A river and lake eutrophication state monitoring system

Abstract

This invention discloses a river and lake eutrophication status monitoring system, belonging to the field of intelligent water environment monitoring technology, including data acquisition, feature fusion, state inference, probability calculation, and map construction modules. The data acquisition module acquires multi-source heterogeneous raw sensing data of the monitored water area, forming multi-dimensional water body feature parameters. The feature fusion module inputs the parameters into a deep feature fusion network, outputting a high-dimensional fused feature vector through adaptive weighting and nonlinear mapping. An inference engine embedded with historical cases trains a causal graph model, activating model nodes and paths based on high-dimensional features, calculating the simulated concentrations of total phosphorus and total nitrogen, and the posterior probability distribution of the comprehensive trophic state index, ultimately constructing a spatiotemporal evolution map of eutrophication. This system can fully explore the inherent correlations of water body characteristics, accurately infer index changes based on probabilistic causal logic, and completely restore the spatiotemporal evolution of eutrophication in water areas, making the monitoring results more consistent with the actual water body state.

Description

Technical Field

[0001] This invention belongs to the field of intelligent water environment monitoring technology, specifically a river and lake eutrophication status monitoring system. Background Technology

[0002] Existing monitoring methods for eutrophication status in rivers and lakes mostly rely on multi-sensor data collection to acquire water body perception data. The multidimensional water body characteristic parameters are processed using only superficial methods such as direct splicing and fixed weighting. Eutrophication status determination often depends on empirical formula calculation, single index threshold judgment, or pure data fitting models without logical connection. The results of total phosphorus, total nitrogen, and comprehensive trophic status index are obtained through conventional calculation methods.

[0003] Conventional monitoring methods are insufficient for in-depth analysis of multi-source, heterogeneous water body characteristic data. Feature information is prone to redundancy, omissions, or distortion, and the intrinsic relationships between different dimensions of water body characteristics cannot be effectively identified. The reasoning process for eutrophication status fails to establish probabilistic causal relationships between multidimensional causes and eutrophication characteristics. The calculated index results do not match the actual evolution of eutrophication in water bodies, failing to accurately reflect the dynamic changes of indicators or to form a complete spatiotemporal representation of eutrophication changes.

[0004] To address the issues of insufficient fusion of multi-source water features and lack of causal logic in state reasoning, it is necessary to conduct deep fusion processing of multi-dimensional water feature parameters through adaptive weighting and nonlinear mapping. By using a causal graph model trained with historical cases, the posterior probability calculation of core eutrophication indicators can be completed through the activation of model nodes and paths, thereby constructing a spatiotemporal evolution map of eutrophication in the monitored water area. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a river and lake eutrophication status monitoring system, comprising: The data acquisition module is used to acquire raw sensing data from multiple sources in the monitored water area, forming a set of multi-dimensional water body characteristic parameters; The feature fusion module inputs the multidimensional water body feature parameters into a deep feature fusion network, which outputs a high-dimensional fused feature vector through adaptive weighting and nonlinear mapping. The state reasoning module feeds the high-dimensional fused feature vector into the eutrophication state reasoning engine. The eutrophication state reasoning engine embeds a causal graph model trained based on historical cases. The causal graph model defines the probabilistic causal relationship between multidimensional causes and eutrophication representation. The probability calculation module, the eutrophication state inference engine activates relevant nodes and paths in the causal graph model based on the high-dimensional fusion feature vector, and calculates a set of posterior probability distributions of core eutrophication indicators, the core eutrophication indicators including at least total phosphorus simulated concentration, total nitrogen simulated concentration and comprehensive nutrient state index. The map construction module constructs a spatiotemporal evolution map of eutrophication in the monitored water area based on the posterior probability distribution.

[0006] Furthermore, the acquisition of multi-source heterogeneous raw sensing data of the monitored water area constitutes a set of multi-dimensional water body characteristic parameters, including: An underwater sensor array deployed in the monitored water area acquires multi-source heterogeneous raw sensing data, which includes at least chlorophyll fluorescence signal, blue-green algae fluorescence signal, dissolved oxygen concentration, turbidity, pH value, water temperature, and incident and emitted light intensity data at a specific wavelength. The original sensing data is subjected to synchronous acquisition and spatiotemporal alignment processing to generate a synchronous dataset with unified timestamps and matching spatial locations; The preliminary estimate of chlorophyll concentration, algal activity indicator, water optical attenuation coefficient, water acid-base state and thermodynamic state are extracted from the synchronous dataset to form a set of multidimensional water characteristic parameters. The process of performing synchronous acquisition and spatiotemporal alignment processing on the original sensing data to generate a synchronous dataset with unified timestamps and matching spatial locations includes: Each sensor in the underwater sensor array is equipped with a high-precision clock module, and the clock is periodically synchronized with a unified time server. In each data acquisition cycle, all sensors are controlled to complete data sampling within a preset synchronization time window, and the same timestamp is added to each sampled data. Obtain the real-time three-dimensional spatial coordinates of each sensor in the underwater sensor array; Based on the real-time three-dimensional spatial coordinates, the sampled data at different spatial locations under the same timestamp are interpolated and mapped to the nodes of the preset rule monitoring grid; All data from different sensors that are mapped to the same monitoring grid node of the aforementioned rule are organized and packaged to form the synchronized dataset.

[0007] Furthermore, preliminary estimates of chlorophyll concentration, algal activity indicators, water optical attenuation coefficient, water pH, and thermodynamic state are extracted from the synchronous dataset to form a set of multidimensional water characteristic parameters, including: Dark current correction and temperature effect compensation were performed on the chlorophyll fluorescence signal, and the corrected signal was converted into a preliminary estimate of the chlorophyll concentration using a calibration curve. The fluorescence signal of the blue-green algae is subjected to spectral separation processing to calculate the contribution ratio of fluorescence of a specific algal species, and the ratio of the contribution ratio to the background fluorescence intensity is used as the algal activity indicator value. Based on the turbidity data and the light intensity attenuation data at a specific wavelength, the optical attenuation coefficient of the water body is calculated according to Beer-Lambert's law. The pH value data is directly used as the acid-base state of the water body; The difference between the water temperature data and the historical average water temperature for the same period is taken as the thermodynamic state.

[0008] Furthermore, the multidimensional water body feature parameters are input into a deep feature fusion network. This deep feature fusion network outputs a high-dimensional fused feature vector through adaptive weighting and nonlinear mapping, including: Each of the multidimensional water body feature parameters is transformed into an initial feature vector with the same dimension through an independent feature embedding sub-network. Calculate the mutual information estimate between every two initial feature vectors, and construct a feature correlation matrix based on the mutual information estimate; Based on the feature correlation matrix, the fusion weight of each of the multidimensional water feature parameters in the current context is dynamically calculated through an attention mechanism; The initial feature vectors are weighted and summed using the fusion weights to obtain a preliminary fusion vector; The initial fusion vector is input into a multilayer perceptron. The multilayer perceptron performs deep nonlinear transformation and information aggregation on the initial fusion vector through multiple fully connected layers and nonlinear activation functions, and finally outputs the high-dimensional fusion feature vector.

[0009] Furthermore, the eutrophication state inference engine activates relevant nodes and paths in the causal graph model based on the high-dimensional fused feature vector, and calculates a set of posterior probability distributions of core eutrophication indicators, including: The causal graph model includes observed variable nodes, latent variable nodes, and eutrophication index nodes, wherein the observed variable nodes correspond to the multidimensional water body characteristic parameters. The high-dimensional fusion feature vector is used as evidence and input into the corresponding observation variable node in the causal graph model; In the causal graph model, probabilistic message passing is performed to propagate probabilistic information from the observed variable node along the causal edge to the latent variable node and the eutrophication index node. The joint probability distribution of all nodes in the causal graph model is updated using the belief propagation algorithm. From the updated joint probability distribution, extract the probability distributions corresponding to the nodes representing the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive nutrient status index, and use them as their posterior probability distributions.

[0010] Furthermore, based on the aforementioned posterior probability distribution, a spatiotemporal evolution map of eutrophication in the monitored water area is constructed, including: A state inversion algorithm with physical and biochemical constraints is used to jointly optimize and solve the posterior probability distribution and the multidimensional water body characteristic parameters; The state inversion algorithm incorporates the mass conservation equation, empirical relationships of photosynthetic rate, and kinetic constraints of nutrient cycling during the solution process, and iteratively corrects the estimated values ​​of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive nutrient state index. Based on the optimal estimate obtained after iterative convergence, the transient eutrophication level of the monitored water body corresponding to the timestamp is determined. Based on the transient eutrophication levels of continuous time series, a spatiotemporal evolution map of eutrophication in the monitored waters is constructed, and its evolution trend and spatial heterogeneity characteristics are identified.

[0011] Furthermore, the state inversion algorithm using a physical and biochemical constraint, which jointly optimizes and solves the posterior probability distribution and the multidimensional water body characteristic parameters, includes: The posterior probability distributions of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive nutrient status index are used as the initial inputs for optimization. A joint loss function is constructed, which includes a data fitting term and a physical constraint term. The data fitting term measures the difference between the optical and chemical indicators derived from the estimated nutrient concentration and the multidimensional water body characteristic parameters. The physical constraint term forces the solution to satisfy the mass conservation equation, the empirical relationship of the photosynthetic rate, and the kinetic constraints of the nutrient cycle. The gradient descent method is used to iteratively adjust the estimated values ​​of the total phosphorus concentration, the total nitrogen concentration, and the comprehensive nutrient status index in order to minimize the joint loss function; When the value of the joint loss function changes less than a preset threshold in multiple consecutive iterations, the iteration is determined to be converged, and the current optimal estimate is output.

[0012] Furthermore, the state inversion algorithm incorporates the mass conservation equation, empirical relationships of photosynthetic rates, and kinetic constraints of nutrient cycling during the solution process, including: The mass conservation equation is an equation in which the rate of change of total phosphorus and total nitrogen within a water volume unit is equal to the sum of the input and output fluxes and the internal source and sink terms. The empirical relationship for photosynthetic rate is a semi-empirical formula for estimating primary productivity rate based on water temperature, light intensity, and nutrient concentration. The kinetic constraints of the nutrient cycle include the adsorption and desorption kinetic equations between dissolved and particulate nutrients, and the nitrification and denitrification reaction rate equations. In each iteration of calculating the physical constraint term of the joint loss function, the current nutrient concentration estimate is substituted into the mass conservation equation, the empirical relationship of the photosynthetic rate, and the kinetic constraints of the nutrient cycle to calculate the predicted theoretical response value. The theoretical response value is compared with the actual response value derived from the multidimensional water body characteristic parameters, and the difference constitutes part of the physical constraint term.

[0013] Further, determining the transient eutrophication level of the monitored water body corresponding to the timestamp based on the optimal estimate obtained after iterative convergence includes: The final value of the comprehensive nutritional status index is extracted from the optimal estimate; The final value of the comprehensive nutrient status index is compared with the preset threshold range for eutrophication level classification. Based on the comparison results, the current state of the monitored water area is classified into one of the following levels: oligotrophic, mesotrophic, slightly eutrophic, moderately eutrophic, or severely eutrophic. The level category is the transient eutrophication level. The transient eutrophication level, the corresponding timestamp, and the spatial location information of the monitoring point are bound and stored together.

[0014] Furthermore, based on the transient eutrophication levels of continuous time series, a spatiotemporal evolution map of eutrophication in the monitored water area is constructed, and its evolution trend and spatial heterogeneity characteristics are identified, including: Collect a series of transient eutrophication levels and their associated information on all monitoring grid nodes within a preset monitoring period; Using time as the axis and space as the surface, the transient eutrophication levels at different time points are plotted into a series of spatial distribution raster maps. Arrange the series of spatial distribution raster maps in chronological order to form a spatiotemporal animation or sequence map of eutrophication, namely the spatiotemporal evolution map of eutrophication. Spatiotemporal statistical analysis was performed on the eutrophication spatiotemporal evolution map to calculate the slope of the change in eutrophication level of the entire water body or a specific area over time, which was used as the evolution trend. The variance or spatial autocorrelation index of eutrophication levels at different spatial locations at the same time point is analyzed as a quantitative indicator of the spatial heterogeneity characteristics.

[0015] Compared with existing technologies, the river and lake eutrophication status monitoring system based on multi-sensor fusion of this invention has achieved the following beneficial effects through practical application: 1. This invention inputs multi-dimensional water body feature parameters from multiple sources and heterogeneous sources into a deep feature fusion network. Through adaptive weighting and nonlinear mapping, it outputs a high-dimensional fusion feature vector, which can match the actual monitoring characteristics of different water body feature parameters. This avoids the loss of feature information caused by conventional shallow fusion methods. Nonlinear mapping can sort out the intrinsic correlation between multi-dimensional features. The high-dimensional fusion feature vector can completely retain the effective information of various water body features, making the expression of feature data fit the real state of rivers and lakes.

[0016] 2. This invention inputs a high-dimensional fused feature vector into the eutrophication state inference engine of the embedded causal graph model. Based on the high-dimensional fused feature vector, the relevant nodes and paths in the causal graph model are activated to calculate the posterior probability distribution of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive trophic state index. The probabilistic causal relationship preset by the causal graph model fits the actual correlation logic between the causes and characteristics of eutrophication. The posterior probability distribution of the indicators can intuitively reflect the dynamic change attributes of the indicators. The eutrophication spatiotemporal evolution map constructed based on this probability distribution can fully show the spatiotemporal change trend of eutrophication in the monitored water area. Attached Figure Description

[0017] Figure 1 This is a time-series diagram of a river and lake eutrophication status monitoring system according to the present invention. Figure 2 A flowchart for deep feature fusion network processing; Figure 3 This is a posterior probability distribution of total phosphorus concentration. Figure 4 A heatmap showing the correlation between monitoring indicators of eutrophication in rivers and lakes; Figure 5 This is a time series trend chart of the core indicators of eutrophication. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1This invention relates to a multi-sensor fusion-based river and lake eutrophication status monitoring system. The system acquires multi-source, heterogeneous raw sensor data from the monitored water area through a data acquisition module, forming a set of multi-dimensional water body feature parameters. A feature fusion module inputs these multi-dimensional water body feature parameters into a deep feature fusion network. This network outputs a high-dimensional fused feature vector through adaptive weighting and nonlinear mapping. A state inference module feeds the high-dimensional fused feature vector into an eutrophication state inference engine. This engine embeds a causal graph model trained based on historical cases, which defines the probabilistic causal relationship between multi-dimensional causes and eutrophication characteristics. A probability calculation module drives the eutrophication state inference engine to activate relevant nodes and paths in the causal graph model based on the high-dimensional fused feature vector, calculating a posterior probability distribution of a set of core eutrophication indicators. These core indicators include at least simulated total phosphorus concentration, simulated total nitrogen concentration, and a comprehensive trophic state index. A map construction module constructs a spatiotemporal evolution map of eutrophication in the monitored water area based on the posterior probability distribution.

[0020] In one embodiment of the present invention, an underwater sensor array deployed in the monitored water area acquires multi-source heterogeneous raw sensing data. This raw sensing data includes at least chlorophyll fluorescence signals, cyanobacterial fluorescence signals, dissolved oxygen concentration, turbidity, pH value, water temperature, and incident and emitted light intensity data at specific wavelengths. Synchronous acquisition and spatiotemporal alignment processing are performed on the raw sensing data to generate a synchronized dataset with unified timestamps and matching spatial locations. A high-precision clock module is configured for each sensor in the underwater sensor array, and clock synchronization with a unified time server is performed periodically. During each data acquisition cycle, all sensors are controlled to complete data sampling within a preset synchronization time window, and each sampled data is stamped with the same timestamp. The real-time three-dimensional spatial coordinates of each sensor in the underwater sensor array are acquired. Based on the real-time three-dimensional spatial coordinates, sampled data at the same timestamp but different spatial locations are interpolated and mapped to nodes of a preset regular monitoring grid. All data from different sensors mapped to the same node of the regular monitoring grid are organized and packaged to form the synchronized dataset. Preliminary estimates of chlorophyll concentration, algal activity indicators, water optical attenuation coefficient, water pH, and thermodynamic state are extracted from the synchronous dataset to form a set of multidimensional water characteristic parameters. Dark current correction and temperature effect compensation are applied to the chlorophyll fluorescence signal, and a calibration curve is used to convert the corrected signal into the preliminary estimate of chlorophyll concentration. Spectral separation processing is performed on the cyanobacterial fluorescence signal to calculate the contribution ratio of fluorescence for a specific algal species; the ratio of this contribution to the background fluorescence intensity is used as the algal activity indicator. The water optical attenuation coefficient is calculated based on Beer-Lambert's law using the turbidity data and light intensity attenuation data for a specific wavelength. The pH value is directly used as the water pH. The difference between the water temperature data and the historical average water temperature for the same period is used as the thermodynamic state.

[0021] In practice, the monitoring involves acquiring raw sensing data of the monitored water area, synchronous alignment processing, and extraction of multidimensional water body characteristic parameters. The underwater sensor array deployed in the monitored water area contains various types of sensor probes. These probes acquire multi-source, heterogeneous raw sensing data. The raw sensing data set includes at least chlorophyll fluorescence signals, cyanobacterial fluorescence signals, dissolved oxygen concentration, turbidity values, pH readings, water temperature readings, and incident and emitted light intensity data at specific wavelengths. Synchronous acquisition and spatiotemporal alignment processing are performed on the raw sensing data to generate a synchronized dataset with unified timestamps and spatially matched locations, which forms the basis for subsequent processing.

[0022] In some embodiments, a high-precision clock module is configured for each sensor in the underwater sensor array. This high-precision clock module periodically synchronizes with a unified time server deployed on the monitoring base station. During each preset data acquisition cycle, the central controller sends a synchronization sampling command to all sensors, controlling them to complete data sampling within a preset synchronization time window. After sampling, the data acquisition unit assigns the same timestamp to each sampled data. Simultaneously, the positioning system integrated into the underwater sensor array acquires the real-time three-dimensional spatial coordinates of each sensor in the array. Based on these real-time three-dimensional spatial coordinates, the data alignment processor maps sampled data from different spatial locations at the same timestamp to nodes of a preset regular monitoring grid using a spatial interpolation algorithm. All data successfully mapped to the same regular monitoring grid node from different sensors are organized and packaged to form a well-structured synchronized dataset. Preliminary estimates of chlorophyll concentration, algal activity indicators, water optical attenuation coefficient, water pH, and thermodynamic state are extracted from the generated synchronized dataset. These parameters collectively constitute a set of multi-dimensional water characteristic parameters. In practice, the data processing unit performs dark current correction and temperature effect compensation on the chlorophyll fluorescence signal. Then, using a pre-calibrated conversion curve, it converts the corrected chlorophyll fluorescence signal values ​​into a preliminary estimate of chlorophyll concentration. The fluorescence signal from cyanobacteria is subjected to spectral separation processing, and the contribution ratio of a specific algal species' fluorescence peak in the total fluorescence signal is calculated. The ratio of this contribution ratio to the background fluorescence intensity is used as an indicator of algal community activity. It can be understood that, based on turbidity data and light intensity attenuation data at a specific wavelength, the optical attenuation coefficient of the water body is calculated using Beer-Lambert's law. The calculation process follows the formula:

[0023] in: This represents the optical attenuation coefficient of the water body to be determined. This indicates the length of the path the light beam takes in the water. and These represent the readings of the incident and emitted light intensities, respectively. The pH data recorded in the synchronous dataset is directly used as a water body acid-base state parameter characterizing the water's acid-base state. The difference between the water temperature data recorded in the synchronous dataset and the pre-calculated and stored historical average water temperature for the same period is used as a thermodynamic state parameter reflecting the water's thermal anomalies. Optionally, during the synchronous dataset generation stage, when a node in a certain regular monitoring grid lacks directly mapped sensor data, an inverse distance weighted interpolation method is used to calculate the estimated value of that node using sensor data from neighboring nodes. In the feature parameter extraction stage, for temperature effect compensation of the chlorophyll fluorescence signal, a linear correction coefficient related to the water temperature reading is used to adjust the original fluorescence signal reading.

[0024] In some embodiments, the synchronous sampling time window of the underwater sensor array can be dynamically configured according to monitoring needs, such as shortening the sampling interval and synchronizing the time window during periods of high algal blooms. Multidimensional water feature parameters extracted from the synchronous dataset, including preliminary estimates of chlorophyll concentration, algal activity indicators, water optical attenuation coefficient, water pH, and thermodynamic state, are organized and output in vector form for use by the subsequent feature fusion module. Optionally, the multidimensional water feature parameters are normalized before output, scaling the numerical range of all parameters to the same interval to eliminate the impact of differences in dimensions and magnitudes of different parameters on subsequent model calculations. Historical average water temperature values ​​for the same period are obtained by querying a long-term monitoring database, which stores the average water temperature of the monitoring points for the same period over the past few years. It is understood that the incident and emitted light intensity data of specific wavelengths in the raw sensing data are typically selected from bands sensitive to phytoplankton pigment absorption, such as the 440 nm blue light band and the 676 nm red light band, to calculate the water optical attenuation coefficient for that specific band. In practice, the data acquisition module periodically executes the entire process from data acquisition and synchronization alignment to feature extraction, thereby continuously producing a multidimensional water body feature parameter sequence over time.

[0025] In one embodiment of the present invention, see [reference] Figure 2 Each of the multidimensional water feature parameters is transformed into an initial feature vector of the same dimension through an independent feature embedding sub-network. The mutual information estimate between any two initial feature vectors is calculated, and a feature correlation matrix is ​​constructed based on the mutual information estimate. Based on the feature correlation matrix, the fusion weight of each multidimensional water feature parameter in the current context is dynamically calculated using an attention mechanism. All initial feature vectors are weighted and summed using the fusion weights to obtain a preliminary fusion vector. This preliminary fusion vector is input into a multilayer perceptron, which performs deep nonlinear transformations and information aggregation on the preliminary fusion vector through multiple fully connected layers and nonlinear activation functions, ultimately outputting the high-dimensional fusion feature vector.

[0026] In specific implementation, multidimensional water body feature parameters are transformed into high-dimensional fused feature vectors through a deep feature fusion network. This deep feature fusion network consists of multiple steps, including feature embedding, correlation calculation, weight fusion, and nonlinear transformation. The multidimensional water body feature parameter is an input vector containing multiple components, such as a vector containing five components: a preliminary estimate of chlorophyll concentration, an indicator of algal activity, the optical attenuation coefficient of the water body, the acid-base state of the water body, and the thermodynamic state. Each component represents a water body feature parameter. Each item in the multidimensional water body feature parameter is passed through an independent feature embedding sub-network. Each feature embedding sub-network is a small fully connected neural network used to convert the scalar form of a single feature parameter into an initial feature vector with the same preset dimension d. In some embodiments, the independent feature embedding sub-networks have the same structure but independent parameters. Each feature embedding sub-network receives a scalar input, processes it through two fully connected layers and an activation function, and outputs a d-dimensional vector. The mutual information estimate between every two initial feature vectors is calculated. The mutual information estimate is used to measure the statistical dependence between the information carried by the two feature vectors. For a set containing n feature vectors, the estimated mutual information between each pair of vectors forms an n×n symmetric matrix, which is the feature correlation matrix. The mutual information estimate can be calculated using an entropy estimation algorithm based on k-nearest neighbors, with the following formula:

[0027] in: This represents the estimated mutual information between eigenvectors X and Y. It is the digamma function. It is the total number of samples. It is the number of nearest neighbors. and These are counts based on a joint spatial distance metric centered at X and Y. Based on the feature relevance matrix, an attention mechanism dynamically calculates the fusion weight of each multidimensional water feature parameter in the current context. The attention mechanism normalizes each row of the feature relevance matrix, resulting in a set of weight coefficients that sum to 1.

[0028] In practical implementation, an attention mechanism is used to dynamically calculate the fusion weight of each multidimensional water feature parameter in the current context. The attention mechanism's operation can be understood as follows: each row of the feature correlation matrix is ​​input into a Softmax function. The Softmax function converts the values ​​of each row into a probability distribution, where the value at each position represents the attention weight of the corresponding feature to the feature represented by the current row. The calculated fusion weights are then used to perform a weighted summation on all initial feature vectors. This weighted summation involves multiplying each initial feature vector by its corresponding fusion weight and then adding all the weighted vectors together to obtain a preliminary fusion vector. Optionally, the preliminary fusion vector has the same dimension as the initial feature vectors, both being d-dimensional. The preliminary fusion vector is then input into a multilayer perceptron (MLP). The MLP performs deep nonlinear transformations and information aggregation on the preliminary fusion vector through multiple fully connected layers and nonlinear activation functions. The MLP typically contains two or more hidden layers using the ReLU activation function. The final output layer can have no activation function or use a linear activation function, outputting the high-dimensional fusion feature vector.

[0029] In some embodiments, the dimension f of the high-dimensional fused feature vector can be set to a value higher than the dimension d of the initial feature vector, for example, d=32, f=128. In specific implementations, the training of the deep feature fusion network is performed jointly with the training of the entire monitoring system. Its parameters are optimized on historical datasets using gradient descent, so that the final output high-dimensional fused feature vector can more effectively serve the probability calculation of the subsequent state inference module. The entire transformation process from multidimensional water feature parameters to high-dimensional fused feature vectors explicitly models and compresses the nonlinear relationships and inter-feature dependencies in the input data. Optionally, the feature embedding sub-network, the attention weight calculation module, and the multilayer perceptron together constitute the set of trainable parameters for the deep feature fusion network. In the model inference stage, after a new set of multidimensional water feature parameters is input, the corresponding high-dimensional fused feature vector can be obtained through the aforementioned fixed forward propagation calculation.

[0030] In one embodiment of the present invention, the causal graph model includes observed variable nodes, latent variable nodes, and eutrophication index nodes, wherein the observed variable nodes correspond to the multidimensional water body characteristic parameters. The high-dimensional fused feature vector is input as evidence into the corresponding observed variable node in the causal graph model. Probabilistic message passing is performed in the causal graph model, propagating probabilistic information from the observed variable node along causal edges to the latent variable node and the eutrophication index node. The joint probability distribution of all nodes in the causal graph model is updated using a belief propagation algorithm. From the updated joint probability distribution, the probability distributions corresponding to the nodes representing the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive trophic state index are extracted as their posterior probability distributions.

[0031] In practical implementation, the causal graph model embedded in the eutrophication state inference engine is utilized to calculate the posterior probability distribution of core eutrophication indicators based on high-dimensional fused feature vectors. The causal graph model is a probabilistic graphical model containing three types of nodes: observed variable nodes, latent variable nodes, and eutrophication indicator nodes. The observed variable nodes correspond to multidimensional water body feature parameters, with each node representing an input water body feature parameter. The high-dimensional fused feature vector is input as evidence to the corresponding observed variable node in the causal graph model. Each dimension or group of dimensions of the high-dimensional fused feature vector is assigned to the observed variable node, serving as the observed value or feature representation of the observed value, thus transforming the continuous feature vector into evidence specifications for the observed variables in the causal graph model.

[0032] In practical implementation, nodes in the causal graph model are connected by directed edges, which represent probabilistic causal relationships between variables. Observed variable nodes are typically located at the beginning or middle of the causal chain, latent variable nodes represent mediating factors that are not directly observed but influence the eutrophication process, and eutrophication index nodes represent core indicators that need to be inferred, including nodes for simulated total phosphorus concentration, simulated total nitrogen concentration, and the comprehensive nutrient status index. Inputting a high-dimensional fused feature vector as evidence into the corresponding observed variable node in the causal graph model means assigning the values ​​of the vector elements to the state of the corresponding observed variable node according to a pre-defined mapping relationship. Probabilistic message passing is performed in the causal graph model, propagating probabilistic information from the observed variable node along the causal edges to the latent variable node and the eutrophication index node. Probabilistic message passing is the core mechanism for inference in the causal graph model; the message is a function of the probability distribution of the variable that is passed between nodes. After receiving evidence, the observed variable node updates its own belief, i.e., the probability distribution, and generates a new message to send to its connected parent and child nodes. These messages are propagated along the directed edges in the causal graph model, sequentially activating and updating the probability information of latent variable nodes and eutrophication index nodes.

[0033] In some embodiments, probabilistic message passing follows the local message passing rules defined by the causal graph model. For any node, after receiving messages from all its neighboring nodes, it calculates its own marginal belief distribution and generates a message to send to the next neighboring node based on this distribution and model parameters. This process iterates among all nodes until all transmitted messages tend to stabilize. The joint probability distribution of all nodes in the causal graph model is updated using a belief propagation algorithm. The belief propagation algorithm is a specific iterative algorithm for implementing the above probabilistic message passing. It cyclically transmits messages among the nodes of the causal graph model using a serial or parallel scheduling method, and continuously updates the marginal probability distribution, i.e., the "belief," of each node based on the transmitted messages. In a specific implementation, the update rule of the belief propagation algorithm can be expressed as:

[0034] in: Indicates from node Send to node About nodes state The news, It is a node The local potential function, It is a connection node and nodes The compatibility function on the edge, Represents a node Except for nodes The set of all neighboring nodes other than From node Other neighboring nodes The received message. The algorithm iteratively performs message calculation and transmission until the change in the message is less than a certain threshold, at which point the belief of each node is determined. This approximates the marginal probability distribution, with the beliefs of all nodes jointly approximating the joint probability distribution of the causal graph model. From the updated joint probability distribution, the probability distributions corresponding to the nodes representing the simulated total phosphorus concentration, simulated total nitrogen concentration, and comprehensive trophic state index are extracted as their posterior probability distributions. After the belief propagation algorithm converges, the eutrophication state inference engine directly reads the marginal belief distributions of the simulated total phosphorus concentration node, simulated total nitrogen concentration node, and comprehensive trophic state index node. These marginal belief distributions are the posterior probability distributions of each core eutrophication indicator given the observational evidence. Optionally, the structure and parameters of the causal graph model are learned through historical monitoring case data, which includes multidimensional water body characteristic parameters and corresponding accurate labels for total phosphorus concentration, total nitrogen concentration, and comprehensive trophic state index obtained through laboratory analysis. During the model usage phase, the learned causal graph model parameters remain fixed, and inference is performed only based on the input high-dimensional fused feature vector.

[0035] It is understandable that the posterior probability distribution is usually represented in the form of a discrete probability mass function or a continuous probability density function. For simulated total phosphorus and total nitrogen concentrations, their posterior probability distributions may be continuous distributions, such as Gaussian distributions, whose mean and variance are determined by the beliefs of the nodes. For the comprehensive trophic state index, its posterior probability distribution may be a discrete probability distribution at different trophic state levels.

[0036] In some embodiments, the causal graph model can be specifically implemented as a Bayesian network. Observational variable nodes correspond to sensor features, latent variable nodes may include concepts that cannot be directly measured, such as "nutrient input flux" and "algal growth rate," and eutrophication index nodes are the target variables. Message passing and belief updates are performed within the framework of the Bayesian network. Optionally, when the causal graph model is large, approximate inference algorithms, such as cyclic belief propagation or sampling-based methods, can be used to efficiently calculate the posterior probability distribution of the nodes. Before inputting the high-dimensional fused feature vector into the observation variable nodes, it may need to be discretized or subjected to probability soft assignment to match the value types of the observation variables in the causal graph model. After extracting the posterior probability distribution from the updated joint probability distribution, the eutrophication state inference engine outputs the posterior probability distributions of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive trophic state index. These posterior probability distributions, as the output of the probability calculation module, are passed to the subsequent graph construction module for further processing. It can be understood that the process of probability message passing and belief propagation is automated and requires no manual intervention. In practice, the eutrophication state inference engine is implemented as a software module. It receives high-dimensional fused feature vectors from the feature fusion module, performs causal graph model inference, and outputs structured posterior probability distribution data.

[0037] See Figure 3 This is a posterior probability distribution chart of total phosphorus concentration, visually reflecting the changes in the accuracy and uncertainty of total phosphorus concentration estimation at each stage. The initial monitoring stage shows the highest peak and narrowest distribution, with a mean of approximately 0.08 mg / L and minimal uncertainty. In the feature fusion stage, the mean shifts to the right to approximately 0.10 mg / L, the distribution widens, and uncertainty slightly increases. In the probabilistic inference stage, the mean further shifts to the right to approximately 0.12 mg / L, the distribution widens, and uncertainty is greatest. In the iterative optimization stage, the mean falls back to approximately 0.11 mg / L, the distribution narrows, and uncertainty decreases. In the final evaluation stage, the mean stabilizes at approximately 0.09 mg / L, with a moderate distribution width, and the results tend to converge. The gradual increase from the initial 0.08 mg / L to 0.12 mg / L in the inference stage reflects the role of feature fusion and probabilistic inference in uncovering potential nutrient loads. The final drop to 0.09 mg / L indicates that the optimization algorithm corrected over-inference, making the results more consistent with the actual water conditions.

[0038] In one embodiment of the present invention, a state inversion algorithm with physical and biochemical constraints is used to jointly optimize and solve the posterior probability distribution and the multidimensional water body characteristic parameters. The posterior probability distributions of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive trophic state index are used as the initial inputs for the optimization solution. A joint loss function is constructed, comprising a data fitting term and a physical constraint term. The data fitting term measures the difference between the optical and chemical indicators derived from the estimated nutrient concentrations and the multidimensional water body characteristic parameters, while the physical constraint term forces the solution results to satisfy the mass conservation equation, the empirical relationship of the photosynthetic rate, and the kinetic constraints of the nutrient cycle. The gradient descent method is used to iteratively adjust the estimated values ​​of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive trophic state index to minimize the joint loss function. When the value of the joint loss function changes less than a preset threshold in multiple consecutive iterations, the iteration is considered to have converged, and the current optimal estimate is output.

[0039] The state inversion algorithm incorporates a mass conservation equation, an empirical relationship for photosynthetic rates, and kinetic constraints for nutrient cycling during the solution process. The mass conservation equation sets the rate of change of total phosphorus and total nitrogen within a water volume unit equal to the sum of input / output fluxes and internal source / sink terms. The empirical relationship for photosynthetic rates is a semi-empirical formula for estimating primary productivity rates based on water temperature, light intensity, and nutrient concentration. The kinetic constraints for nutrient cycling include adsorption / desorption kinetic equations between dissolved and particulate nutrients, and nitrification / denitrification reaction rate equations. In each iteration calculating the physical constraints of the joint loss function, the current estimated nutrient concentration is substituted into the mass conservation equation, the empirical relationship for photosynthetic rates, and the kinetic constraints of nutrient cycling to calculate the predicted theoretical response value. The theoretical response value is compared with the actual response value derived from the multidimensional water body characteristic parameters; the difference constitutes part of the physical constraint term.

[0040] In practical implementation, a state inversion algorithm with physical and biochemical constraints is used to jointly optimize and solve for the posterior probability distribution and multidimensional water body characteristic parameters, thereby obtaining the optimal estimates of the core eutrophication indicators. The state inversion algorithm uses the posterior probability distributions of simulated total phosphorus concentration, simulated total nitrogen concentration, and comprehensive trophic state index as the initial inputs for optimization. These posterior probability distributions provide probabilistic prior information on the values ​​of the core indicators. A joint loss function is constructed, including a data fitting term and a physical constraint term. The data fitting term measures the difference between the theoretical optical and chemical indicators derived from the nutrient concentrations and comprehensive trophic state index estimated in the current iteration and the measured multidimensional water body characteristic parameters. The physical constraint term forces the solution results to satisfy the mass conservation equation, the empirical relationship of photosynthetic rate, and the kinetic constraints of nutrient cycling.

[0041] In some embodiments, the joint loss function The specific form is defined as the data fitting term. With physical constraints The weighted sum, i.e.:

[0042] in: and These are pre-defined hyperparameters that balance the weights of the two components. Data fitting term. It can be defined in the form of mean square error, for example ,in It is the i-th theoretical optical or chemical index value calculated using a semi-empirical formula from the currently estimated simulated total phosphorus concentration, simulated total nitrogen concentration, and comprehensive nutrient status index. It corresponds to the i-th observation index value obtained from the multidimensional water body characteristic parameters. This is the weighting coefficient for this item. Physical constraint item. It consists of the sum of squares of the degree of violation of multiple sub-constraints, and each sub-constraint corresponds to a physical and biochemical law.

[0043] In practice, the gradient descent method is employed to iteratively adjust the estimates of total phosphorus concentration, total nitrogen concentration, and comprehensive trophic state index to minimize the joint loss function. In each iteration, the gradient descent method calculates the gradient of the joint loss function with respect to the estimates of total phosphorus concentration, total nitrogen concentration, and comprehensive trophic state index, and then updates these estimates in the reverse direction of the gradient. When the value of the joint loss function changes less than a preset threshold in multiple consecutive iterations, the iteration is considered to have converged, and the current optimal estimates of total phosphorus concentration, total nitrogen concentration, and comprehensive trophic state index are output.

[0044] The state inversion algorithm incorporates a mass conservation equation, an empirical relation on photosynthetic rates, and kinetic constraints on nutrient cycling during the solution process. The mass conservation equation sets the rate of change of total phosphorus and total nitrogen within a water volume unit equal to the sum of input-output fluxes and internal source-sink terms. The empirical relation on photosynthetic rates is a semi-empirical formula for estimating primary productivity rates based on water temperature, light intensity, and nutrient concentration. The kinetic constraints on nutrient cycling include the adsorption-desorption kinetics between dissolved and particulate nutrients, and the nitrification and denitrification reaction rate equations. In each iteration, when calculating the physical constraints of the joint loss function, the current simulated total phosphorus concentration estimate, total nitrogen concentration estimate, and comprehensive trophic state index estimate are substituted into the mass conservation equation, the empirical relation on photosynthetic rates, and the kinetic constraints of nutrient cycling to calculate the predicted theoretical response value. The theoretical response value is compared with the actual response value derived from multidimensional water body characteristic parameters; the difference constitutes part of the physical constraint term. It is understandable that deriving actual response values ​​from multidimensional water body characteristic parameters involves the conversion of raw data. For example, deriving the actual net photosynthetic rate from the rate of change in dissolved oxygen concentration, and deriving the actual carbonic acid system change from pH value change to indirectly reflect certain processes. Refer to Table 1, which shows the comparison between the theoretical and actual response values ​​of some sub-items of the physical constraint term in one iteration calculation.

[0045] Table 1: Calculation Table of Physical Constraints

[0046] Optionally, gradient descent can employ variants with adaptive learning rates, such as the Adam optimizer, to accelerate convergence and avoid getting trapped in local minima. A preset convergence threshold is set based on the required monitoring accuracy; for example, the change in the joint loss function value over 10 consecutive iterations should be less than 1e-6. The specific form of the mass conservation equation needs to be determined based on the hydraulic conditions of the monitored water area, and may employ a fully mixed reactor model or a one-dimensional advection diffusion reaction equation. In some embodiments, the empirical relationship for the photosynthetic rate can be expressed as a formula of the following form:

[0047] in: It is the predicted primary productivity. It is the maximum photosynthetic rate. It is the initial light energy utilization efficiency. It is the intensity of light. and These are the currently estimated total phosphorus and total nitrogen concentrations. and It is the half-saturation constant. It is the temperature coefficient. It's the water temperature. This is the optimal temperature. The actual photosynthetic rate is derived from multidimensional water body characteristic parameters. This can be obtained through the dissolved oxygen diurnal variation method or chlorophyll fluorescence light response curve; the corresponding sub-term in the physical constraint term is... .

[0048] In practical implementation, the kinetic constraints of the nutrient cycle, such as the nitration reaction rate equation, can be expressed using a first-order kinetic equation:

[0049] in: It is the nitration rate. It is a rate constant affected by water temperature. This refers to the concentration of ammonium nitrogen. The actual nitrification rate... The derivation may be based on specific pH variation patterns or an assumption of the consumption rate of ammonium nitrogen under steady state. Adsorption-desorption kinetics can be expressed using linear isotherms. Simplified description, where It refers to the concentration of nutrients in the particulate phase. It refers to the concentration of nutrients in the dissolved phase. These are the allocation coefficients. Optional, they are the weighting coefficients in the joint loss function. and Cross-validation can be used to adjust the algorithm, balancing the trade-off between data fit and consistency with physical laws. The outputs of the state inversion algorithm—the optimal estimates of total phosphorus concentration, total nitrogen concentration, and comprehensive trophic state index—will serve as the basis for determining the transient eutrophication level. The solution process of the state inversion algorithm is iterative and automated. In its implementation, the algorithm uses the statistical characteristics (such as the mean) of the posterior probability distribution output by the probability calculation module as the initial estimate, and uses multidimensional water body characteristic parameters as the observation comparison benchmark. It continuously adjusts the estimate by minimizing the joint loss function until both data fit and physical-biochemical constraints are simultaneously met.

[0050] See Figure 4 This is a heatmap showing the correlation between eutrophication monitoring indicators in rivers and lakes. Colors represent the degree of linear correlation between variables. Total phosphorus and total nitrogen are strongly positively correlated with CSI, verifying that nutrients are the core cause of eutrophication. Dissolved oxygen is strongly negatively correlated with CSI, serving as a direct monitoring indicator of eutrophication status. Easily observable indicators such as chlorophyll and turbidity are highly correlated with nutrients and CSI, and difficult-to-measure indicators can be indirectly inferred through multi-sensor fusion. A rapid decrease in dissolved oxygen concentration and a simultaneous increase in chlorophyll / turbidity can serve as an early warning signal of eutrophication outbreaks. This heatmap clearly reveals the linear correlation network between river and lake water characteristics and eutrophication indicators, providing crucial prior knowledge of variable correlations for multi-sensor fusion monitoring and the construction of causal graph models.

[0051] In one embodiment of the present invention, the final value of the comprehensive trophic state index is extracted from the optimal estimate. The final value of the comprehensive trophic state index is compared with a preset threshold range for eutrophication level classification. Based on the comparison result, the current state of the monitored water area is classified into one of the following levels: oligotrophic, mesotrophic, slightly eutrophic, moderately eutrophic, or severely eutrophic. This level category is the transient eutrophication level. The transient eutrophication level, the corresponding timestamp, and the spatial location information of the monitoring point are bound and stored. A series of transient eutrophication levels and their bound information are collected from all monitoring grid nodes within a preset monitoring period. Using time as the axis and space as the surface, the transient eutrophication levels at different time points are plotted into a series of spatial distribution raster maps. The series of spatial distribution raster maps are arranged in chronological order to form a spatiotemporal animation or sequence diagram of the eutrophication state, i.e., the spatiotemporal evolution map of eutrophication. Spatiotemporal statistical analysis was performed on the aforementioned eutrophication spatiotemporal evolution map to calculate the slope of eutrophication level change over time for the entire water body or a specific region, which was used as the evolution trend. The variance or spatial autocorrelation index of eutrophication levels at different spatial locations at the same time point was analyzed as a quantitative indicator of the spatial heterogeneity characteristics.

[0052] In practice, the optimal estimate obtained from state inversion determines the transient eutrophication level, and a spatiotemporal evolution map of eutrophication is constructed based on the level information of continuous time series. The final value of the comprehensive trophic state index is extracted from the optimal estimate output by the state inversion algorithm; this final value is a scalar representing the comprehensive trophic level of the water body. The final value of the comprehensive trophic state index is compared with a preset eutrophication level classification threshold range, which is usually set according to national or industry standards, dividing continuous index values ​​into several non-overlapping intervals. Based on the comparison results, the current state of the monitored water body is classified into one of the following categories: oligotrophic, mesotrophic, slightly eutrophic, moderately eutrophic, or severely eutrophic. This determined level category is the transient eutrophication level. The transient eutrophication level characterizes the eutrophication state of the monitored water body at a specific point in time. The transient eutrophication level, the corresponding data acquisition timestamp, and the spatial location information of the monitoring point are bound and stored to form a complete record containing time, space, and state attributes.

[0053] In some embodiments, preset threshold ranges for eutrophication levels can be stored in a lookup table. For example, a composite trophic state index (CSI) less than 30 corresponds to oligotrophic, greater than or equal to 30 and less than 50 to mesotrophic, greater than or equal to 50 and less than 60 to slightly eutrophic, greater than or equal to 60 and less than 70 to moderately eutrophic, and greater than or equal to 70 to severely eutrophic. The binding storage operation records the transient eutrophication level, timestamp, latitude and longitude coordinates of the monitoring point, and water depth information in a database or a data file of a specific format for subsequent spatiotemporal analysis. A series of transient eutrophication levels and their binding information are collected from all monitoring grid nodes within a preset monitoring period. The preset monitoring period can be one day, one week, or one month, depending on the objectives of the monitoring task. Using time as the axis and space as the surface, a series of spatial distribution raster maps are drawn from the transient eutrophication levels at different time points. Each spatial distribution raster map represents the spatial distribution of eutrophication status in the monitored water area at a specific sampling time, and the value of each pixel in the map represents the transient eutrophication level of its corresponding regular monitoring grid node at that time.

[0054] In practice, spatial distribution raster maps can be created using the drawing functions of geographic information system (GIS) software libraries or scientific computing libraries. Arranging a series of spatial distribution raster maps chronologically can form a spatiotemporal animation or sequence map of eutrophication status, which constitutes an intuitive visualization of the spatiotemporal evolution of eutrophication. The spatiotemporal evolution map of eutrophication dynamically displays the spatial distribution of eutrophication status within the monitored water body and its changes over time. Spatiotemporal statistical analysis of the eutrophication spatiotemporal evolution map is performed to calculate the slope of eutrophication level changes over the entire water body or a specific area, serving as a quantitative indicator of the evolution trend. The slope can be calculated by performing linear regression on the eutrophication level time series for each spatial location; the slope of the regression line represents the long-term trend of eutrophication status at that location. Analyzing the variance or spatial autocorrelation index of eutrophication levels between different spatial locations at the same time point serves as a quantitative indicator of spatial heterogeneity. Spatial variance reflects the dispersion of eutrophication status across an entire water body at a given time, while spatial autocorrelation quantifies the spatial clustering or dispersion pattern of eutrophication. It can be understood that calculating the slope of eutrophication level change across the entire water body over time can employ a holistic analysis method based on the average of all pixel time series values. The specific formula can be expressed as:

[0055] in: This represents the average slope of the change in the eutrophication level of the entire water body over time. It represents the total number of time points. It is a time point sequence number. It is the average of the time sequence numbers. It is a point in time. The average transient eutrophication level at all monitoring points in the entire water body. It is all The average value. This indicates a trend of increasing eutrophication levels and worsening eutrophication. This indicates an improving trend. Optionally, spatial heterogeneity characteristics can also be analyzed by calculating a semivariogram to describe the variability of eutrophication levels with spatial distance. Specific regions can be designated sub-basins, water source protection areas, or algal bloom areas, and their evolution trend can be analyzed separately. Spatiotemporal animations can be played at a controlled speed via web pages or desktop applications, facilitating observation of the evolution process by management personnel.

[0056] In some embodiments, the identification results of evolution trends can be represented as a spatial distribution trend map, where the color of each pixel represents the magnitude and direction of its slope. Quantitative indicators of spatial heterogeneity characteristics can be compared with historical contemporaneous periods or contemporaneous periods in different years to assess the stability or changes in spatial patterns. The completed spatiotemporal evolution map of eutrophication and its statistical analysis results can be used to generate monitoring reports or drive early warning decisions. Optionally, in addition to the five main categories, the classification of eutrophication levels can be further subdivided into more subcategories according to management needs. Before mapping, the transient eutrophication levels of time series can be filtered to smooth random fluctuations and highlight the main trends. When drawing the spatial distribution raster map, spatial interpolation can be used to fill in grid nodes with missing data to generate a continuous spatial distribution surface. The entire process from extracting the final value of the comprehensive trophic state index from the optimal estimate to completing the construction and analysis of the spatiotemporal evolution map of eutrophication is a key step in realizing the monitoring and assessment of river and lake eutrophication status from point to surface, and from static to dynamic. In specific implementations, the map construction module automatically executes all the calculation and visualization processes described in this embodiment.

[0057] See Figure 5 This is a time series trend chart of core eutrophication indicators, showing the dynamic changes of these indicators over a one-year period (365 days), intuitively reflecting the seasonal patterns of nutrients and overall nutrient status. Total phosphorus concentration exhibits a typical bimodal seasonal fluctuation: the first peak (approximately 80 μg / L) is reached around day 75, and the second peak (approximately 80 μg / L) is reached around day 350; the trough occurs around day 210 (approximately 20 μg / L), with large fluctuations reflecting seasonal differences in nutrient input and consumption. The overall nutrient status index fluctuates in phase with total phosphorus concentration, peaking around days 50 and 230 (approximately 50), and troughing around days 140 and 330 (approximately 30); its fluctuation range is smaller than that of total phosphorus concentration, indicating that it is the result of multiple factors acting together and provides some buffer against changes in nutrient levels. The total nitrogen concentration remained at a very low and stable level throughout the year (around 1 mg / L), with almost no significant fluctuations; it was not significantly correlated with total phosphorus or the comprehensive nutrient status index, indicating that total nitrogen was not a limiting factor for eutrophication during this monitoring period.

[0058] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A monitoring system for eutrophication status of rivers and lakes, characterized in that, include: The data acquisition module is used to acquire raw sensing data from multiple sources in the monitored water area, forming a set of multi-dimensional water body characteristic parameters; The feature fusion module inputs the multidimensional water body feature parameters into a deep feature fusion network, which outputs a high-dimensional fused feature vector through adaptive weighting and nonlinear mapping. The state reasoning module feeds the high-dimensional fused feature vector into the eutrophication state reasoning engine. The eutrophication state reasoning engine embeds a causal graph model trained based on historical cases. The causal graph model defines the probabilistic causal relationship between multidimensional causes and eutrophication representation. The probability calculation module, the eutrophication state inference engine activates relevant nodes and paths in the causal graph model based on the high-dimensional fusion feature vector, and calculates a set of posterior probability distributions of core eutrophication indicators, the core eutrophication indicators including at least total phosphorus simulated concentration, total nitrogen simulated concentration and comprehensive nutrient state index. The map construction module constructs a spatiotemporal evolution map of eutrophication in the monitored water area based on the posterior probability distribution.

2. The river and lake eutrophication status monitoring system as described in claim 1, characterized in that, The acquired raw sensing data of the multi-source heterogeneous water area constitutes a set of multi-dimensional water body characteristic parameters, including: An underwater sensor array deployed in the monitored water area acquires multi-source heterogeneous raw sensing data, which includes at least chlorophyll fluorescence signal, blue-green algae fluorescence signal, dissolved oxygen concentration, turbidity, pH value, water temperature, and incident and emitted light intensity data at a specific wavelength. The original sensing data is subjected to synchronous acquisition and spatiotemporal alignment processing to generate a synchronous dataset with unified timestamps and matching spatial locations; The preliminary estimate of chlorophyll concentration, algal activity indicator, water optical attenuation coefficient, water acid-base state and thermodynamic state are extracted from the synchronous dataset to form a set of multidimensional water characteristic parameters. The process of performing synchronous acquisition and spatiotemporal alignment processing on the original sensing data to generate a synchronous dataset with unified timestamps and matching spatial locations includes: Each sensor in the underwater sensor array is equipped with a high-precision clock module, and the clock is periodically synchronized with a unified time server. In each data acquisition cycle, all sensors are controlled to complete data sampling within a preset synchronization time window, and the same timestamp is added to each sampled data. Obtain the real-time three-dimensional spatial coordinates of each sensor in the underwater sensor array; Based on the real-time three-dimensional spatial coordinates, the sampled data at different spatial locations under the same timestamp are interpolated and mapped to the nodes of the preset rule monitoring grid; All data from different sensors that are mapped to the same monitoring grid node of the aforementioned rule are organized and packaged to form the synchronized dataset.

3. The river and lake eutrophication status monitoring system as described in claim 2, characterized in that, Preliminary estimates of chlorophyll concentration, algal activity indicators, water optical attenuation coefficient, water pH, and thermodynamic state are extracted from the synchronous dataset to form a set of multidimensional water characteristic parameters, including: Dark current correction and temperature effect compensation were performed on the chlorophyll fluorescence signal, and the corrected signal was converted into a preliminary estimate of the chlorophyll concentration using a calibration curve. The fluorescence signal of the blue-green algae is subjected to spectral separation processing to calculate the contribution ratio of fluorescence of a specific algal species, and the ratio of the contribution ratio to the background fluorescence intensity is used as the algal activity indicator value. Based on the turbidity data and the light intensity attenuation data at a specific wavelength, the optical attenuation coefficient of the water body is calculated according to Beer-Lambert's law. The pH value data is directly used as the acid-base state of the water body; The difference between the water temperature data and the historical average water temperature for the same period is taken as the thermodynamic state.

4. The river and lake eutrophication status monitoring system as described in claim 3, characterized in that, The multidimensional water body feature parameters are input into a deep feature fusion network. This network, through adaptive weighting and nonlinear mapping, outputs a high-dimensional fused feature vector, including: Each of the multidimensional water body feature parameters is transformed into an initial feature vector with the same dimension through an independent feature embedding sub-network. Calculate the mutual information estimate between every two initial feature vectors, and construct a feature correlation matrix based on the mutual information estimate; Based on the feature correlation matrix, the fusion weight of each of the multidimensional water feature parameters in the current context is dynamically calculated through an attention mechanism; The initial feature vectors are weighted and summed using the fusion weights to obtain a preliminary fusion vector; The initial fusion vector is input into a multilayer perceptron. The multilayer perceptron performs deep nonlinear transformation and information aggregation on the initial fusion vector through multiple fully connected layers and nonlinear activation functions, and finally outputs the high-dimensional fusion feature vector.

5. The river and lake eutrophication status monitoring system as described in claim 4, characterized in that, The eutrophication state inference engine activates relevant nodes and paths in the causal graph model based on the high-dimensional fused feature vector, and calculates a set of posterior probability distributions for core eutrophication indicators, including: The causal graph model includes observed variable nodes, latent variable nodes, and eutrophication index nodes, wherein the observed variable nodes correspond to the multidimensional water body characteristic parameters. The high-dimensional fusion feature vector is used as evidence and input into the corresponding observation variable node in the causal graph model; In the causal graph model, probabilistic message passing is performed to propagate probabilistic information from the observed variable node along the causal edge to the latent variable node and the eutrophication index node. The joint probability distribution of all nodes in the causal graph model is updated using the belief propagation algorithm. From the updated joint probability distribution, extract the probability distributions corresponding to the nodes representing the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive nutrient status index, and use them as their posterior probability distributions.

6. The river and lake eutrophication status monitoring system as described in claim 5, characterized in that, Based on the aforementioned posterior probability distribution, a spatiotemporal evolution map of eutrophication in the monitored water area is constructed, including: A state inversion algorithm with physical and biochemical constraints is used to jointly optimize and solve the posterior probability distribution and the multidimensional water body characteristic parameters; The state inversion algorithm incorporates the mass conservation equation, empirical relationships of photosynthetic rate, and kinetic constraints of nutrient cycling during the solution process, and iteratively corrects the estimated values ​​of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive nutrient state index. Based on the optimal estimate obtained after iterative convergence, the transient eutrophication level of the monitored water body corresponding to the timestamp is determined. Based on the transient eutrophication levels of continuous time series, a spatiotemporal evolution map of eutrophication in the monitored waters is constructed, and its evolution trend and spatial heterogeneity characteristics are identified.

7. The river and lake eutrophication status monitoring system as described in claim 6, characterized in that, The state inversion algorithm utilizing a physical and biochemical constraint jointly optimizes and solves the posterior probability distribution and the multidimensional water body characteristic parameters, including: The posterior probability distributions of the simulated total phosphorus concentration, the simulated total nitrogen concentration, and the comprehensive nutrient status index are used as the initial inputs for optimization. A joint loss function is constructed, which includes a data fitting term and a physical constraint term. The data fitting term measures the difference between the optical and chemical indicators derived from the estimated nutrient concentration and the multidimensional water body characteristic parameters. The physical constraint term forces the solution to satisfy the mass conservation equation, the empirical relationship of the photosynthetic rate, and the kinetic constraints of the nutrient cycle. The gradient descent method is used to iteratively adjust the estimated values ​​of the total phosphorus concentration, the total nitrogen concentration, and the comprehensive nutrient status index in order to minimize the joint loss function; When the value of the joint loss function changes less than a preset threshold in multiple consecutive iterations, the iteration is determined to be converged, and the current optimal estimate is output.

8. The river and lake eutrophication status monitoring system as described in claim 7, characterized in that, The state inversion algorithm incorporates the mass conservation equation, empirical relationships on photosynthetic rates, and kinetic constraints on nutrient cycling during the solution process, including: The mass conservation equation is an equation in which the rate of change of total phosphorus and total nitrogen within a water volume unit is equal to the sum of the input and output fluxes and the internal source and sink terms. The empirical relationship for photosynthetic rate is a semi-empirical formula for estimating primary productivity rate based on water temperature, light intensity, and nutrient concentration. The kinetic constraints of the nutrient cycle include the adsorption and desorption kinetic equations between dissolved and particulate nutrients, and the nitrification and denitrification reaction rate equations. In each iteration of calculating the physical constraint term of the joint loss function, the current nutrient concentration estimate is substituted into the mass conservation equation, the empirical relationship of the photosynthetic rate, and the kinetic constraints of the nutrient cycle to calculate the predicted theoretical response value. The theoretical response value is compared with the actual response value derived from the multidimensional water body characteristic parameters, and the difference constitutes part of the physical constraint term.

9. The river and lake eutrophication status monitoring system as described in claim 8, characterized in that, The step of determining the transient eutrophication level of the monitored water body at the specified timestamp based on the optimal estimate obtained after iterative convergence includes: The final value of the comprehensive nutritional status index is extracted from the optimal estimate; The final value of the comprehensive nutrient status index is compared with the preset threshold range for eutrophication level classification. Based on the comparison results, the current state of the monitored water area is classified into one of the following levels: oligotrophic, mesotrophic, slightly eutrophic, moderately eutrophic, or severely eutrophic. The level category is the transient eutrophication level. The transient eutrophication level, the corresponding timestamp, and the spatial location information of the monitoring point are bound and stored together.

10. A river and lake eutrophication status monitoring system as described in claim 9, characterized in that, The transient eutrophication levels based on continuous time series are used to construct a spatiotemporal evolution map of eutrophication in the monitored waters, and to identify its evolution trend and spatial heterogeneity characteristics, including: Collect a series of transient eutrophication levels and their associated information on all monitoring grid nodes within a preset monitoring period; Using time as the axis and space as the surface, the transient eutrophication levels at different time points are plotted into a series of spatial distribution raster maps. Arrange the series of spatial distribution raster maps in chronological order to form a spatiotemporal animation or sequence map of eutrophication, namely the spatiotemporal evolution map of eutrophication. Spatiotemporal statistical analysis was performed on the eutrophication spatiotemporal evolution map to calculate the slope of the change in eutrophication level of the entire water body or a specific area over time, which was used as the evolution trend. The variance or spatial autocorrelation index of eutrophication levels at different spatial locations at the same time point is analyzed as a quantitative indicator of the spatial heterogeneity characteristics.

Images