Lake water quality monitoring method and system based on deep learning

Abstract

The invention provides a lake water quality monitoring method and system based on deep learning, and relates to the field of deep learning. The method comprises the following steps: firstly, collecting original sensor data, remote sensing images, hydrodynamic force and meteorological data, carrying out space-time alignment processing, and constructing a lake surface rasterization space-time diagram; the graph is input into a model composed of a spectrum attention encoder, a graph convolutional network and a time sequence unit, lake global water quality index concentration distribution is predicted, a dynamic threshold value is set based on historical data, the standard exceeding condition is recognized in real time, pollution early warning is triggered, and high-precision monitoring and intelligent management of lake water quality are achieved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to a lake water quality monitoring method and system based on deep learning. Background Art

[0002] Lake water quality monitoring relies primarily on two methods: offline manual sampling followed by laboratory analysis, which can simultaneously measure multiple indicators such as COD, total phosphorus, total nitrogen, and chlorophyll-a. The second method involves deploying fixed multi-parameter sensors, such as those for pH, dissolved oxygen, turbidity, and conductivity, to collect data in near real time. While the former offers high accuracy, it is time-consuming, costly, and suffers from low spatiotemporal resolution. While the latter allows for minute-by-minute monitoring, it is limited by the number of locations, sensor drift, and power and communication requirements.

[0003] With the widespread adoption of remote sensing, the Internet of Things (IoT), and edge computing, hyperspectral / multispectral drone and satellite imagery can now cover the entire lake. Low-power 5G / NB-IoT wireless networks increase sensor deployment density, and cloud-edge collaborative platforms meet the demands of large-scale data throughput and model inference. Deep learning, with its strong ability to represent complex nonlinear relationships, is gradually replacing empirical regression and shallow machine learning models for inverting high-precision water quality concentration fields and automatically identifying algal blooms or sudden pollution events.

[0004] At present, most solutions are still limited to fixed-point sensors or single remote sensing data, making it difficult to balance vertical accuracy and area coverage; and traditional convolution or regression models lack comprehensive modeling of lake hydrodynamics, wind fields and pollutant fluxes into the lake, and are prone to failure when faced with seasonal changes in water optical properties; sensor drift, scaling, communication interruptions and on-site calibration requirements increase the operating and maintenance burden. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a lake water quality monitoring method and system based on deep learning, which can achieve high-precision, continuous temporal and spatial monitoring of lake water quality and intelligent pollution early warning, and improve the real-time and scientific nature of water environment management.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A lake water quality monitoring method based on deep learning, comprising:

[0008] Collect multi-source data of the target area; the multi-source data includes raw sensor data, remote sensing image data, hydrodynamic data and meteorological data;

[0009] Performing time synchronization, spatial registration, anomaly processing, and missing value compensation on the multi-source data to form a spatiotemporally aligned feature tensor;

[0010] A space-time graph is established with lake surface grid cells as nodes and hydrodynamic coupling weights as edges;

[0011] The spatiotemporal graph is input into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake.

[0012] If any indicator of the concentration distribution of water quality indicators in the entire lake exceeds the dynamic threshold calculated based on historical monitoring data, a water pollution warning will be triggered.

[0013] Preferably, collecting multi-source data of the target area includes:

[0014] Deploy multi-parameter water quality sensors in the lakes and estuaries in the target area to obtain the raw sensor data in real time; the raw sensor data includes dissolved oxygen, pH, turbidity, conductivity and temperature;

[0015] Use multispectral imaging equipment to obtain remote sensing image data covering the lake surface;

[0016] Collecting hydrodynamic data and meteorological data; the hydrodynamic data includes: flow velocity, flow direction and stratification depth; the meteorological data includes: wind speed, wind direction, rainfall and temperature;

[0017] The original sensor data, the remote sensing image data, the hydrodynamic data and the meteorological data are determined as the multi-source data.

[0018] Preferably, the multi-source data is subjected to time synchronization, spatial registration, anomaly processing and missing value compensation to form a spatiotemporally aligned feature tensor, including:

[0019] Converting the timestamps of the multi-source data into Coordinated Universal Time;

[0020] Resampling or linear interpolation is performed on various types of data in the multi-source data according to a preset time resolution Δt to obtain a first aligned data set D1;

[0021] Project all geographic coordinates involved in D1 to the WGS-84 coordinate system;

[0022] According to the spatial resolution of the lake grid unit Δx×Δy, the remote sensing pixels are reprojected to the nearest neighbor, and the sensor point values ​​are mapped to the corresponding grid to obtain the second aligned dataset D2; where Δx is the spatial resolution of the grid in the longitude direction, and Δy is the spatial resolution of the grid in the latitude direction;

[0023] The 3σ-rule and box plot method were used to detect outliers in D2;

[0024] Perform drift correction based on the median of adjacent moments on the detected outliers, and remove those that cannot be corrected to generate the cleaned data set D3;

[0025] Spline interpolation is used to compensate for time series gaps in D3 with a gap length ≤ K consecutive sampling periods; the maximum continuous time step threshold for gap compensation is:

[0026] For gaps with a length greater than K sampling periods or holes caused by remote sensing cloud cover, spatiotemporal kriging interpolation is used to fill them and obtain the complete dataset D4;

[0027] D4 is stacked into the feature tensor T(t, y, x, c) in a fixed order of time, latitude, longitude, and feature channel; where t is the time index variable, x is the index of the grid in the longitude direction, y is the index of the grid in the latitude direction, and c is the feature channel number.

[0028] Preferably, a spatiotemporal graph is established with lake surface grid cells as nodes and hydrodynamic coupling weights as edges, including:

[0029] Read the lake boundary vector data and perform rasterization with a preset spatial resolution Δx×Δy in the WGS-84 coordinate system to generate a set N of raster cells covering the lake surface;

[0030] Record the center coordinates of each grid cell to obtain the node coordinate table P(N);

[0031] Index the feature tensor T(t,y,x,c), assign the feature vector at time t and whose coordinates fall into each grid cell to the corresponding node, and obtain the node set V(t) with node attributes;

[0032] Calculate the instantaneous velocity vector u at the center of each grid at time t based on the hydrodynamic data i (t);

[0033] For any spatially adjacent nodes i and j, calculate the coupling coefficient Among them, d ij is the node center distance, L is the characteristic diffusion length, θ ij (t) is the angle between the average direction of flow velocity and the direction of the line connecting the nodes;

[0034] Will w ij (t) Normalize to get the weighted adjacency matrix

[0035] right Apply a threshold ε and retain w ij (t)≥ε, and obtain the sparse adjacency matrix A(t), forming a single-time spatial graph Gs(t)=(V(t),A(t));

[0036] Splice T0 consecutive spatial graphs {Gs(t)} into a sequence of temporal graphs according to the time index;

[0037] Between adjacent times t and t+1, a time edge per unit time step is established based on the node identifier, and a fixed weight of 1 is assigned to obtain the complete spatiotemporal graph G st ={G s (t),A temp (t,t+1)}(t=1,…,T0-1); where A temp (·) is the temporal adjacency matrix.

[0038] Preferably, the spatiotemporal graph is input into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake, including:

[0039] A channel attention mechanism is constructed for the remote sensing band data contained in each node in the spatiotemporal graph, an attention weight vector is generated according to the importance of each band to the water quality response, the spectral characteristics of the node are weighted and enhanced, and a node input vector that integrates the attention information is obtained;

[0040] Based on the sparse adjacency matrix A of the spatiotemporal graph and the node input vector, a graph convolution model is constructed to model the spatial structure dependency between nodes and obtain the node spatial feature H (L) ; The update of node features in each layer of graph convolution is expressed as:

[0041]

[0042] in, is the normalized adjacency matrix, H (l) is the node feature representation of the lth layer, W (l) is the trainable weight matrix of this layer, and σ(·) is the activation function;

[0043] The node space feature H (L) The evolution of time steps t=1, 2, ..., T0 constitutes a time series input to the temporal modeling unit, which models the temporal dependency of features through a gated recurrent unit or a long short-term memory network, and outputs a temporally enhanced node representation vector;

[0044] The node representation vector is input into the regression prediction layer, and the concentration values ​​of multiple types of water quality indicators of each node at the current time t are inverted respectively. The output form is the water quality indicator concentration distribution M(t, y, x, v) of the entire lake; where t represents the time step, y and x represent the latitude and longitude position indexes of the lake surface grid, and v represents the type of water quality indicator.

[0045] Preferably, the water quality indicators are chemical oxygen demand (COD), total phosphorus (TP) and chlorophyll-a.

[0046] Preferably, if any indicator of the concentration distribution of water quality indicators in the entire lake exceeds the dynamic threshold calculated based on historical monitoring data, a water pollution warning is triggered, including:

[0047] Obtain historical water quality monitoring data of lakes in the target area over the past N years and construct historical concentration series according to indicator type v;

[0048] Calculate the 95th percentile P of historical concentration for each indicator 95 As the dynamic threshold, and adjusting the correction factor according to the current seasonal conditions, the dynamic threshold function θ is obtained v :

[0049] θ v =P 95 (v)·(1+δ s )

[0050] Among them, δ s is the preset correction coefficient related to the current season, θ v is the dynamic threshold of index v;

[0051] The lake-wide water quality index concentration distribution tensor M(t,y,x,v) is compared with the dynamic threshold θ v Perform channel-by-channel comparison;

[0052] If there exists any index v at any spatial position (y,x) that satisfies:

[0053] M(t,y,x,v)>θ v

[0054] It is determined that the local abnormality of the indicator exceeds the standard;

[0055] Count all the grid points that exceed the standard n v The total number of effective grid points of this indicator N v The ratio of the excess standard is calculated as r v =n v / N v ;

[0056] The warning level is determined based on the excess ratio and the extent of the excess.

[0057] Preferably, the warning level is determined based on the exceeding ratio and the exceeding range, including:

[0058] If r v If the value exceeds 5% and the maximum excess ratio exceeds 1.2, a Level 1 warning will be triggered;

[0059] If r v >10% or the maximum excess multiple is >1.5, a second-level warning is triggered;

[0060] If r v >20% or if any point exceeds the standard by more than 2.0 times, a level 3 warning will be triggered.

[0061] A lake water quality monitoring system based on deep learning, including:

[0062] A data acquisition unit, configured to acquire multi-source data of a target area; the multi-source data includes raw sensor data, remote sensing image data, hydrodynamic data, and meteorological data;

[0063] A data preprocessing unit, configured to perform time synchronization, spatial registration, anomaly processing, and missing value compensation on the multi-source data to form a spatiotemporally aligned feature tensor;

[0064] A graph structure construction unit is used to build a spatiotemporal graph using lake surface grid cells as nodes and hydrodynamic coupling weights as edges;

[0065] a water quality prediction modeling unit, configured to input the spatiotemporal graph into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake;

[0066] The early warning judgment and release unit is used to trigger a water pollution warning if any indicator of the concentration distribution of water quality indicators in the entire lake exceeds the dynamic threshold calculated based on historical monitoring data.

[0067] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0068] The present invention constructs a unified spatiotemporal aligned feature tensor by fusing raw sensor data, remote sensing image data, hydrodynamic data and meteorological data, and establishes a dynamic spatiotemporal graph structure based on lake surface grid units, which can effectively characterize the spatiotemporal distribution characteristics of lake water quality. It further introduces a deep learning model composed of a spectral attention encoder, a graph convolutional network and a time series unit, which improves the expression ability and modeling accuracy of multi-source heterogeneous data in water quality prediction. At the same time, it dynamically calculates water quality thresholds based on historical monitoring data to realize the identification of exceeding standards and early warning triggering of key water quality indicators. It has high spatial resolution, high response efficiency and intelligent risk identification capabilities, which is conducive to supporting accurate monitoring and rapid intervention of lake water environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0070] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0071] Figure 2 A schematic diagram of the system structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0073] The purpose of this invention is to provide a lake water quality monitoring method and system based on deep learning, which can realize high-precision, spatiotemporal continuous monitoring of lake water quality and intelligent pollution early warning, and improve the real-time and scientific nature of water environment management.

[0074] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] Figure 1 A flow chart of the method provided in the embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a lake water quality monitoring method based on deep learning, comprising:

[0076] Step 100: Collect multi-source data of the target area; the multi-source data includes raw sensor data, remote sensing image data, hydrodynamic data and meteorological data;

[0077] Step 200: Perform time synchronization, spatial registration, anomaly processing, and missing value compensation on multi-source data to form a spatiotemporally aligned feature tensor;

[0078] Step 300: Establish a spatiotemporal graph using lake surface grid cells as nodes and hydrodynamic coupling weights as edges;

[0079] Step 400: Input the spatiotemporal graph into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake.

[0080] Step 500: If any indicator of the concentration distribution of water quality indicators in the entire lake exceeds the dynamic threshold calculated based on historical monitoring data, a water pollution warning is triggered.

[0081] In step 100 of this embodiment, preferably, multi-parameter water quality sensors are deployed in the main body of the lake and its estuaries to achieve real-time monitoring of key indicators such as dissolved oxygen (DO), pH, turbidity, conductivity and temperature. The sensor layout positions are optimized in combination with the water exchange path and the pollution input channel to ensure a high-time response capability to the dynamic changes in water quality in key areas. Unlike traditional single-point sampling, this solution forms spatial coverage through multi-point deployment, and at the same time complements remote sensing image information to provide a dense observation basis for the subsequent construction of a spatiotemporal graph structure. The raw sensor data is directly involved in the construction process of the feature tensor after being collected with a unified timestamp. It is one of the core time-varying elements that drive deep model learning and is irreplaceable.

[0082] In terms of obtaining remote sensing information of the lake surface, this embodiment preferably uses an aviation or UAV platform equipped with a multispectral sensor to obtain image data including multiple bands such as red, green, blue, and near-infrared, and performs geometric correction in combination with ground control points. The remote sensing image covers the entire surface of the water area and has high spatial continuity; at the same time, hydrodynamic data (including flow velocity, flow direction and stratification depth) and meteorological data (including wind speed, wind direction, rainfall and temperature) are further introduced as key physical inputs for edge weight calculation and water quality change driving factors in the mapping process. The above four types of data sources together constitute the multi-source data system in the present invention, which supports the construction of a spatiotemporal alignment feature tensor that integrates spectral features, sensor features and environmental field driving factors, ensuring the integrity and characterization ability of the input structure of the deep learning model, which is directly related to the model reasoning accuracy and early warning reliability.

[0083] In step 200 of this preferred embodiment, first, by uniformly converting all data timestamps to Coordinated Universal Time (UTC), and resampling or linearly interpolating various types of data with a fixed step size of Δt, a first aligned dataset D1 is constructed to achieve cross-device and cross-platform temporal consistency. Subsequently, all spatial coordinates are uniformly projected to the WGS-84 coordinate system, and the lake surface is divided into regular grids in the longitude direction Δx and the latitude direction Δy according to the accuracy requirements of water body monitoring. On this common grid, the remote sensing pixels are subjected to nearest neighbor reprojection, and the sensor observations are assigned to the corresponding grids using a spatial mapping function to generate a second aligned dataset D2, thereby fusing point and surface observation information within a unified spatial framework, laying the foundation for subsequent spatiotemporal modeling.

[0084] In response to possible abnormal fluctuations and missing measurements in D3, this solution uses the "3σ-rule + box plot" dual discrimination to eliminate extreme outliers, and uses the median of adjacent moments for drift correction. Records that cannot be corrected are directly eliminated to form a cleaned dataset D3. Cubic spline interpolation is used for missing openings in D3 (continuous missing ≤K time steps) to maintain local trend continuity; for spatial holes caused by long gaps or cloud cover, spatiotemporal kriging interpolation is introduced to estimate missing values ​​using the neighborhood covariance structure and output the complete dataset D4. Finally, D4 is stacked into a feature tensor T(t,y,x,c) in the four-dimensional order of tyxc to ensure strict alignment of the data in time, space, and multi-channel feature dimensions, providing high-quality, same-scale input for the deep learning model, significantly improving the accuracy and stability of subsequent water quality predictions.

[0085] In step 300 of the preferred embodiment, first, the vector data of the lake boundary is read, and a regular rasterization operation is performed on the lake area at a preset spatial resolution in the WGS-84 coordinate system to generate a set of two-dimensional grid cells covering the water area. The center coordinates of each grid cell are recorded as node coordinates to form a spatial node index table; then, based on the constructed feature tensor, the feature vectors of each time point and spatial position are indexed and matched, and assigned to the corresponding grid nodes, thereby establishing a node set containing spatiotemporal attributes. In this process, the spatial resolution is determined by the ground pixel size of the remote sensing image (such as 50 meters to 100 meters), and the time index comes from the unified time step after the alignment of multiple source data (such as every 10 minutes) to ensure the consistency of the spatial and temporal dimension structures.

[0086] Furthermore, the instantaneous velocity vector at each grid center at each time step is obtained based on hydrodynamic observation data. This data is typically obtained from acoustic current meters deployed at the lake center or estuary and interpolated to the entire lake grid. For any pair of spatially adjacent grid nodes, the degree of coupling between them is calculated based on physical factors such as the distance between them, the velocity intensity, and the angle between the flow direction and the line connecting the nodes. This serves as the edge weight of the spatial graph. This coupling weight is normalized, and a threshold for edge weight is set to filter out low-strength connections, resulting in a sparse spatial adjacency relationship. This constructs a spatial graph at a single time point. Finally, multiple consecutive spatial graphs are concatenated in chronological order using a temporal index, and temporal edges are established between nodes at the same location in adjacent time steps. This creates a complete spatiotemporal graph structure that incorporates both spatial diffusion structure and temporal evolution. The diffusion long scale is derived from the spatial correlation analysis of historical velocity history, and the edge weight threshold is determined by a combination of graph connectivity and model performance. These parameters are derived from statistical derivation of real-world observational data and fine-tuned during model training.

[0087] In step 400 of this preferred embodiment, a channel attention mechanism is first constructed for the remote sensing image band data contained in each node in the spatiotemporal graph. By analyzing the sensitivity of different bands to water quality indicators, a band attention weight vector is generated. This weight is used to weight the node's multispectral features, thereby enhancing its responsiveness to key band information (such as red edge, near-infrared, and other reflectances related to algae concentration), forming the node input vector for fused spectral attention. This input not only includes remote sensing features but can also incorporate raw sensor data or environmental variables, enabling node features to express optical, water quality, and hydrodynamic characteristics while possessing enhanced discriminative power.

[0088] Subsequently, based on the sparse adjacency matrix and the constructed node input vector, a graph convolutional network is used to encode and model the spatial structure information. The graph convolution operation learns the diffusion and transmission relationship between nodes in different grids of the lake surface layer by layer through weighted aggregation of the node's own features and the features of its neighboring nodes. The node feature representation output by each layer of graph convolution is updated through a trainable weight matrix and a nonlinear activation function, ultimately obtaining a node spatial feature representation with spatial dependency information. This spatial representation can describe the mutual influence and propagation trend of water quality indicators between different grids, providing basic support for capturing regional pollution trends.

[0089] After obtaining the spatial features of the nodes, they are organized in time series and input into a time series modeling unit, such as a gated recurrent unit (GRU) or a long short-term memory network (LSTM), to model the evolution of the nodes under continuous time steps. This time series unit can effectively retain the dynamic trend information of water quality changes at different nodes, and output a node representation vector containing time series memory features. Finally, the representation vector is input into the regression prediction layer to predict the concentration value of the water quality index at the current moment node by node, forming a multivariate concentration distribution map of the entire lake. The water quality indicators are preferably chemical oxygen demand (COD), total phosphorus (TP) and chlorophyll-a, and can be extended to other water quality parameters such as total nitrogen, ammonia nitrogen or cyanobacteria density according to actual monitoring needs. The distribution map has continuity in the time and space dimensions, and can provide reliable support for water quality assessment and decision-making response.

[0090] In step 500 of this preferred embodiment, to achieve intelligent identification and graded early warning of abnormal lake water quality conditions, historical monitoring data for several consecutive years (e.g., five years) is first obtained within the target area. Concentration change sequences are constructed for different water quality indicators, such as chemical oxygen demand, total phosphorus, and chlorophyll-a. The 95th percentile of each indicator is calculated using statistical methods as a baseline threshold. To enhance the threshold's adaptability to seasonal environmental changes, a correction coefficient is introduced to dynamically adjust the threshold. This correction coefficient can be set based on the current hydrological period (e.g., flood season, dry season) or factors that affect water quality fluctuations, such as temperature and light, thereby obtaining a dynamic threshold function that reflects the current background conditions.

[0091] After the model outputs the concentration distribution of water quality indicators across the entire lake, it is compared channel by channel with the dynamic thresholds described above. If any indicator is found to exceed the corresponding threshold at any spatial location, it is determined to be a local exceedance event for that indicator. The proportion of grid cells exceeding the standard for that indicator across the entire lake surface is further calculated to obtain the exceedance ratio. Combining the maximum exceedance multiple and this ratio, a warning level can be divided: when the exceedance ratio is greater than 5% and the maximum exceedance multiple exceeds 1.2, a level 1 warning is triggered; if the exceedance ratio exceeds 10% or the maximum exceedance multiple exceeds 1.5, a level 2 warning is triggered; if the exceedance ratio exceeds 20% or if any point exceeds 2.0, a level 3 warning is triggered. The warning information will automatically generate structured data for push and visual display, enabling timely response to the risk of sudden changes in water quality.

[0092] Corresponding to the above method, such as Figure 2 As shown, this embodiment also provides a lake water quality monitoring system based on deep learning, including:

[0093] A data acquisition unit, configured to acquire multi-source data of a target area; the multi-source data includes raw sensor data, remote sensing image data, hydrodynamic data, and meteorological data;

[0094] A data preprocessing unit, configured to perform time synchronization, spatial registration, anomaly processing, and missing value compensation on the multi-source data to form a spatiotemporally aligned feature tensor;

[0095] A graph structure construction unit is used to build a spatiotemporal graph using lake surface grid cells as nodes and hydrodynamic coupling weights as edges;

[0096] a water quality prediction modeling unit, configured to input the spatiotemporal graph into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake;

[0097] The early warning judgment and release unit is used to trigger a water pollution warning if any indicator of the concentration distribution of water quality indicators in the entire lake exceeds the dynamic threshold calculated based on historical monitoring data.

[0098] The beneficial effects of the present invention are as follows:

[0099] (1) This invention constructs a multi-source data system with unified spatiotemporal scales by fusing raw sensor data, remote sensing imagery, hydrodynamic information, and meteorological data. This effectively overcomes the technical bottlenecks of existing water quality monitoring methods, such as low spatial resolution, single observation elements, and weak information fusion capabilities. It provides a solid data foundation for comprehensive perception and modeling of lake water quality. In particular, the complementary use of remote sensing and sensor data allows for capturing large-scale trends in lake surface changes while retaining the ability to detect localized mutations with high sensitivity.

[0100] (2) This paper innovatively introduces a spectral attention mechanism, a graph convolutional structure, and a temporal neural network to construct an end-to-end deep learning model. This not only improves the ability to fit the complex nonlinear evolution of water quality indicators, but also strengthens the ability to jointly model the spatial diffusion structure and temporal evolution laws. By establishing a dynamic spatiotemporal graph, the internal diffusion process of the lake driven by hydrodynamics is explicitly introduced into the modeling framework, achieving an effective combination of structural priors and data-driven methods, significantly improving prediction accuracy and model generalization.

[0101] (3) In terms of anomaly detection and risk warning, this paper proposes a dynamic threshold construction method based on historical monitoring statistical characteristics and seasonal adjustment factors, which solves the problems of fixed threshold methods with high false alarm rates and poor adaptability in multiple seasons and hydrological periods. Combining a dual-index evaluation mechanism of spatial coverage and exceedance multiples, this method achieves quantitative determination of the degree of water quality anomaly and multi-level response, effectively enhancing the practicality and management operability of the system.

[0102] (4) The present invention realizes a closed loop of the entire process from multi-source data fusion, spatiotemporal modeling, intelligent prediction to dynamic early warning. It has high precision, high robustness and strong scalability. It can be widely used in water quality monitoring and ecological safety assessment of lakes, reservoirs and other water bodies, providing scientific support and technical guarantee for water environment governance, pollution tracing and emergency response.

[0103] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0104] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A lake water quality monitoring method based on deep learning, characterized in that: include: Collect multi-source data of the target area; the multi-source data includes raw sensor data, remote sensing image data, hydrodynamic data and meteorological data; Performing time synchronization, spatial registration, anomaly processing, and missing value compensation on the multi-source data to form a spatiotemporally aligned feature tensor; A space-time graph is established with lake surface grid cells as nodes and hydrodynamic coupling weights as edges; The spatiotemporal graph is input into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake. If any indicator of the concentration distribution of water quality indicators in the entire lake exceeds the dynamic threshold calculated based on historical monitoring data, a water pollution warning will be triggered.

2. The lake water quality monitoring method based on deep learning according to claim 1 is characterized in that Collect multi-source data of the target area, including: Deploy multi-parameter water quality sensors in the lakes and estuaries in the target area to obtain the raw sensor data in real time; the raw sensor data includes dissolved oxygen, pH, turbidity, conductivity and temperature; Use multispectral imaging equipment to obtain remote sensing image data covering the lake surface; Collecting hydrodynamic data and meteorological data; the hydrodynamic data includes: flow velocity, flow direction and stratification depth; the meteorological data includes: wind speed, wind direction, rainfall and temperature; The original sensor data, the remote sensing image data, the hydrodynamic data and the meteorological data are determined as the multi-source data.

3. The lake water quality monitoring method based on deep learning according to claim 1 is characterized in that, The multi-source data is subjected to time synchronization, spatial registration, anomaly processing, and missing value compensation to form a spatiotemporally aligned feature tensor, including: Converting the timestamps of the multi-source data into Coordinated Universal Time; Resampling or linear interpolation is performed on various types of data in the multi-source data according to a preset time resolution Δt to obtain a first aligned data set D1; Project all geographic coordinates involved in D1 to the WGS-84 coordinate system; According to the spatial resolution of the lake grid unit Δx×Δy, the remote sensing pixels are reprojected to the nearest neighbor, and the sensor point values ​​are mapped to the corresponding grid to obtain the second aligned dataset D2; where Δx is the spatial resolution of the grid in the longitude direction, and Δy is the spatial resolution of the grid in the latitude direction; The 3σ-rule and box plot method were used to detect outliers in D2; Perform drift correction based on the median of adjacent moments on the detected outliers, and remove those that cannot be corrected to generate the cleaned data set D3; Spline interpolation is used to compensate for time series gaps in D3 with a gap length ≤ K consecutive sampling periods; the maximum continuous time step threshold for gap compensation is: For gaps with a length greater than K sampling periods or holes caused by remote sensing cloud cover, spatiotemporal kriging interpolation is used to fill them and obtain the complete dataset D4; D4 is stacked into the feature tensor T(t, y, x, c) in a fixed order of time, latitude, longitude, and feature channel; where t is the time index variable, x is the index of the grid in the longitude direction, y is the index of the grid in the latitude direction, and c is the feature channel number.

4. The lake water quality monitoring method based on deep learning according to claim 1 is characterized in that, A space-time graph is established with lake surface grid cells as nodes and hydrodynamic coupling weights as edges, including: Read the lake boundary vector data and perform rasterization with a preset spatial resolution Δx×Δy in the WGS-84 coordinate system to generate a set N of raster cells covering the lake surface; Record the center coordinates of each grid cell to obtain the node coordinate table P(N); Index the feature tensor T(t,y,x,c), assign the feature vector at time t and whose coordinates fall into each grid cell to the corresponding node, and obtain the node set V(t) with node attributes; Calculate the instantaneous flow velocity vector u at the center of each grid at time t based on the hydrodynamic data i (t); For any spatially adjacent nodes i and j, calculate the coupling coefficient Among them, d ij is the node center distance, L is the characteristic diffusion length, θ ij (t) is the angle between the average direction of flow velocity and the direction of the line connecting the nodes; Will w ij (t) Normalize to get the weighted adjacency matrix right Apply a threshold ε and retain w ij (t)≥ε, and obtain the sparse adjacency matrix A(t), forming a single-time spatial graph Gs(t)=(V(t),A(t)); Splice T0 consecutive spatial graphs {Gs(t)} into a sequence of temporal graphs according to the time index; Between adjacent times t and t+1, a time edge per unit time step is established based on the node identifier, and a fixed weight of 1 is assigned to obtain the complete spatiotemporal graph G st ={G s (t),A temp (t,t+1)}t=1,…,T0-1); where A temp (·) is the temporal adjacency matrix.

5. The lake water quality monitoring method based on deep learning according to claim 4 is characterized in that, The spatiotemporal graph is input into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake, including: A channel attention mechanism is constructed for the remote sensing band data contained in each node in the spatiotemporal graph, an attention weight vector is generated according to the importance of each band to the water quality response, the spectral characteristics of the node are weighted and enhanced, and a node input vector that integrates the attention information is obtained; Based on the sparse adjacency matrix A of the spatiotemporal graph and the node input vector, a graph convolution model is constructed to model the spatial structure dependency between nodes and obtain the node spatial feature H (L) ; The update of node features in each layer of graph convolution is expressed as: in, is the normalized adjacency matrix, H (l) is the node feature representation of the lth layer, W (l) is the trainable weight matrix of this layer, and σ(·) is the activation function; The node space feature H (L) The evolution of time steps t=1, 2, ..., T0 constitutes a time series input to the temporal modeling unit, which models the temporal dependency of features through a gated recurrent unit or a long short-term memory network, and outputs a temporally enhanced node representation vector; The node representation vector is input into the regression prediction layer, and the concentration values ​​of multiple types of water quality indicators of each node at the current time t are inverted respectively. The output form is the water quality indicator concentration distribution M(t, y, x, v) of the entire lake; where t represents the time step, y and x represent the latitude and longitude position indexes of the lake surface grid, and v represents the type of water quality indicator.

6. The lake water quality monitoring method based on deep learning according to claim 5 is characterized in that: The types of water quality indicators include chemical oxygen demand (COD), total phosphorus (TP) and chlorophyll-a.

7. The lake water quality monitoring method based on deep learning according to claim 5 is characterized in that: If any of the indicators in the concentration distribution of water quality indicators in the entire lake area exceeds the dynamic threshold calculated based on historical monitoring data, a water pollution warning will be triggered, including: Obtain historical water quality monitoring data of lakes in the target area over the past N years and construct historical concentration series according to indicator type v; Calculate the 95th percentile P of historical concentration for each indicator 95 As the dynamic threshold, and adjusting the correction factor according to the current seasonal conditions, the dynamic threshold function θ is obtained v : i v =P 95 (v)·(1+δ s ) Among them, δ s is the preset correction coefficient related to the current season, θ v is the dynamic threshold of index v; The lake-wide water quality index concentration distribution tensor M(t,y,x,v) is compared with the dynamic threshold θ v Perform channel-by-channel comparison; If there exists any index v at any spatial position (y,x) that satisfies: M(t,y,x,v)>θ v It is determined that the local abnormality of the indicator exceeds the standard; Count all the grid points that exceed the standard n v The total number of effective grid points of this indicator N v The ratio of the excess standard is calculated as r v =n v / N v ; The warning level is determined based on the excess ratio and the extent of the excess.

8. The lake water quality monitoring method based on deep learning according to claim 7 is characterized in that: The warning level is determined based on the excess ratio and extent, including: If r v If the value exceeds 5% and the maximum excess ratio exceeds 1.2, a Level 1 warning will be triggered; If r v >10% or the maximum excess multiple is >1.5, a second-level warning is triggered; If r v >20% or if any point exceeds the standard by more than 2.0 times, a level 3 warning will be triggered.

9. A lake water quality monitoring system based on deep learning, characterized in that: include: A data acquisition unit, configured to acquire multi-source data of a target area; the multi-source data includes raw sensor data, remote sensing image data, hydrodynamic data, and meteorological data; A data preprocessing unit, configured to perform time synchronization, spatial registration, anomaly processing, and missing value compensation on the multi-source data to form a spatiotemporally aligned feature tensor; A graph structure construction unit is used to build a spatiotemporal graph using lake surface grid cells as nodes and hydrodynamic coupling weights as edges; a water quality prediction modeling unit, configured to input the spatiotemporal graph into an end-to-end model consisting of a spectral attention encoder, a graph convolutional network, and a temporal unit to obtain the concentration distribution of water quality indicators across the entire lake; The early warning judgment and release unit is used to trigger a water pollution warning if any indicator of the concentration distribution of water quality indicators in the entire lake exceeds the dynamic threshold calculated based on historical monitoring data.

Images