A river and lake water ecological environment intelligent monitoring method based on multi-source perception data fusion

Abstract

This invention relates to the field of river and lake aquatic ecological environment monitoring technology, specifically to a smart monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion, comprising the following steps: S1, collecting multi-source data and completing anomaly removal, temporal alignment, and spatial registration; S2, extracting and fusing water quality temporal, upstream and downstream spatial, and remote sensing features; S3, using D-S evidence, graph convolution, and dynamic weighting for quality assessment to obtain the fusion result; S4, determining the water quality level, predicting trends and indicators, issuing warnings and generating reports when anomalies occur; S5, adding new data and incrementally updating parameters. This invention improves the accuracy, stability, timely warning, and long-term adaptability of river and lake aquatic ecological environment monitoring through spatiotemporal alignment, feature fusion, adaptive weight adjustment, trend prediction, and incremental updates of multi-source monitoring data.

Description

Technical Field

[0001] This invention relates to the field of river and lake water ecological environment monitoring technology, and in particular to a smart monitoring method for river and lake water ecological environment based on multi-source sensing data fusion. Background Technology

[0002] With the continuous advancement of river and lake ecological environment protection and comprehensive watershed management, water ecological environment monitoring has become an important technical means for water resource management, water pollution prevention and control, and ecological restoration. Currently, river and lake water ecological environment monitoring typically utilizes water quality sensors, hydrological monitoring equipment, meteorological monitoring equipment, and remote sensing monitoring platforms to continuously monitor water body physicochemical indicators, hydrological conditions, and the surrounding environment. Combined with data analysis, this enables water quality assessment, pollution early warning, and prediction of ecological environment change trends, providing data support for river and lake ecological environment governance and scientific decision-making.

[0003] However, most existing river and lake aquatic ecological environment monitoring technologies are designed for single types of monitoring data or employ simple data fusion methods. Data from different sources often exhibit variations in sampling frequency, spatial resolution, and data quality, hindering the effective collaborative utilization of multi-source monitoring data. Furthermore, they lack adaptive adjustment mechanisms to address dynamic changes in the reliability of different data sources, making them susceptible to anomalies, missing data, and noise interference, resulting in insufficient accuracy and stability of the fusion results. In addition, existing monitoring systems generally lack the ability to continuously learn from historical monitoring information, making it difficult to adapt to seasonal changes and long-term evolution of the river and lake aquatic ecological environment. The accuracy of predicting water quality trends and providing early warnings of anomalies still needs improvement. Summary of the Invention

[0004] This invention provides a smart monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion. By uniformly preprocessing water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data, and combining multimodal feature extraction, deep fusion of multi-source data, adaptive fusion weight adjustment, water quality level determination, trend prediction, and incremental updates, it realizes intelligent monitoring and dynamic early warning of river and lake aquatic ecological environment, improves the accuracy of multi-source monitoring data fusion, the stability of monitoring results, the timeliness of anomaly early warning, and the system's adaptability to long-term changes in river and lake aquatic ecological environment.

[0005] A smart monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion includes the following steps:

[0006] S1. Collect monitoring data of the target water area, including water quality data, hydrological data, meteorological data, satellite remote sensing data and UAV aerial survey data. Use the local outlier factor algorithm to remove abnormal data, use the dynamic time warping algorithm to align the data of different sampling frequencies in time, and use interpolation and coordinate transformation to perform spatial registration to obtain preprocessed multi-source sensing data.

[0007] S2, based on the preprocessed multi-source sensing data, extract the water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features of each monitoring point, and use a cross-modal attention mechanism to map the water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features to a unified embedding space to generate a fused spatiotemporal feature representation.

[0008] S3. Based on the fusion spatiotemporal feature representation, the DS evidence theory is used to synthesize evidence for water quality data, hydrological data and meteorological data of the same monitoring point. A graph convolutional network is used to spatially propagate and aggregate the fusion results of adjacent monitoring points. The quality assessment score is calculated according to the data missing rate, signal-to-noise ratio and timeliness of each data source. The fusion weight of each data source is dynamically adjusted to obtain the deep fusion result of multi-source data.

[0009] S4. Based on the deep fusion results of the multi-source data, the water quality level is determined, and the water quality change trend and key indicator values ​​in the future time period are predicted according to the deep fusion results of historical multi-source data. When the monitoring indicators exceed the preset threshold or the prediction results show a trend of water quality deterioration, an early warning is triggered and a dynamic monitoring report is generated.

[0010] S5 adds newly collected monitoring data to the training set and incrementally updates the relevant parameters used to extract water quality temporal evolution characteristics, upstream and downstream spatial neighborhood characteristics, and water body remote sensing characteristics to generate fused spatiotemporal feature representations, as well as the relevant parameters used for evidence synthesis, spatial propagation and aggregation, and fusion weight adjustment.

[0011] Optionally, S1 includes:

[0012] S11, water quality sensors, hydrological sensors, and meteorological sensors are deployed in the target water area. The water quality sensors are used to collect data on pH value, dissolved oxygen, turbidity, conductivity, ammonia nitrogen, and total phosphorus. The hydrological sensors are used to collect data on water level, flow velocity, and flow rate. The meteorological sensors are used to collect data on air temperature, air pressure, wind speed, and light intensity. At the same time, satellite remote sensing data covering the target water area and UAV aerial survey data are accessed, and the collection time, collection location, and data source identifiers are added to each monitoring data to generate the original multi-source sensing dataset of the target water area.

[0013] S12, Based on the original multi-source sensing dataset, a local neighborhood is constructed for the continuous monitoring data collected by each sensor, and the local outlier factor algorithm is used to calculate the local outlier factor value of each monitoring data. Monitoring data with local outlier factor values ​​greater than the anomaly judgment threshold are marked as abnormal data and removed to obtain multi-source sensing data after anomaly removal.

[0014] S13, based on the multi-source sensing data after anomaly removal, taking a unified target time series as a benchmark, a dynamic time warping algorithm is used to calculate the cumulative matching distance between data sequences with different sampling frequencies and the target time series, and the time alignment path is determined according to the minimum cumulative matching distance, and each data sequence is mapped to a unified time node to obtain time-aligned multi-source sensing data;

[0015] S14. Based on the time-aligned multi-source sensing data, a standardized spatial grid covering the target water area is established. According to the spatial distance between each monitoring point and the standardized spatial grid node, the water quality data, hydrological data and meteorological data are interpolated and calculated to map the data of each monitoring point to the corresponding standardized spatial grid node.

[0016] S15, the spatial coordinates of the satellite remote sensing data and UAV aerial survey data are transformed to a target coordinate system consistent with the standardized spatial grid. The transformed satellite remote sensing data and UAV aerial survey data are then matched with the interpolated water quality data, hydrological data and meteorological data according to their spatial location to obtain preprocessed multi-source sensing data.

[0017] Optionally, S2 includes:

[0018] S21, Based on the preprocessed multi-source sensing data, water quality time series are constructed according to monitoring points and collection time. The pH value, dissolved oxygen, turbidity, conductivity, ammonia nitrogen and total phosphorus data of each monitoring point at continuous time are input into a bidirectional long short-term memory network. The water quality change process is encoded along the forward and reverse time directions respectively. The forward hidden state and the reverse hidden state are spliced ​​together to obtain the water quality time series evolution characteristics of each monitoring point. The water quality time series evolution characteristics include trend characteristics, periodic characteristics and abrupt change characteristics.

[0019] S22, construct a spatiotemporal relationship graph of the target water area based on the preprocessed multi-source sensing data, take each monitoring point as a node, and establish spatial edges according to the geographical proximity and water flow direction correlation between monitoring points. Take the water quality temporal evolution characteristics of each monitoring point as node features and input them into the graph attention network. Calculate the attention coefficient between the target monitoring point and its neighboring monitoring points, and perform weighted aggregation of the node features of neighboring monitoring points according to the attention coefficient to obtain the upstream and downstream spatial neighborhood features that characterize the propagation relationship of upstream and downstream water quality changes.

[0020] S23, Based on the preprocessed satellite remote sensing data and UAV aerial survey data, band operations are performed on the multispectral remote sensing image covering the target water area to extract the normalized water index, chlorophyll a concentration inversion value and suspended matter concentration inversion value. The normalized water index, chlorophyll a concentration inversion value and suspended matter concentration inversion value are combined to form a remote sensing feature image. Convolutional encoding is performed through a convolutional neural network to obtain the water remote sensing features of the target water area.

[0021] S24, the water quality temporal evolution features, upstream and downstream spatial neighborhood features, and water body remote sensing features are linearly projected respectively, and the correlation weights between different modal features are calculated using a cross-modal attention mechanism. Based on the correlation weights, the projected water quality temporal evolution features, upstream and downstream spatial neighborhood features, and water body remote sensing features are weighted and fused to generate a fused spatiotemporal feature representation.

[0022] Optionally, S3 includes:

[0023] S31, based on the fused spatiotemporal feature representation, water quality data, hydrological data and meteorological data corresponding to the same monitoring point are obtained respectively. The water quality data, hydrological data and meteorological data are used as independent evidence to determine the water quality level. The conflict coefficient and trust function between each independent evidence are calculated using the DS evidence theory. The independent evidence is synthesized according to the conflict coefficient and trust function to obtain the comprehensive judgment result of each monitoring point.

[0024] S32, based on the constructed spatiotemporal relationship graph, uses the comprehensive discrimination results of each monitoring point as node features, and constructs a graph convolutional network according to the geographical adjacency relationship and water flow direction relationship between monitoring points. The graph convolutional network performs spatial propagation and feature aggregation on the comprehensive discrimination results of each monitoring point, so that water quality information between adjacent monitoring points can be transferred to each other, and the spatially smoothed water quality distribution state can be obtained.

[0025] S33. Based on the spatially smoothed water quality distribution, the missing data rate, signal-to-noise ratio, and timeliness of water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data are statistically analyzed. The quality assessment score of each data source is calculated based on the missing data rate, signal-to-noise ratio, and timeliness. A gating mechanism is used to dynamically adjust the weight of each data source in the fusion process according to the quality assessment score of each data source. When the quality assessment score of any data source decreases, the fusion weight of the corresponding data source is automatically reduced to obtain the deep fusion result of multi-source data.

[0026] Optionally, S31 includes:

[0027] S311, Based on the fused spatiotemporal feature representation, water quality data, hydrological data and meteorological data corresponding to the same monitoring point are obtained respectively, and according to the water quality level judgment criteria corresponding to each type of data, the basic probability assignment of water quality data, hydrological data and meteorological data to each water quality level is calculated, and the basic probability assignment of each type of data to the water quality level is used as water quality evidence, hydrological evidence and meteorological evidence respectively, forming an independent evidence set for the same monitoring point;

[0028] S312, Based on the set of independent evidence, the DS evidence theory is used to calculate the conflict coefficient between each independent piece of evidence and the trust function corresponding to each water quality level, and to determine the degree of consistency between different independent pieces of evidence.

[0029] S313, Based on the conflict coefficient and the trust function, perform DS evidence synthesis on the water quality evidence, hydrological evidence and meteorological evidence, calculate the comprehensive basic probability assignment for each water quality level, and take the judgment result corresponding to the water quality level with the largest comprehensive basic probability assignment as the comprehensive judgment result for the monitoring point.

[0030] Optionally, S32 includes:

[0031] S321, based on the constructed spatiotemporal relationship graph, each monitoring point is used as a graph node, and the geographical adjacency relationship and water flow direction relationship between monitoring points are used as graph edges. The comprehensive judgment result of each monitoring point is converted into a node feature vector, and a node feature matrix and an adjacency matrix are constructed. The node feature vector is composed of the comprehensive basic probability assignment corresponding to each water quality level, and the adjacency matrix is ​​used to characterize whether there is a geographical adjacency relationship or a water flow direction relationship between any two monitoring points.

[0032] S322, add self-connection relationships for each monitoring point based on the adjacency matrix, and normalize the adjacency matrix after adding self-connection to obtain a normalized adjacency matrix. Input the node feature matrix and the normalized adjacency matrix into the graph convolutional network so that each monitoring point receives comprehensive discrimination information from its adjacent upstream monitoring points, downstream monitoring points and geographically adjacent monitoring points, thus completing the spatial propagation of water quality information between monitoring points.

[0033] S323, using a graph convolutional network, the node features of each monitoring point and its neighboring monitoring points are weighted and aggregated according to the normalized adjacency matrix. The aggregation results are then subjected to feature transformation and nonlinear mapping to obtain the spatial fusion features of each monitoring point. The spatial fusion features of each monitoring point are then mapped to the standardized grid of the target water area according to their spatial location to form a spatially smoothed water quality distribution state.

[0034] Optionally, S33 includes:

[0035] S331, based on the spatially smoothed water quality distribution state, extract the data records of water quality data, hydrological data, meteorological data, satellite remote sensing data and UAV aerial survey data within the current evaluation period, and count the amount of data to be collected, the amount of effective data, the signal power, the noise power and the data update time of each data source, and calculate the data missing rate, signal-to-noise ratio and timeliness of each data source to form a set of quality indicators for each data source.

[0036] S332, based on the data missing rate, signal-to-noise ratio and timeliness of each data source, the signal-to-noise ratio is normalized, and the data integrity, normalized signal-to-noise ratio and timeliness are weighted and calculated to obtain the quality assessment score corresponding to each data source. Among them, the higher the data missing rate, the lower the data integrity, the higher the signal-to-noise ratio and timeliness, the higher the quality assessment score corresponding to the data source.

[0037] S333, the quality assessment scores corresponding to each data source are input into the gating mechanism to calculate the gating value of each data source in the current fusion process, and the gating value is normalized to obtain the fusion weight of each data source. Based on the fusion weight, the feature representations corresponding to water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data are weighted and fused, and combined with the spatially smoothed water quality distribution state to obtain the deep fusion result of multi-source data. When the quality assessment score of any data source decreases, its gating value and fusion weight decrease accordingly.

[0038] Optionally, S4 includes:

[0039] S41, Based on the deep fusion results of the multi-source data, the deep fusion results of the multi-source data corresponding to each monitoring point are input into the trained multi-class support vector machine, and the classification score of the water quality corresponding to the deep fusion results of the multi-source data is calculated respectively. The water quality category with the largest classification score is determined as the water quality level of the corresponding monitoring point, and the water quality level determination result of the target water area is formed.

[0040] S42, extract the deep fusion results of historical multi-source data within multiple consecutive monitoring periods according to the monitoring time sequence, construct a sequence of deep fusion results of historical multi-source data, input the sequence of deep fusion results of historical multi-source data into a time-series prediction model based on the Transformer architecture, calculate the degree of temporal correlation between historical monitoring periods through self-attention, and perform feature aggregation based on the degree of temporal correlation of each historical monitoring period to predict the water quality change trend and key indicator prediction values ​​of each monitoring point within a predetermined time period in the future. Key indicators include pH value, dissolved oxygen, turbidity, conductivity, ammonia nitrogen and total phosphorus.

[0041] S43, compare the current key indicator monitoring values ​​and the predicted key indicator values ​​within a predetermined future time period of each monitoring point with the corresponding preset indicator thresholds, and determine whether there is a water quality deterioration trend based on the predicted changes of the key indicators. When any current key indicator monitoring value exceeds the corresponding preset indicator threshold, or any key indicator predicted value exceeds the corresponding preset indicator threshold within a predetermined future time period, or the prediction result shows a water quality deterioration trend, trigger an early warning signal, and generate a dynamic monitoring report including a spatiotemporal evolution diagram and trend analysis chart based on the water quality level determination result, the predicted key indicator values, and the early warning result.

[0042] Optionally, S43 includes:

[0043] S431, acquire the monitoring values ​​of pH, dissolved oxygen, turbidity, conductivity, ammonia nitrogen and total phosphorus at each monitoring point at the current monitoring time, as well as the predicted values ​​of the corresponding key indicators in the future predetermined time period, compare the current monitoring values ​​of key indicators and the predicted values ​​of key indicators at each prediction time with the corresponding preset indicator thresholds, determine whether each key indicator exceeds the limit at present and whether it will exceed the predicted limit in the future predetermined time period, and obtain the current indicator exceedance judgment result and the predicted indicator exceedance judgment result;

[0044] S432, calculate the predicted changes of each key indicator relative to the current monitoring value at each monitoring point in the future predetermined time period, and compare the predicted increase of turbidity, conductivity, ammonia nitrogen and total phosphorus, the predicted decrease of dissolved oxygen and the deviation of the predicted pH value from the normal range with the corresponding deterioration judgment thresholds according to the water quality impact direction corresponding to each key indicator. When any key indicator meets the corresponding deterioration judgment condition, the key indicator is marked as a deterioration indicator. When at least two of the multiple key indicators are marked as deterioration indicators, it is determined that there is a water quality deterioration trend at the corresponding monitoring point.

[0045] S433, combining the current indicator exceeding the limit judgment result, the predicted indicator exceeding the limit judgment result, and the water quality deterioration trend judgment result, when the current monitoring value of any key indicator exceeds the corresponding preset indicator threshold, or the predicted value of any key indicator exceeds the corresponding preset indicator threshold within a predetermined time period in the future, or when there is a water quality deterioration trend at the corresponding monitoring point, an early warning signal is triggered. The water quality level judgment result, the predicted value of the key indicator, and the early warning signal are correlated according to the monitoring time and the spatial location of the monitoring point to generate a spatiotemporal evolution map and trend analysis chart including the spatial distribution of water quality level, the spatiotemporal changes of key indicators, the predicted trend of key indicators, and the early warning location, forming a dynamic monitoring report.

[0046] Optionally, S5 includes:

[0047] S51. Acquire newly collected water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data according to the predetermined update cycle. Perform preprocessing using abnormal data removal, time-series alignment, and spatial registration. Then, according to the monitoring point, monitoring time, and standardized spatial grid, associate the preprocessed new monitoring data with the corresponding actual results of water quality level and actual values ​​of key indicators to form a new training sample batch. Add the new training sample batch to the original training set to obtain the updated training set.

[0048] S52, the newly added training samples are sequentially input into the processing process for extracting water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features and generating fused spatiotemporal feature representations, as well as the processing process for evidence synthesis, spatial propagation and aggregation and fusion weight adjustment, to obtain the multi-source data deep fusion results corresponding to the newly added training samples. The level determination loss is calculated based on the actual water quality level results and water quality level determination results of the newly added training samples, and the indicator prediction loss is calculated based on the actual values ​​and predicted values ​​of key indicators. The level determination loss and indicator prediction loss are weighted to obtain the incremental update loss of the current update cycle.

[0049] S53, based on the incremental update loss, calculate the loss gradient of the first parameter set used to extract water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features and generate fused spatiotemporal feature representations, and the loss gradient of the second parameter set used for evidence synthesis, spatial propagation and aggregation and fusion weight adjustment. Update the first parameter set and the second parameter set according to the corresponding incremental learning rate, and use the updated parameters for multimodal spatiotemporal feature extraction and multi-source data deep fusion in the next monitoring cycle.

[0050] The beneficial effects of this invention are:

[0051] This invention constructs a multi-source sensing system composed of water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data. By combining anomaly data removal, temporal alignment, and spatial registration, it achieves unified preprocessing of monitoring data from different sources, with different sampling frequencies, and different spatial resolutions. This effectively eliminates the impact of anomalies and spatiotemporal differences on monitoring results. Simultaneously, by extracting water quality temporal evolution characteristics, upstream and downstream spatial neighborhood characteristics, and water body remote sensing characteristics, and fusing them using a cross-modal attention mechanism, it can fully explore the temporal variation patterns, spatial propagation patterns, and remote sensing information characteristics of river and lake aquatic ecological environments, improve the efficiency of collaborative utilization of multi-source data, and provide more comprehensive and accurate fused characteristics for subsequent ecological environment monitoring.

[0052] This invention employs the DS evidence theory to synthesize evidence from water quality, hydrological, and meteorological data. It also combines graph convolutional networks to achieve spatial propagation and aggregation of water quality information between monitoring points. Furthermore, it dynamically calculates fusion weights based on the data missing rate, signal-to-noise ratio, and timeliness of each data source, enabling adaptive fusion of different quality data sources. This effectively reduces the impact of uncertainties and conflicts between different data sources on monitoring results, improves the reliability, stability, and spatial continuity of deep fusion results of multi-source data, and makes the monitoring results of river and lake water ecological environment more consistent with actual water body changes.

[0053] This invention determines water quality levels based on deep fusion of multi-source data and predicts water quality change trends and key indicator predictions within a predetermined time period by combining historical monitoring data. It comprehensively analyzes current monitoring values, predicted values, and water quality deterioration trends to achieve early warning of aquatic ecological environment anomalies and automatically generates dynamic monitoring reports. At the same time, by continuously adding new monitoring data to the training set, the relevant parameters used for multimodal feature extraction and multi-source data fusion are incrementally updated, enabling the monitoring model to continuously adapt to the seasonal changes and long-term evolution of river and lake aquatic ecological environments, thereby improving the prediction accuracy, environmental adaptability, and intelligent monitoring level of the system during long-term operation. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a schematic diagram of the monitoring method according to an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram illustrating the generation of multi-source data deep fusion results according to an embodiment of the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0058] like Figures 1-2 As shown, a smart monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion includes the following steps:

[0059] S1. Collect monitoring data of the target water area, including water quality data, hydrological data, meteorological data, satellite remote sensing data and UAV aerial survey data. Use the local outlier factor algorithm to remove abnormal data, use the dynamic time warping algorithm to align the data of different sampling frequencies in time, and use interpolation and coordinate transformation to perform spatial registration to obtain preprocessed multi-source sensing data.

[0060] S2, based on the preprocessed multi-source sensing data, extracts the water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features of each monitoring point, and uses a cross-modal attention mechanism to map the water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features to a unified embedding space to generate a fused spatiotemporal feature representation.

[0061] S3, based on the fusion of spatiotemporal feature representation, uses DS evidence theory to synthesize evidence from water quality data, hydrological data and meteorological data of the same monitoring point, uses graph convolutional network to spatially propagate and aggregate the fusion results of adjacent monitoring points, and calculates quality assessment scores based on the data missing rate, signal-to-noise ratio and timeliness of each data source, dynamically adjusts the fusion weight of each data source, and obtains the deep fusion result of multi-source data.

[0062] S4 determines the water quality level based on the deep fusion results of multi-source data, and predicts the water quality change trend and key indicator values ​​in the future within a predetermined time period based on the deep fusion results of historical multi-source data. When the monitoring indicators exceed the preset threshold or the prediction results show a trend of water quality deterioration, an early warning is triggered and a dynamic monitoring report is generated.

[0063] S5 adds newly collected monitoring data to the training set and incrementally updates the relevant parameters used to extract water quality temporal evolution characteristics, upstream and downstream spatial neighborhood characteristics, and water body remote sensing characteristics to generate fused spatiotemporal feature representations, as well as the relevant parameters used for evidence synthesis, spatial propagation and aggregation, and fusion weight adjustment.

[0064] S1 includes:

[0065] S11. Water quality sensors, hydrological sensors, and meteorological sensors are deployed in the target water area. The water quality sensors are used to collect data on pH value, dissolved oxygen, turbidity, conductivity, ammonia nitrogen, and total phosphorus. The hydrological sensors are used to collect data on water level, flow velocity, and flow rate. The meteorological sensors are used to collect data on air temperature, air pressure, wind speed, and light intensity. At the same time, satellite remote sensing data covering the target water area and UAV aerial survey data are connected. The collection time, collection location, and data source identifiers are added to each monitoring data to generate the original multi-source sensing dataset of the target water area.

[0066] S12, based on the original multi-source sensing dataset, a local neighborhood is constructed for the continuous monitoring data collected by each sensor, and the local outlier factor algorithm is used to calculate the local outlier factor value of each monitoring data. Monitoring data with local outlier factor values ​​greater than the anomaly judgment threshold are marked as outlier data and removed, resulting in the multi-source sensing data after anomaly removal, represented as:

[0067] ;

[0068] in, For monitoring data The local outlier value, For monitoring data of nearest neighbor set for The number of data in the nearest neighbor set For monitoring data Locally achievable density, For monitoring data Locally achievable density, when At that time, determine the monitoring data This is abnormal data. This is the threshold for anomaly detection;

[0069] ;

[0070] in, , These represent the mean and standard deviation of the local outlier values ​​for all monitored data. This is the threshold adjustment coefficient;

[0071] S13, based on the multi-source sensing data after anomaly removal, and using a unified target time series as a benchmark, a dynamic time warping algorithm is used to calculate the cumulative matching distance between data sequences with different sampling frequencies and the target time series. The time alignment path is then determined based on the minimum cumulative matching distance, mapping each data sequence to a unified time node, resulting in time-aligned multi-source sensing data, represented as follows:

[0072] ;

[0073] in, For data sequences The Middle Data and target time series The Middle The cumulative distance between data points For monitoring data With target data Local distance between them The cumulative distance in the vertical direction, The cumulative distance in the horizontal direction, The path with the smallest cumulative distance in the diagonal direction is used as the time-aligned path.

[0074] S14. Based on the time-aligned multi-source sensing data, a standardized spatial grid covering the target water area is established. According to the spatial distance between each monitoring point and the standardized spatial grid node, interpolation calculations are performed on the water quality data, hydrological data, and meteorological data, mapping the data from each monitoring point to the corresponding standardized spatial grid node, as shown below:

[0075] ;

[0076] ;

[0077] in, For standardized spatial grid nodes The corresponding interpolation results, For the first Monitoring data from each monitoring point For the first Each monitoring point corresponds to a standardized spatial grid node. Distance weights, The number of monitoring points participating in the interpolation calculation. For the first The monitoring point and the first Spatial distance between standardized spatial grid nodes The distance decay exponent;

[0078] S15. The spatial coordinates of the satellite remote sensing data and UAV aerial survey data are transformed to a target coordinate system consistent with the standardized spatial grid. Then, the transformed satellite remote sensing data and UAV aerial survey data are matched with the interpolated water quality data, hydrological data, and meteorological data according to their spatial location, resulting in preprocessed multi-source sensing data. The coordinate transformation relationship is expressed as follows:

[0079] ;

[0080] in, , The coordinates of satellite remote sensing data or UAV aerial survey data in the original coordinate system. , These are the coordinates after transformation to the target coordinate system. , , , These are the parameters of the coordinate transformation matrix, used to represent rotation, scaling, and orientation transformations. , These are the coordinate translation parameters.

[0081] S2 includes:

[0082] S21. Based on the preprocessed multi-source sensing data, water quality time series are constructed according to monitoring points and collection times. The pH, dissolved oxygen, turbidity, conductivity, ammonia nitrogen, and total phosphorus data at each monitoring point are input into a bidirectional long short-term memory network. The water quality change process is encoded along both the forward and reverse time axes. The forward and reverse hidden states are concatenated to obtain the water quality time series evolution characteristics of each monitoring point. These characteristics include trend features, periodic features, and abrupt change features, expressed as follows:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] in, For the first The monitoring point at the 1st Water quality data vector at each time point It is a positive long short-term memory network. It is an inverse long short-term memory network. Positive hidden state, In reverse hidden state, For the first The monitoring point at the 1st The bidirectional encoding results at each time point, This is a vector concatenation operation. The length of the time series. To aggregate the encoding results at each time point, For the first The temporal evolution characteristics of water quality at each monitoring point;

[0088] S22. Based on the preprocessed multi-source sensing data, a spatiotemporal relationship graph of the target water area is constructed. Each monitoring point is used as a node, and spatial edges are established according to the geographical adjacency and water flow direction correlation between monitoring points. The water quality temporal evolution characteristics of each monitoring point are used as node features input to the graph attention network. The attention coefficient between the target monitoring point and its neighboring monitoring points is calculated, and the node features of neighboring monitoring points are weighted and aggregated according to the attention coefficient to obtain the upstream and downstream spatial neighborhood features representing the propagation relationship of upstream and downstream water quality changes, as shown below:

[0089] ;

[0090] ;

[0091] ;

[0092] in, For the first The monitoring point and the first Attention score between monitoring points For the first The time-series evolution characteristics of water quality at each monitoring point The node feature transformation matrix, For attention parameter vectors, This is the transpose of the attention parameter vector. It is a non-linear activation function. For the first The monitoring point for the first Attention coefficient at each monitoring point In order to be with the first A set of monitoring points that are geographically adjacent or related by water flow direction. The monitoring point number is the number of the monitoring point in the monitoring point set. For activation function, For the first The upstream and downstream spatial neighborhood characteristics of each monitoring point;

[0093] S23. Based on preprocessed satellite remote sensing data and UAV aerial survey data, band operations are performed on the multispectral remote sensing image covering the target water area to extract the normalized water index, chlorophyll a concentration inversion value, and suspended matter concentration inversion value. These values ​​are then combined to form a remote sensing feature image, which is then convolutionally encoded using a convolutional neural network to obtain the water remote sensing features of the target water area, represented as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] in, The normalized water index, The reflectance of the green band in multispectral remote sensing images. The reflectivity is in the near-infrared band. This refers to the chlorophyll a concentration inversion value obtained from remote sensing image band calculations. The suspended matter concentration inversion value is obtained from the calculation of remote sensing image bands. The image is composed of remote sensing features derived from normalized water index, chlorophyll a concentration, and suspended matter concentration. This refers to the encoding process of a convolutional neural network. Features of water bodies as remote sensing characteristics;

[0098] The encoding process of a convolutional neural network can be represented as follows:

[0099] ;

[0100] The initial input is: ;

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] in, For the first Feature maps output by each convolutional layer For the first The convolutional kernel weights of each convolutional layer For the first The bias parameters of each convolutional layer It is a non-linear activation function. The feature map is the initial input. For the first The first convolutional layer Each output channel is in position eigenvalues ​​at that location For the first Feature maps output by each convolutional layer For the next level The feature values ​​of each channel within the corresponding convolution region. For the first The connection in the nth convolutional layer The input channel and the first The convolutional kernel weights for each output channel, For the first The bias parameters corresponding to each output channel. , These are the height and width of the convolution kernel, respectively. For the first The number of channels in each convolutional layer , These are the row and column indices within the convolution kernel, respectively. It is a non-linear activation function. For the first In the first pooling layer Each channel is located in The pooling results at that point For pooled window size, For pooling step size, The convolutional feature values ​​within the pooling window, For the first The remote sensing feature values ​​obtained by global average pooling of each channel This is the last convolutional layer in the convolutional neural network. , These represent the height and width of the last layer feature map, respectively. This represents the number of channels in the last layer of the feature map. This is a water body remote sensing feature vector composed of remote sensing feature values ​​from each channel;

[0106] S24. The temporal evolution features of water quality, upstream and downstream spatial neighborhood features, and remote sensing features of water bodies are linearly projected to give the three types of features the same feature dimension. A cross-modal attention mechanism is then used to calculate the correlation weights between different modal features. Based on these correlation weights, the projected temporal evolution features of water quality, upstream and downstream spatial neighborhood features, and remote sensing features of water bodies are weighted and fused to generate a fused spatiotemporal feature representation, expressed as:

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] in, The characteristics of water quality evolution over time. Characteristics of upstream and downstream spatial neighborhoods, For remote sensing features of water bodies, , , These are, respectively, the temporal embedding features, spatial embedding features, and remote sensing embedding features mapped to a unified embedding space. , , These are the projection matrices for the three types of features, , , These are the corresponding bias vectors. , These are the query modality and the followed modality, respectively. , , , These are the query matrix, key matrix, and value matrix, respectively. , , These are parameter matrices for query, key, and value, respectively. To query the feature dimensions of the vector and key vector, For modality For modes Cross-modal attention weights, For modality Fusion mode Features following information For feature concatenation operations, This is the final generated spatiotemporal feature representation.

[0117] S3 includes:

[0118] S31. Based on the fusion of spatiotemporal feature representation, water quality data, hydrological data and meteorological data corresponding to the same monitoring point are obtained respectively. The water quality classification results of water quality data, hydrological data and meteorological data are used as independent evidence. The conflict coefficient and trust function between the independent evidence are calculated by using DS evidence theory. Based on the conflict coefficient and trust function, the independent evidence is synthesized to obtain the comprehensive classification result of each monitoring point, so as to reduce the impact of uncertainty and conflict between different data sources on the water quality judgment result.

[0119] S32, based on the constructed spatiotemporal relationship graph, uses the comprehensive discrimination results of each monitoring point as node features, and constructs a graph convolutional network according to the geographical adjacency relationship and water flow direction relationship between monitoring points. The graph convolutional network performs spatial propagation and feature aggregation on the comprehensive discrimination results of each monitoring point, so that water quality information between adjacent monitoring points can be transferred to each other, and the spatially smoothed water quality distribution state can be obtained to characterize the spatial continuity of water quality changes in the upstream and downstream of the target water area.

[0120] S33, based on the spatially smoothed water quality distribution, statistically analyzes the missing data rate, signal-to-noise ratio, and timeliness of water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data. It then calculates the quality assessment score for each data source based on these metrics. A gating mechanism is employed to dynamically adjust the weight of each data source during the fusion process according to its corresponding quality assessment score. When the quality assessment score of any data source decreases, the fusion weight of the corresponding data source is automatically reduced, resulting in a deep fusion result of the multi-source data.

[0121] S31 includes:

[0122] S311, based on the fusion of spatiotemporal feature representations, water quality data, hydrological data, and meteorological data corresponding to the same monitoring point are acquired separately. According to the water quality grade determination criteria corresponding to each type of data, the basic probability assignments of the water quality data, hydrological data, and meteorological data to each water quality grade are calculated. These basic probability assignments of each type of data to the water quality grade are then used as water quality evidence, hydrological evidence, and meteorological evidence, respectively, forming an independent set of evidence for the same monitoring point, represented as:

[0123] ;

[0124] in, For the first Class data on water quality levels Basic probability assignment, For the first Class of data on water quality levels The discrimination score, For the first Each water quality level Number of water quality grades;

[0125] Water quality level Based on the Class I, Class II, Class III, Class IV, and Class V water quality standards corresponding to each water quality indicator in the "Surface Water Environmental Quality Standard" (GB 3838), water quality grade sets are established respectively, represented as follows:

[0126] ;

[0127] in, The water quality is classified as Class I. The water quality is Class II. The water quality is Class III. The water quality is Class IV. The water quality is Class V, as described in this embodiment. ;

[0128] ;

[0129] in, For the first The data corresponds to the actual monitored values ​​of the monitoring indicators. For the first The data corresponds to the monitoring indicators in the water quality level. The standard reference value in For the first The scale parameters of the monitoring indicators corresponding to the data categories;

[0130] S312, based on a set of independent evidence, uses the DS evidence theory to calculate the conflict coefficient between each independent piece of evidence and the trust function corresponding to each water quality level, determining the degree of consistency between different independent pieces of evidence, providing a basis for evidence synthesis, expressed as follows:

[0131] ;

[0132] ;

[0133] in, The conflict coefficient between two sets of independent evidence. For the first set of independent evidence against the proposition Basic probability assignment, For the second set of independent evidence against the proposition Basic probability assignment, , These are two propositions, The two propositions are mutually exclusive. For the proposition Trust function, For the proposition Basic probability assignment, For the proposition Belongs to the proposition ;

[0134] S313, based on the conflict coefficient and confidence function, performs DS evidence synthesis on water quality evidence, hydrological evidence, and meteorological evidence, calculates the comprehensive basic probability assignment for each water quality level, and takes the judgment result corresponding to the water quality level with the highest comprehensive basic probability assignment as the comprehensive discrimination result for that monitoring point, expressed as:

[0135] ;

[0136] ;

[0137] in, Proposition after evidence synthesis The overall basic probability assignment, For the proposition With proposition Jointly support the proposition , The water quality level is determined by comprehensively assessing the results.

[0138] S32 includes:

[0139] S321, based on the constructed spatiotemporal relationship graph, each monitoring point is treated as a graph node, and the geographical adjacency and water flow direction correlation between monitoring points are treated as graph edges. The comprehensive judgment results of each monitoring point are converted into node feature vectors, and a node feature matrix and an adjacency matrix are constructed. The node feature vector is composed of the comprehensive basic probability assignment corresponding to each water quality level. The adjacency matrix is ​​used to characterize whether there is a geographical adjacency or water flow direction correlation between any two monitoring points, and is expressed as follows:

[0140] ;

[0141] ;

[0142] ;

[0143] in, For the first The node feature vector of each monitoring point The node feature matrix, For the number of monitoring points, For the first The monitoring point and the first The graph connection relationship between monitoring points, when there is a geographical adjacency or a correlation in water flow direction between them. ,otherwise ;

[0144] S322: Self-connection relationships are added to each monitoring point based on the adjacency matrix, and the adjacency matrix after adding self-connections is normalized to obtain a normalized adjacency matrix. The node feature matrix and the normalized adjacency matrix are input into the graph convolutional network, so that each monitoring point receives comprehensive discrimination information from its upstream, downstream, and geographically adjacent monitoring points, completing the spatial propagation of water quality information between monitoring points, represented as:

[0145] ;

[0146] ;

[0147] ;

[0148] ;

[0149] in, This is the adjacency matrix between monitoring points. To increase the adjacency matrix after self-connection, It is the identity matrix. For the first The node degree corresponding to each monitoring point For the normalized adjacency matrix, It is a degree matrix;

[0150] S323 utilizes a graph convolutional network to weighted aggregate the node features of each monitoring point and its neighboring monitoring points based on a normalized adjacency matrix. The aggregation results are then subjected to feature transformation and nonlinear mapping to obtain the spatial fusion features of each monitoring point. These spatial fusion features are then mapped to a standardized grid of the target water area according to their spatial location, forming a spatially smoothed water quality distribution, represented as follows:

[0151] ;

[0152] ;

[0153] ;

[0154] in, , The first , The node feature matrix of a layered graph convolutional network. For the first The feature transformation matrix of the layer, For the first Layer bias parameters, It is a non-linear activation function. For initial node features, The spatially fused features are obtained from each monitoring point after spatial propagation and feature aggregation, representing the spatially smoothed water quality distribution. This is the feature matrix of the nodes in the last layer. denoted as the total number of layers in the graph convolutional network.

[0155] S33 includes:

[0156] S331, based on the spatially smoothed water quality distribution, extracts data records from water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data within the current assessment period. Calculates the required data volume, effective data volume, signal power, noise power, and data update time for each data source. Calculates the data missing rate, signal-to-noise ratio, and timeliness for each data source, forming a set of quality indicators for each data source, expressed as:

[0157] ;

[0158] ;

[0159] ;

[0160] in, These are the data source serial numbers, corresponding to water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data, respectively. For the first Data missing rate of each data source For the current evaluation period, the first The effective data volume of each data source For the current evaluation period, the first The amount of data to be collected from each data source. For the first Signal-to-noise ratio of each data source For the first The effective average power of the signal from each data source For the first The average noise power of each data source. For the first The timeliness of each data source For the current assessment time and the first The time interval between the most recent updates of each data source For the first The reference update cycle for each data source;

[0161] S332, based on the data missing rate, signal-to-noise ratio (SNR), and timeliness of each data source, normalizes the SNR and weights the data completeness, normalized SNR, and timeliness to obtain the quality assessment score for each data source. A higher data missing rate, lower data completeness, higher SNR, and higher timeliness result in a higher quality assessment score for the data source, as expressed below:

[0162] ;

[0163] ;

[0164] ;

[0165] in, For the first Data integrity of each data source For the first Normalized signal-to-noise ratio of each data source , These represent the minimum and maximum signal-to-noise ratios of each data source. To prevent positive numbers with a denominator of zero, For the first Quality assessment scores for each data source , , These are the evaluation coefficients for data completeness, normalized signal-to-noise ratio, and timeliness, respectively.

[0166] S333: Input the quality assessment scores corresponding to each data source into the gating mechanism, calculate the gating value of each data source in the current fusion process, and normalize the gating values ​​to obtain the fusion weight of each data source. Based on the fusion weight, perform weighted fusion on the feature representations corresponding to water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data, and combine it with the spatially smoothed water quality distribution status to obtain the deep fusion result of multi-source data. When the quality assessment score of any data source decreases, its gating value and fusion weight decrease accordingly, as shown below:

[0167] ;

[0168] ;

[0169] ;

[0170] ;

[0171] in, For the first The gate value of each data source, For the first The gating adjustment coefficient of each data source, For the first Gating bias parameters of each data source for The fusion weight of each data source To integrate the data source sequence numbers in the weight normalization process, Total number of data sources For the first Feature representations corresponding to each data source The quality adaptive fusion features are obtained by weighting each data source according to the fusion weight. For feature concatenation operations, This is the result of deep fusion of multi-source data.

[0172] S4 includes:

[0173] S41, based on the deep fusion results of multi-source data, the deep fusion results of multi-source data corresponding to each monitoring point are input into the trained multi-class support vector machine. The classification score of the water quality corresponding to the deep fusion results of multi-source data is calculated respectively, and the water quality category with the highest classification score is determined as the water quality level of the corresponding monitoring point, forming the water quality level determination result of the target water area, expressed as:

[0174] ;

[0175] ;

[0176] in, For the first Each monitoring point corresponds to a water quality level. Category score, For the first The deep fusion results of multi-source data from multiple monitoring points Water quality grades in multi-class support vector machines The corresponding classification parameter vector, This is the transpose of the classification parameter vector. Water quality level The corresponding classification bias, For the first Water quality rating results for each monitoring point;

[0177] S42. Historical multi-source data deep fusion results are extracted from multiple consecutive monitoring periods according to the monitoring time sequence to construct a historical multi-source data deep fusion result sequence. This sequence is then input into a time-series prediction model based on the Transformer architecture. The degree of temporal correlation between historical monitoring periods is calculated through self-attention, and features are aggregated based on the degree of temporal correlation of each historical monitoring period to predict the water quality change trend and key indicator prediction values ​​for each monitoring point within a predetermined future time period. Key indicators include pH, dissolved oxygen, turbidity, conductivity, ammonia nitrogen, and total phosphorus, expressed as:

[0178] ;

[0179] ;

[0180] ;

[0181] ;

[0182] ;

[0183] ;

[0184] ;

[0185] ;

[0186] in, For the first Each monitoring point at the current predicted time Previous historical multi-source data deep fusion result sequence, For the first The monitoring point at the 1st The results of deep fusion of multi-source data at the monitoring time. To input the number of historical monitoring periods for the time series prediction model, , , These are the query matrix, key matrix, and value matrix in the trend prediction process. , , These are the query mapping matrix, key mapping matrix, and value mapping matrix of the trend prediction model, respectively. This is the temporal attention matrix between historical monitoring periods. For the feature dimensions of query vectors and key vectors during trend prediction, These are the temporal aggregated features obtained through temporal correlation calculation. For the first The monitoring point in the future Predicted values ​​of key indicators at each prediction time. Future No. The output mapping matrix corresponding to each prediction time. This is the corresponding output bias vector. This refers to the predicted time sequence number within the predetermined time period. For the first The first monitoring point Key indicators in the future The predicted change at each prediction time relative to the current time. These are the predicted values ​​for the corresponding key indicators. To determine the current monitoring value of the corresponding key indicators, based on the positive or negative direction of the predicted change and the direction of the water quality impact of the corresponding key indicators, the water quality improvement trend, stable trend, or deterioration trend is determined.

[0187] S43 compares the current monitoring values ​​of key indicators and the predicted values ​​of key indicators within a predetermined time period at each monitoring point with the corresponding preset indicator thresholds. It then determines whether there is a trend of water quality deterioration based on the predicted changes in the key indicators. When any current monitoring value of a key indicator exceeds the corresponding preset indicator threshold, or any predicted value of a key indicator exceeds the corresponding preset indicator threshold within a predetermined time period, or when the prediction results show a trend of water quality deterioration, an early warning signal is triggered. Based on the water quality level determination results, the predicted values ​​of key indicators, and the early warning results, a dynamic monitoring report including a spatiotemporal evolution diagram and trend analysis charts is generated.

[0188] S43 includes:

[0189] S431, acquire the pH, dissolved oxygen, turbidity, conductivity, ammonia nitrogen, and total phosphorus monitoring values ​​of each monitoring point at the current monitoring time, as well as the predicted values ​​of the corresponding key indicators within a predetermined future time period. Compare the current key indicator monitoring values ​​and the predicted values ​​of the key indicators at each prediction time with the corresponding preset indicator thresholds to determine whether each key indicator currently exceeds the limit and whether predicted exceedance will occur within the predetermined future time period. Obtain the current indicator exceedance determination result and the predicted indicator exceedance determination result, specifically including:

[0190] (1) For pH values ​​with interval constraints, the first The current exceedance determination results for each monitoring point are as follows:

[0191] ;

[0192] No. The predicted pH values ​​exceeding the limit at each monitoring point are as follows:

[0193] ;

[0194] (2) For turbidity, conductivity, ammonia nitrogen, and total phosphorus, which are subject to upper limit constraints, the first The first monitoring point The current exceedance results for the key indicators are as follows:

[0195] ;

[0196] The corresponding prediction over-limit judgment result is:

[0197] ;

[0198] (3) For dissolved oxygen with a lower limit constraint, the first The current exceedance determination results for each monitoring point are as follows:

[0199] ;

[0200] The corresponding prediction over-limit judgment result is:

[0201] ;

[0202] in, This ensures that at least one predicted time within a future predetermined time period satisfies the corresponding condition. This ensures that all predicted times within a future predetermined time period meet the corresponding conditions. For the first The pH value of each monitoring point at the current monitoring time. For the first The monitoring point in the future Predicted pH value at each predicted time point. For the first The first monitoring point The monitoring values ​​of key indicators at the current monitoring time. For the first The first monitoring point Key indicators in the future The predicted value at each prediction time. For the first The dissolved oxygen monitoring value at each monitoring point at the current monitoring time. For the first The monitoring point in the future Predicted dissolved oxygen values ​​at each prediction time. For the first Current exceedance determination results of pH values ​​at each monitoring point For the first Predicted exceedance results of pH values ​​at each monitoring point For the first The first monitoring point The current exceedance determination results of key indicators subject to upper limit constraints. For the first The first monitoring point The prediction results of the key indicators subject to upper limit constraints. For the first The current results of dissolved oxygen exceeding the limit at each monitoring point. For the first The predicted exceedance results of dissolved oxygen at each monitoring point. This represents the lower limit threshold of the normal pH range. This represents the upper limit threshold of the normal pH range. For the first The key indicator of the item adopts the upper limit constraint and the corresponding preset upper limit threshold. hour, ,when If the target body of water is a river, then If the target water area is a lake or reservoir, then ,when hour, , , These represent the average turbidity and standard deviation over historical stable monitoring periods, respectively. hour, , , These represent the average conductivity and standard deviation over a historical stable monitoring period, respectively. This is the preset lower limit threshold for dissolved oxygen. If the water quality level is Class I, If the water quality level is Class II, If the water quality level is Class III, If the water quality level is Class IV, If the water quality level is Class V, ;

[0203] S432, calculate the predicted changes of each key indicator relative to the current monitoring value at each monitoring point within a predetermined future time period. Based on the water quality impact direction corresponding to each key indicator, compare the predicted increases of turbidity, conductivity, ammonia nitrogen, and total phosphorus, the predicted decrease of dissolved oxygen, and the deviation of the predicted pH value from the normal range with the corresponding deterioration judgment thresholds. When any key indicator meets the corresponding deterioration judgment condition, that key indicator is marked as a deteriorating indicator. When at least two of the multiple key indicators are marked as deteriorating indicators, it is determined that the corresponding monitoring point has a water quality deterioration trend, specifically including:

[0204] (1) For turbidity, conductivity, ammonia nitrogen, and total phosphorus, the deterioration markers for the corresponding indicators are:

[0205] ;

[0206] ;

[0207] (2) The deterioration of dissolved oxygen is marked by:

[0208] ;

[0209] ;

[0210] (3) pH value in the future The deviation of each predicted time point from the normal range is:

[0211] ;

[0212] Deterioration of pH value is marked as:

[0213] ;

[0214] ;

[0215] (4) No. The number of deterioration indicators at each monitoring point is:

[0216] ;

[0217] (5) The result of the water quality deterioration trend assessment is as follows: ;

[0218] in, For the first The first monitoring point A key indicator of deterioration, For the first The threshold for judging the deterioration of key indicators. As a marker of deterioration in dissolved oxygen, The threshold for determining the deterioration of dissolved oxygen. This represents the deviation of the predicted pH value from the normal range. A marker of pH deterioration. The threshold for determining pH deterioration. For the first The number of key indicators showing a deteriorating trend at each monitoring point The result of the water quality deterioration trend assessment. This indicates a trend of water quality deterioration. This is a marker of deteriorating turbidity. As a marker of deterioration in electrical conductivity, As a marker of ammonia nitrogen deterioration, As a marker of deterioration in total phosphorus, , The mean and standard deviation of the change in the direction of deterioration. , These represent the mean and standard deviation of the change in the direction of dissolved oxygen deterioration, respectively. , These are the mean and standard deviation of the change in pH value in the direction of deterioration, respectively.

[0219] S433, combining the current indicator exceedance judgment results, the predicted indicator exceedance judgment results, and the water quality deterioration trend judgment results, when the current monitoring value of any key indicator exceeds the corresponding preset indicator threshold, or the predicted value of any key indicator exceeds the corresponding preset indicator threshold within a predetermined time period in the future, or when there is a water quality deterioration trend at the corresponding monitoring point, an early warning signal is triggered. The water quality level judgment results, the predicted values ​​of key indicators, and the early warning signal are correlated according to the monitoring time and the spatial location of the monitoring point to generate a spatiotemporal evolution map and trend analysis chart including the spatial distribution of water quality level, the spatiotemporal changes of key indicators, the predicted trend of key indicators, and the early warning location, forming a dynamic monitoring report.

[0220] S5 includes:

[0221] S51. According to the predetermined update cycle, newly collected water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data are acquired. Preprocessing is performed using outlier removal, temporal alignment, and spatial registration. Then, according to monitoring points, monitoring time, and standardized spatial grids, the preprocessed new monitoring data are correlated with the corresponding actual water quality levels and key indicator values ​​to form a new training sample batch. This new training sample batch is added to the original training set to obtain the updated training set, represented as:

[0222] ;

[0223] in, For the first The updated training set is formed in each update cycle. For the first The training set used in each update cycle For the first Each update cycle generates a batch of new training samples;

[0224] S52, the newly added training samples are sequentially input into the processing steps for extracting water quality temporal evolution features, upstream and downstream spatial neighborhood features, and water body remote sensing features to generate a fused spatiotemporal feature representation, as well as the processing steps for evidence synthesis, spatial propagation and aggregation, and fusion weight adjustment. This yields the deep fusion results of multi-source data corresponding to the newly added training samples. The grade determination loss is calculated based on the actual water quality grade results and the water quality grade determination results of the newly added training samples, and the indicator prediction loss is calculated based on the actual values ​​and predicted values ​​of key indicators. The grade determination loss and indicator prediction loss are weighted to obtain the incremental update loss for the current update cycle, expressed as:

[0225] ;

[0226] ;

[0227] ;

[0228] in, For the first The loss in level determination per update cycle, For the first The indicator predicts the loss over each update cycle. For the first Incremental update loss per update cycle, This refers to the number of samples in the newly added training sample batch. The number of water quality grades For the first Each newly added training sample is assigned a value indicating whether it belongs to a specific water quality level; if it does, the value is 1, otherwise 0. For the first The predicted probability that each newly added training sample belongs to a water quality level. For key indicator quantity, This refers to the predicted number of time slots within a future predetermined time period. For the first The first newly added training sample Key indicators in the future The predicted value at each prediction time. These are the actual values ​​of the corresponding key indicators. , These are the weighting coefficients for the grade determination loss and the indicator prediction loss, respectively.

[0229] S53, based on incremental update loss, calculate the loss gradient of the first parameter set used to extract water quality temporal evolution features, upstream and downstream spatial neighborhood features, and water body remote sensing features to generate a fused spatiotemporal feature representation, and the loss gradient of the second parameter set used for evidence synthesis, spatial propagation and aggregation, and fusion weight adjustment. Update the first and second parameter sets according to the corresponding incremental learning rate, and use the updated parameters for multimodal spatiotemporal feature extraction and multi-source data deep fusion in the next monitoring cycle, as shown below:

[0230] ;

[0231] ;

[0232] in, For the first The first parameter set after each update cycle is completed includes relevant parameters used to extract water quality temporal evolution features, upstream and downstream spatial neighborhood features, and water body remote sensing features to generate a fused spatiotemporal feature representation. The first set of parameters before the update. For the first The second parameter set, updated after each update cycle, includes parameters related to evidence synthesis, spatial propagation and aggregation, and fusion weight adjustment. The second set of parameters before the update. The incremental learning rate is the value of the first set of parameters. The incremental learning rate is the second set of parameters. To incrementally update the gradient of the loss relative to the first set of parameters, This is used to incrementally update the gradient of the loss relative to the second set of parameters.

[0233] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0234] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A smart monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion, characterized in that, Includes the following steps: S1. Collect monitoring data of the target water area, including water quality data, hydrological data, meteorological data, satellite remote sensing data and UAV aerial survey data. Use the local outlier factor algorithm to remove abnormal data, use the dynamic time warping algorithm to align the data of different sampling frequencies in time, and use interpolation and coordinate transformation to perform spatial registration to obtain preprocessed multi-source sensing data. S2, based on the preprocessed multi-source sensing data, extract the water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features of each monitoring point, and use a cross-modal attention mechanism to map the water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features to a unified embedding space to generate a fused spatiotemporal feature representation. S3. Based on the fusion spatiotemporal feature representation, the DS evidence theory is used to synthesize evidence for water quality data, hydrological data and meteorological data of the same monitoring point. A graph convolutional network is used to spatially propagate and aggregate the fusion results of adjacent monitoring points. The quality assessment score is calculated according to the data missing rate, signal-to-noise ratio and timeliness of each data source. The fusion weight of each data source is dynamically adjusted to obtain the deep fusion result of multi-source data. S4. Based on the deep fusion results of the multi-source data, the water quality level is determined, and the water quality change trend and key indicator values ​​in the future time period are predicted according to the deep fusion results of historical multi-source data. When the monitoring indicators exceed the preset threshold or the prediction results show a trend of water quality deterioration, an early warning is triggered and a dynamic monitoring report is generated. S5 adds newly collected monitoring data to the training set and incrementally updates the relevant parameters used to extract water quality temporal evolution characteristics, upstream and downstream spatial neighborhood characteristics, and water body remote sensing characteristics to generate fused spatiotemporal feature representations, as well as the relevant parameters used for evidence synthesis, spatial propagation and aggregation, and fusion weight adjustment.

2. The intelligent monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 1, characterized in that, S1 includes: S11, water quality sensors, hydrological sensors, and meteorological sensors are deployed in the target water area. The water quality sensors are used to collect data on pH value, dissolved oxygen, turbidity, conductivity, ammonia nitrogen, and total phosphorus. The hydrological sensors are used to collect data on water level, flow velocity, and flow rate. The meteorological sensors are used to collect data on air temperature, air pressure, wind speed, and light intensity. At the same time, satellite remote sensing data covering the target water area and UAV aerial survey data are accessed, and the collection time, collection location, and data source identifiers are added to each monitoring data to generate the original multi-source sensing dataset of the target water area. S12, Based on the original multi-source sensing dataset, a local neighborhood is constructed for the continuous monitoring data collected by each sensor, and the local outlier factor algorithm is used to calculate the local outlier factor value of each monitoring data. Monitoring data with local outlier factor values ​​greater than the anomaly judgment threshold are marked as abnormal data and removed to obtain multi-source sensing data after anomaly removal. S13, based on the multi-source sensing data after anomaly removal, taking a unified target time series as a benchmark, a dynamic time warping algorithm is used to calculate the cumulative matching distance between data sequences with different sampling frequencies and the target time series, and the time alignment path is determined according to the minimum cumulative matching distance, and each data sequence is mapped to a unified time node to obtain time-aligned multi-source sensing data; S14. Based on the time-aligned multi-source sensing data, a standardized spatial grid covering the target water area is established. According to the spatial distance between each monitoring point and the standardized spatial grid node, the water quality data, hydrological data and meteorological data are interpolated and calculated to map the data of each monitoring point to the corresponding standardized spatial grid node. S15, the spatial coordinates of the satellite remote sensing data and UAV aerial survey data are transformed to a target coordinate system consistent with the standardized spatial grid. The transformed satellite remote sensing data and UAV aerial survey data are then matched with the interpolated water quality data, hydrological data and meteorological data according to their spatial location to obtain preprocessed multi-source sensing data.

3. The intelligent monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 2, characterized in that, S2 includes: S21, Based on the preprocessed multi-source sensing data, water quality time series are constructed according to monitoring points and collection time. The pH value, dissolved oxygen, turbidity, conductivity, ammonia nitrogen and total phosphorus data of each monitoring point at continuous time are input into a bidirectional long short-term memory network. The water quality change process is encoded along the forward and reverse time directions respectively. The forward hidden state and the reverse hidden state are spliced ​​together to obtain the water quality time series evolution characteristics of each monitoring point. The water quality time series evolution characteristics include trend characteristics, periodic characteristics and abrupt change characteristics. S22, construct a spatiotemporal relationship graph of the target water area based on the preprocessed multi-source sensing data, take each monitoring point as a node, and establish spatial edges according to the geographical proximity and water flow direction correlation between monitoring points. Take the water quality temporal evolution characteristics of each monitoring point as node features and input them into the graph attention network. Calculate the attention coefficient between the target monitoring point and its neighboring monitoring points, and perform weighted aggregation of the node features of neighboring monitoring points according to the attention coefficient to obtain the upstream and downstream spatial neighborhood features that characterize the propagation relationship of upstream and downstream water quality changes. S23, Based on the preprocessed satellite remote sensing data and UAV aerial survey data, band operations are performed on the multispectral remote sensing image covering the target water area to extract the normalized water index, chlorophyll a concentration inversion value and suspended matter concentration inversion value. The normalized water index, chlorophyll a concentration inversion value and suspended matter concentration inversion value are combined to form a remote sensing feature image. Convolutional encoding is performed through a convolutional neural network to obtain the water remote sensing features of the target water area. S24, the water quality temporal evolution features, upstream and downstream spatial neighborhood features, and water body remote sensing features are linearly projected respectively, and the correlation weights between different modal features are calculated using a cross-modal attention mechanism. Based on the correlation weights, the projected water quality temporal evolution features, upstream and downstream spatial neighborhood features, and water body remote sensing features are weighted and fused to generate a fused spatiotemporal feature representation.

4. The intelligent monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 3, characterized in that, S3 includes: S31, based on the fused spatiotemporal feature representation, water quality data, hydrological data and meteorological data corresponding to the same monitoring point are obtained respectively. The water quality data, hydrological data and meteorological data are used as independent evidence to determine the water quality level. The conflict coefficient and trust function between each independent evidence are calculated using the DS evidence theory. The independent evidence is synthesized according to the conflict coefficient and trust function to obtain the comprehensive judgment result of each monitoring point. S32, based on the constructed spatiotemporal relationship graph, uses the comprehensive discrimination results of each monitoring point as node features, and constructs a graph convolutional network according to the geographical adjacency relationship and water flow direction relationship between monitoring points. The graph convolutional network performs spatial propagation and feature aggregation on the comprehensive discrimination results of each monitoring point, so that water quality information between adjacent monitoring points can be transferred to each other, and the spatially smoothed water quality distribution state can be obtained. S33. Based on the spatially smoothed water quality distribution, the missing data rate, signal-to-noise ratio, and timeliness of water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data are statistically analyzed. The quality assessment score of each data source is calculated based on the missing data rate, signal-to-noise ratio, and timeliness. A gating mechanism is used to dynamically adjust the weight of each data source in the fusion process according to the quality assessment score of each data source. When the quality assessment score of any data source decreases, the fusion weight of the corresponding data source is automatically reduced to obtain the deep fusion result of multi-source data.

5. The intelligent monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 4, characterized in that, S31 includes: S311, Based on the fused spatiotemporal feature representation, water quality data, hydrological data and meteorological data corresponding to the same monitoring point are obtained respectively, and according to the water quality level judgment criteria corresponding to each type of data, the basic probability assignment of water quality data, hydrological data and meteorological data to each water quality level is calculated, and the basic probability assignment of each type of data to the water quality level is used as water quality evidence, hydrological evidence and meteorological evidence respectively, forming an independent evidence set for the same monitoring point; S312, Based on the set of independent evidence, the DS evidence theory is used to calculate the conflict coefficient between each independent piece of evidence and the trust function corresponding to each water quality level, and to determine the degree of consistency between different independent pieces of evidence. S313, Based on the conflict coefficient and the trust function, perform DS evidence synthesis on the water quality evidence, hydrological evidence and meteorological evidence, calculate the comprehensive basic probability assignment for each water quality level, and take the judgment result corresponding to the water quality level with the largest comprehensive basic probability assignment as the comprehensive judgment result for the monitoring point.

6. The intelligent monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 5, characterized in that, S32 includes: S321, based on the constructed spatiotemporal relationship graph, each monitoring point is used as a graph node, and the geographical adjacency relationship and water flow direction relationship between monitoring points are used as graph edges. The comprehensive judgment result of each monitoring point is converted into a node feature vector, and a node feature matrix and an adjacency matrix are constructed. The node feature vector is composed of the comprehensive basic probability assignment corresponding to each water quality level, and the adjacency matrix is ​​used to characterize whether there is a geographical adjacency relationship or a water flow direction relationship between any two monitoring points. S322, add self-connection relationships for each monitoring point based on the adjacency matrix, and normalize the adjacency matrix after adding self-connection to obtain a normalized adjacency matrix. Input the node feature matrix and the normalized adjacency matrix into the graph convolutional network so that each monitoring point receives comprehensive discrimination information from its adjacent upstream monitoring points, downstream monitoring points and geographically adjacent monitoring points, thus completing the spatial propagation of water quality information between monitoring points. S323, using a graph convolutional network, the node features of each monitoring point and its neighboring monitoring points are weighted and aggregated according to the normalized adjacency matrix. The aggregation results are then subjected to feature transformation and nonlinear mapping to obtain the spatial fusion features of each monitoring point. The spatial fusion features of each monitoring point are then mapped to the standardized grid of the target water area according to their spatial location to form a spatially smoothed water quality distribution state.

7. The intelligent monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 6, characterized in that, S33 includes: S331, based on the spatially smoothed water quality distribution state, extract the data records of water quality data, hydrological data, meteorological data, satellite remote sensing data and UAV aerial survey data within the current evaluation period, and count the amount of data to be collected, the amount of effective data, the signal power, the noise power and the data update time of each data source, and calculate the data missing rate, signal-to-noise ratio and timeliness of each data source to form a set of quality indicators for each data source. S332, based on the data missing rate, signal-to-noise ratio and timeliness of each data source, the signal-to-noise ratio is normalized, and the data integrity, normalized signal-to-noise ratio and timeliness are weighted and calculated to obtain the quality assessment score corresponding to each data source. Among them, the higher the data missing rate, the lower the data integrity, the higher the signal-to-noise ratio and timeliness, the higher the quality assessment score corresponding to the data source. S333, the quality assessment scores corresponding to each data source are input into the gating mechanism to calculate the gating value of each data source in the current fusion process, and the gating value is normalized to obtain the fusion weight of each data source. Based on the fusion weight, the feature representations corresponding to water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data are weighted and fused, and combined with the spatially smoothed water quality distribution state to obtain the deep fusion result of multi-source data. When the quality assessment score of any data source decreases, its gating value and fusion weight decrease accordingly.

8. The intelligent monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 7, characterized in that, S4 includes: S41, Based on the deep fusion results of the multi-source data, the deep fusion results of the multi-source data corresponding to each monitoring point are input into the trained multi-class support vector machine, and the classification score of the water quality corresponding to the deep fusion results of the multi-source data is calculated respectively. The water quality category with the largest classification score is determined as the water quality level of the corresponding monitoring point, and the water quality level determination result of the target water area is formed. S42, extract the deep fusion results of historical multi-source data within multiple consecutive monitoring periods according to the monitoring time sequence, construct a sequence of deep fusion results of historical multi-source data, input the sequence of deep fusion results of historical multi-source data into a time-series prediction model based on the Transformer architecture, calculate the degree of temporal correlation between historical monitoring periods through self-attention, and perform feature aggregation based on the degree of temporal correlation of each historical monitoring period to predict the water quality change trend and key indicator prediction values ​​of each monitoring point within a predetermined time period in the future. Key indicators include pH value, dissolved oxygen, turbidity, conductivity, ammonia nitrogen and total phosphorus. S43, compare the current key indicator monitoring values ​​and the predicted key indicator values ​​within a predetermined future time period of each monitoring point with the corresponding preset indicator thresholds, and determine whether there is a water quality deterioration trend based on the predicted changes of the key indicators. When any current key indicator monitoring value exceeds the corresponding preset indicator threshold, or any key indicator predicted value exceeds the corresponding preset indicator threshold within a predetermined future time period, or the prediction result shows a water quality deterioration trend, trigger an early warning signal, and generate a dynamic monitoring report including a spatiotemporal evolution diagram and trend analysis chart based on the water quality level determination result, the predicted key indicator values, and the early warning result.

9. A smart monitoring method for river and lake aquatic ecological environment based on multi-source sensing data fusion as described in claim 8, characterized in that, S43 includes: S431, acquire the monitoring values ​​of pH, dissolved oxygen, turbidity, conductivity, ammonia nitrogen and total phosphorus at each monitoring point at the current monitoring time, as well as the predicted values ​​of the corresponding key indicators in the future predetermined time period, compare the current monitoring values ​​of key indicators and the predicted values ​​of key indicators at each prediction time with the corresponding preset indicator thresholds, determine whether each key indicator exceeds the limit at present and whether it will exceed the predicted limit in the future predetermined time period, and obtain the current indicator exceedance judgment result and the predicted indicator exceedance judgment result; S432, calculate the predicted changes of each key indicator relative to the current monitoring value at each monitoring point in the future predetermined time period, and compare the predicted increase of turbidity, conductivity, ammonia nitrogen and total phosphorus, the predicted decrease of dissolved oxygen and the deviation of the predicted pH value from the normal range with the corresponding deterioration judgment thresholds according to the water quality impact direction corresponding to each key indicator. When any key indicator meets the corresponding deterioration judgment condition, the key indicator is marked as a deterioration indicator. When at least two of the multiple key indicators are marked as deterioration indicators, it is determined that there is a water quality deterioration trend at the corresponding monitoring point. S433, combining the current indicator exceeding the limit judgment result, the predicted indicator exceeding the limit judgment result, and the water quality deterioration trend judgment result, when the current monitoring value of any key indicator exceeds the corresponding preset indicator threshold, or the predicted value of any key indicator exceeds the corresponding preset indicator threshold within a predetermined time period in the future, or when there is a water quality deterioration trend at the corresponding monitoring point, an early warning signal is triggered. The water quality level judgment result, the predicted value of the key indicator, and the early warning signal are correlated according to the monitoring time and the spatial location of the monitoring point to generate a spatiotemporal evolution map and trend analysis chart including the spatial distribution of water quality level, the spatiotemporal changes of key indicators, the predicted trend of key indicators, and the early warning location, forming a dynamic monitoring report.

10. A method for intelligent monitoring of river and lake aquatic ecological environment based on multi-source sensing data fusion according to claim 9, characterized in that, S5 includes: S51. Acquire newly collected water quality data, hydrological data, meteorological data, satellite remote sensing data, and UAV aerial survey data according to the predetermined update cycle. Perform preprocessing using abnormal data removal, time-series alignment, and spatial registration. Then, according to the monitoring point, monitoring time, and standardized spatial grid, associate the preprocessed new monitoring data with the corresponding actual results of water quality level and actual values ​​of key indicators to form a new training sample batch. Add the new training sample batch to the original training set to obtain the updated training set. S52, the newly added training samples are sequentially input into the processing process for extracting water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features and generating fused spatiotemporal feature representations, as well as the processing process for evidence synthesis, spatial propagation and aggregation and fusion weight adjustment, to obtain the multi-source data deep fusion results corresponding to the newly added training samples. The level determination loss is calculated based on the actual water quality level results and water quality level determination results of the newly added training samples, and the indicator prediction loss is calculated based on the actual values ​​and predicted values ​​of key indicators. The level determination loss and indicator prediction loss are weighted to obtain the incremental update loss of the current update cycle. S53, based on the incremental update loss, calculate the loss gradient of the first parameter set used to extract water quality temporal evolution features, upstream and downstream spatial neighborhood features and water body remote sensing features and generate fused spatiotemporal feature representations, and the loss gradient of the second parameter set used for evidence synthesis, spatial propagation and aggregation and fusion weight adjustment. Update the first parameter set and the second parameter set according to the corresponding incremental learning rate, and use the updated parameters for multimodal spatiotemporal feature extraction and multi-source data deep fusion in the next monitoring cycle.

Images