one-dimensional convolutional layers, with the convolution kernels sliding along the feature dimension to capture feature combination patterns within adjacent or local areas (such as the correlation between several consecutive bands). Pooling layers are then used to reduce dimensionality and enhance feature robustness. Finally, fully connected layers map the learned features to continuous predicted values ​​of COD, DIN, or SRP. The model is trained end-to-end, minimizing the mean squared error (MSE) loss function, and its performance (e.g., RMSE, R2, MAE) is evaluated on training, validation, and test sets.

[0112] 5. Performance Comparison: On the test set, root mean square error (RMSE) and coefficient of determination (R²) were used. 2 The mean absolute error (MAE), mean squared error (MSE), and mean absolute percentage error (MAPE) were used as evaluation indicators. The experimental results (taking COD prediction as an example) are shown in Table 3. Figure 4 A visual comparison chart of the performance of the three models is shown, indicating that the present invention is significantly superior to the comparative models in all indicators.

[0113]

[0114] S4, Eutrophication Status Monitoring

[0115] Using the trained model, by inputting the spectral, spatial, and temporal characteristics of any pixel in the target sea area, the concentrations of COD, DIN, and SRP at that point can be predicted. Subsequently, the eutrophication index (EI) is calculated according to the following formula:

[0116] ;

[0117] The units for COD, DIN, and SRP are all mg / L. Marine areas can be evaluated and classified based on EI values ​​(e.g., EI ≥ 1 indicates eutrophication), and spatial distribution maps can be generated.

[0118] S5, Quantitative Assessment of Uncertainty

[0119] During the training and prediction process of the Transformer, the Dropout layer is enabled. For a cell to be predicted, while keeping the SVR model parameters fixed during the prediction phase, the following operations are performed:

[0120] 1. Perform N forward propagation predictions (e.g., N=50). In each propagation, due to the Dropout layer in the Transformer network randomly deactivating some neurons, the network structure will have slight differences, resulting in the SVR model receiving N slightly different high-dimensional feature vectors and ultimately obtaining N slightly different predicted values ​​{y1,y2,...,yN}.

[0121] 2. Final prediction value: Take the average of these N prediction values ​​as the most likely predicted concentration of the pixel.

[0122] 3. Uncertainty: Calculate the standard deviation σ of these N predicted values. The larger the σ value, the more unstable the model's prediction of that point, and the higher the uncertainty. Finally, an "uncertainty map" corresponding to the prediction results map can be generated. The highlighted areas in the map are areas where the model is unsure, requiring close attention or on-site verification.

[0123] 4. Visualization of prediction results (taking COD prediction as an example).

[0124] Figure 5 The graph shows the COD predictions of the Transformer-SVR hybrid model over a period of time. It can be seen from the graph that the Transformer-SVR hybrid model has the highest prediction accuracy and the best fit, followed by the general deep learning model, while the machine learning model has the worst prediction performance.

[0125] In summary, the above embodiments used the coastal waters of Qingdao as the verification area, integrating MODIS remote sensing imagery, measured water quality and depth data to construct a multimodal dataset containing spectral features, spatial location, and periodic temporal encoding. A hybrid model consisting of a Transformer network and Support Vector Regression (SVR) cascaded together was trained to retrieve water quality indicators. The Transformer network, through its self-attention mechanism, uniformly models and deeply fuses multi-source features such as spectral, spatial, and temporal data. A joint training strategy was employed to optimize the model, and comparative experiments demonstrated that its accuracy significantly outperformed traditional empirical models and general deep learning models. Finally, the model not only outputs the spatial distribution of eutrophication status but also generates an uncertainty quantification map of the prediction results using Monte Carlo random deactivation technology, achieving a dual improvement in monitoring accuracy and decision reliability.

[0126] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for intelligent monitoring of coastal water eutrophication, characterized in that, The method comprises the following steps: S1, data acquisition and preprocessing: based on remote sensing images and water quality measured data, the spectral reflectance, spatial parameters and time parameters of the sampling points are extracted and matched; S2, constructing a multi-dimensional feature data set: transforming the spectral reflectance obtained in step S1 to generate diversified spectral features, and combining the spatial parameters and the time parameters to construct a water quality inversion data set containing three modalities of spectrum, space and time; S3, model construction and training: a hybrid prediction model composed of a Transformer network and a support vector regression (SVR) model is constructed and trained, wherein the Transformer network is used to perform deep feature fusion on the data set constructed in step S2, and the SVR model is used to predict water quality indicators according to the fusion feature vector output by the Transformer network; S4, eutrophication state monitoring: using the trained hybrid prediction model to predict the water quality indicators of the target sea area, and calculating the eutrophication index (EI) according to the prediction results to realize the monitoring of the eutrophication state of the sea area; S5, uncertainty quantification and evaluation: when using the hybrid prediction model for prediction, the uncertainty of the prediction results is quantified and evaluated, and the uncertainty measure of the predicted value of the water quality indicators is generated.

2. The method of claim 1, wherein, In step S1, the spatial parameters include longitude, latitude and water depth; the time parameters include the number of days in a year, and the number of days in a year is processed by periodic sine and cosine transformation.

3. The method of claim 1, wherein, The Transformer network in step S3 is used to extract high-dimensional features from input data, and its structure includes input embedding, positional encoding, multi-head self-attention layer, residual connection and layer normalization, and feed forward network.

4. The method of claim 3, wherein, The encoder of the Transformer network performs global interaction and dynamic weighting on spectral features, spatial features and time features through its internal multi-head self-attention layer, realizing deep intelligent fusion of multi-source features.

5. The method of claim 1, wherein, In step S4, the eutrophication index EI is calculated from the predicted values of chemical oxygen demand COD, dissolved inorganic nitrogen DIN and soluble reactive phosphorus SRP by the formula .

6. The method of claim 1, wherein, The uncertainty quantification and evaluation in step S5 adopts the Monte Carlo random inactivation method, which includes: in the prediction stage, enabling the random inactivation layer in the Transformer network to perform forward propagation, obtaining a group of prediction results output by the SVR model; by analyzing the statistical distribution of the group of prediction results, the uncertainty measure corresponding to each prediction value is calculated.

Images