Real-time fusion and dynamic monitoring method for multi-source heterogeneous water quality data based on domestic AI remote sensing cloud platform

Abstract

The present application relates to the technical field of water environment treatment, in particular to a multi-source heterogeneous water quality data real-time fusion and dynamic monitoring method based on a domestic AI remote sensing cloud platform, comprising the following steps: integrating multi-source heterogeneous water quality data collected by domestic high-resolution satellite remote sensing, unmanned aerial vehicle remote sensing and ground sensors, after standardization preprocessing, extracting spatial and time sequence characteristics through convolutional neural network and long short-term memory network respectively, introducing an attention mechanism to realize real-time deep fusion of multi-source data; based on the fused data, a dynamic monitoring model combining a hybrid kernel function support vector machine and a hidden Markov model is constructed to realize accurate water quality inversion and trend prediction. The whole process of this method adapts to the domestic software and hardware ecology, significantly improves the accuracy, real-time performance and comprehensiveness of water quality monitoring, and provides independent and controllable technical support for water resource protection and water environment treatment.

Description

Technical Field

[0001] This invention relates to the field of water environment management technology, specifically a method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform. Background Technology

[0002] With rapid economic and social development, water pollution has become increasingly prominent, placing higher demands on the accuracy, real-time nature, and comprehensiveness of monitoring technologies for water quality management in river basins, lakes, and reservoirs. Traditional water quality monitoring mainly relies on manual sampling and laboratory analysis, which has inherent drawbacks such as limited monitoring range, significant time lag, high monitoring costs, and difficulty in achieving large-scale dynamic continuous monitoring. These methods can no longer meet the practical application needs of modern refined water environment management and emergency response to sudden water pollution incidents. In recent years, the rapid integration of artificial intelligence, remote sensing, and in-situ sensor technologies has provided core technological support for the intelligent upgrading of water quality monitoring. Intelligent water quality monitoring methods based on multi-source data fusion have become a key research direction in this field.

[0003] In the prior art, CN120493079A discloses a water quality monitoring method and system based on artificial intelligence. The core technical content of this solution is as follows: The method first acquires a comprehensive dataset composed of sensor data, satellite images, and meteorological parameters. It then performs standardization, temporal and spatial alignment, and data missing and anomaly handling on the multi-source data sequentially, and completes data fusion through weighted averaging. Secondly, based on the water flow velocity and pollution concentration gradient in the comprehensive dataset, a dynamic sampling algorithm is used to adjust the sampling frequency and location. The adjusted data undergoes spatiotemporal interpolation to obtain a preprocessed dataset with high-density spatiotemporal coverage. Subsequently, principal component analysis is used to extract key features of sensor values, image textures, and meteorological parameters from the preprocessed dataset, and entropy is used to... A weighted feature matrix is ​​constructed by assigning weights to obtain a fused feature set. Then, it is determined whether the dimensionality of the fused feature set exceeds a preset threshold. If it does, a random forest algorithm is used to reduce the dimensionality and classify the fused features. Model parameters are optimized through cross-validation to obtain the pollution concentration prediction result. Next, a convolutional neural network is used to analyze the spatial pattern of the pollution concentration prediction result, and the pollution migration path is calculated by combining it with water flow field data. The location of the pollution source is determined through iterative optimization, resulting in the pollution source tracing result. Finally, it is determined whether the confidence level of the pollution source tracing result is higher than a preset threshold. If it is higher, a Bayesian network is used to construct a real-time early warning model, fusing the pollution concentration prediction result and the pollution source tracing result to generate a dynamic early warning signal. This early warning signal is then input into a push algorithm to transmit the early warning data to the decision support system. This scheme realizes intelligent processing of the water quality monitoring process, improves the accuracy of pollution concentration prediction and pollution source tracing to a certain extent, and provides a reference for the development of intelligent water quality monitoring technology.

[0004] However, the technical solutions disclosed in the aforementioned prior art documents still have obvious technical defects in practical applications, specifically in the following aspects: First, at the level of multi-source heterogeneous data fusion, this scheme only adopts a linear fusion method that combines principal component analysis with entropy weighting. It does not design a targeted feature extraction and nonlinear deep fusion scheme for the spatial distribution characteristics and time series characteristics of the three types of heterogeneous data: satellite remote sensing, UAV remote sensing, and ground sensor data. As a result, it cannot give full play to the complementary advantages of multi-source heterogeneous data, and the depth and accuracy of data fusion are insufficient. In complex water environments, information silos are likely to occur, which limits the accuracy of water quality parameter inversion.

[0005] Second, in terms of adapting to domestic technologies, the application of remote sensing data in this solution only involves general satellite images. It has not been adapted and optimized for the technical characteristics of domestic high-resolution series satellites and domestic UAV remote sensing payloads, nor has it constructed water quality sensitive feature factors adapted to domestic remote sensing data. At the same time, it lacks full-process adaptation to domestic AI remote sensing cloud platforms, domestic AI acceleration chips, domestic operating systems and deep learning frameworks, and cannot achieve domestic independent control of the entire water quality monitoring process. There are obvious shortcomings in the level of independent development of key technologies.

[0006] Third, in terms of dynamic monitoring and real-time performance, the monitoring model of this scheme only realizes static prediction of pollution concentration and post-event source tracing. It has not built a time-series prediction model that can capture the dynamic changes in water quality in real time, and cannot make advance predictions of water quality change trends. At the same time, the scheme does not adopt a lightweight computing architecture with cloud-edge collaboration, and the end-to-end latency of data processing and model inference is high, which makes it difficult to meet the actual needs of minute-level emergency monitoring of sudden water pollution events.

[0007] Fourth, at the data preprocessing level, this scheme does not design a systematic preprocessing scheme for the spatiotemporal heterogeneity of multi-source heterogeneous data. It only uses basic linear interpolation and K-nearest neighbor filling methods to handle missing data, which cannot effectively eliminate the matching bias and noise interference of multi-source data in the spatiotemporal dimension, resulting in insufficient stability and robustness of subsequent feature extraction and model prediction.

[0008] Based on the technical bottlenecks of the existing technologies, there is an urgent need to develop a water quality monitoring method that can achieve deep nonlinear fusion of multi-source heterogeneous water quality data, is compatible with domestic AI remote sensing technology systems, and has high real-time dynamic monitoring and trend prediction capabilities, so as to break through the limitations of existing technologies and meet the core needs of refined and intelligent management of the current water environment. Summary of the Invention

[0009] The purpose of this invention is to provide a method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform, in order to solve the problems mentioned in the background art, such as insufficient multi-source heterogeneous data fusion capability, limited real-time dynamic monitoring performance, and weak domestic adaptability of existing technologies.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform includes the following steps: S1 Multi-Source Heterogeneous Water Quality Data Acquisition: Satellite remote sensing multispectral / hyperspectral data of the target water area is acquired through domestic Gaofen series satellites, UAV remote sensing data of key areas of the target water area is acquired through UAVs equipped with multispectral imaging payloads, and in-situ water quality time series data is acquired through ground sensor arrays deployed in the target water area. All acquired data is transmitted to the domestic AI remote sensing cloud platform in real time. S2 Multi-Source Heterogeneous Data Standardization Preprocessing: Data cleaning, spatiotemporal registration, and format normalization are performed sequentially on the collected multi-source data. Wavelet transform denoising and Kalman filter temporal smoothing are then applied to obtain a standardized multi-source dataset with matching spatiotemporal dimensions. S3 Real-time fusion of multi-source data based on domestic AI framework: In the domestic AI remote sensing cloud platform, spatial features of remote sensing images are extracted by convolutional neural network (CNN), temporal features of ground time series data are extracted by long short-term memory network (LSTM), and attention mechanism is introduced to dynamically allocate the fusion weights of multi-source features to complete the real-time fusion of multi-source heterogeneous water quality data and obtain a fused feature dataset. S4 Water Quality Dynamic Monitoring Model Construction and Execution: Based on the fused feature dataset, water quality parameters are regressed and inverted using Support Vector Machine (SVM), and a water quality time series change prediction model is constructed using Hidden Markov Model (HMM). Real-time dynamic monitoring and trend prediction of water quality in the target water area are realized in the domestic AI remote sensing cloud platform.

[0011] Preferably, in step S2, the spatiotemporal registration process uses a unified spatiotemporal reference model to map multi-source data to the same spatiotemporal coordinate system. The calculation formula for the unified spatiotemporal reference model is as follows: ; in, To standardize water quality data in a spatiotemporal coordinate system, This is the original data. The time-dimensional linear registration transformation matrix is... The affine registration transformation matrix is ​​the spatial dimension. The wavelet transform denoising uses a hard threshold function, the formula of which is: ; in, These are the denoised wavelet coefficients. These are the original wavelet coefficients. An adaptive threshold calculated based on noise variance. To decompose the scale, This is the translation coefficient.

[0012] Preferably, in step S1, the satellite remote sensing data are multispectral and hyperspectral image data from the domestic Gaofen-6 and Gaofen-7 satellites. After radiometric calibration, atmospheric correction, and water body masking, the spectral reflectance of the sensitive bands of the water body is extracted to construct the spectral feature factors for water quality parameter inversion. The calculation formula for the spectral feature factors is as follows: ; ; in, It is a chlorophyll a concentration-sensitive characteristic factor. As a sensitive characteristic factor of suspended solids concentration, wavelength Remote sensing reflectance of water bodies at a given location.

[0013] Preferably, step S1, the acquisition and processing of UAV remote sensing data, specifically includes the following sub-steps: S121 Track Planning: Based on the vector boundary of the target water area and the key areas of interest identified by satellite remote sensing, an adaptive track is generated, with a heading overlap of 80% and a lateral overlap of 70%. At the same time, denser sections are set for shoreline inflection points, sewage outlets into rivers, and algae accumulation areas. S122 Data Acquisition: Controls a fixed-wing UAV carrying a multispectral imaging payload to complete flight operations according to a planned trajectory, and simultaneously acquires multispectral image data, POS positioning and attitude data, and flight attitude data of the target water area; S123 Image Preprocessing: Radiometric calibration, orthorectification, image stitching, band registration, and water masking are performed sequentially on the acquired raw images to remove land and non-water areas, resulting in a sub-meter spatial resolution spectral feature dataset of key areas of the target water body, supplementing the local detail features of the satellite remote sensing data.

[0014] Preferably, step S1, the deployment and data acquisition of the ground sensor array, specifically includes the following sub-steps: S131 Monitoring Point Deployment: Monitoring points are deployed at the inlet, outlet, key pollution control area, shoreline sensitive area and background control area of ​​the target water area. Each point is equipped with an integrated multi-parameter water quality monitoring buoy to form a ground sensor array covering the entire water area. S132 Sensor Calibration: Before data acquisition, all sensors are calibrated in the laboratory and on-site in-situ. The calibration parameters are uploaded to the domestic AI remote sensing cloud platform and a sensor calibration log is established. S133 Time Series Data Acquisition: In-situ time series data of pH, dissolved oxygen, conductivity, turbidity, chemical oxygen demand, total phosphorus, and total nitrogen are synchronously acquired through a sensor array, with an acquisition frequency of no less than once per hour. S134 Real-time Data Transmission: The collected in-situ data is uploaded to the domestic AI remote sensing cloud platform in real time via a 5G / NB-IoT wireless network and an encrypted transmission channel, while data caching and backup are completed at the local edge node.

[0015] Preferably, in step S3, the dynamic weight allocation calculation formula for the attention mechanism is as follows: ; ; in, For the first The fusion weights of features from similar data sources For the first Salience score of data source features. For the trainable weight matrix of the attention mechanism, This represents the total number of data source categories. For the first Feature vectors extracted from class data sources, This is the final output fused feature dataset.

[0016] As a preferred option, the real-time fusion of multi-source data in step S3 adopts a cloud-edge collaborative computing architecture, specifically executing the following sub-steps: S31 edge node preprocessing: On domestically produced edge computing nodes deployed at the monitoring site, preliminary cleaning, outlier removal, local feature extraction and data compression of multi-source heterogeneous data are completed to reduce data transmission bandwidth usage; S32 encrypted data transmission: The edge nodes encrypt the preprocessed feature data using the national cryptographic algorithm and then upload it to the cloud center node of the domestic AI remote sensing cloud platform; S33 Cloud Global Fusion Processing: The cloud central node receives feature data uploaded by multiple nodes, completes global registration in the spatiotemporal dimension, performs global fusion of multi-source features based on CNN-LSTM-attention mechanism, and generates a fused feature dataset; S34 Result Synchronous Backhaul: The cloud synchronously backhauls the fusion processing results and model inference instructions to the corresponding edge nodes, ensuring that the end-to-end data processing delay does not exceed 15 minutes.

[0017] Preferably, in step S4, the Support Vector Machine (SVM) uses a hybrid kernel function to perform water quality parameter regression and inversion. The calculation formula for the hybrid kernel function is as follows: ; in, For radial basis kernel functions, For polynomial kernel functions, These are the kernel function weighting coefficients, with values ​​ranging from 1 to 2. , The input is the fused feature vector. These are support vectors.

[0018] Preferably, in step S4, the state transition probability matrix of the Hidden Markov Model (HMM) is solved iteratively using the Baum-Welch algorithm, with the objective function being: ; in, These are the triplet parameters of the HMM model. Let be the initial state probability vector. Here is the state transition probability matrix. For the observation probability matrix, The model is a sequence of water quality parameter observations. It divides the water quality status into five discrete states: excellent, good, medium, poor, and bad, and outputs the probability of water quality status changes and the predicted values ​​of water quality parameters for the next 24 to 72 hours.

[0019] Preferably, in step S2, the outlier detection and correction in the data cleaning process includes the following sub-steps: S21 Single Data Source Gross Error Removal: Based on the 3σ criterion, the mean and standard deviation of time series data from a single data source are calculated, and invalid gross error data exceeding the range of mean ± 3 times the standard deviation are removed; S22 Multi-source data cross-validation: compares multi-source data from the same spatiotemporal location and marks abnormal data whose deviation from the measured values ​​of ground sensors exceeds a preset threshold; S23 Outlier Correction: For marked outlier data, linear interpolation of valid data from adjacent time periods at the same location is used to correct the outlier data. The corrected data is then incorporated into a standardized multi-source dataset.

[0020] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention addresses the technical pain points of insufficient fusion depth and inadequate utilization of feature complementarity in existing water quality monitoring technologies by constructing a deep fusion technology system for multi-source data, which integrates spatial feature extraction, temporal feature mining, and dynamic weight fusion. Compared to the existing fusion method that combines principal component analysis with linear weighting, this invention uses convolutional neural networks to specifically extract the spatial distribution features of water bodies from satellite and UAV remote sensing data, and uses long short-term memory networks to mine the dynamic change patterns of water quality in ground sensor temporal data. At the same time, it introduces an attention mechanism to achieve dynamic adaptive allocation of multi-source feature weights, effectively breaking through the technical bottleneck of poor adaptability of traditional linear fusion methods to heterogeneous data features, and significantly improving the accuracy and reliability of multi-source data fusion. Meanwhile, this invention effectively eliminates the spatiotemporal heterogeneity and noise interference of multi-source data through a standardized preprocessing process of wavelet transform denoising, Kalman filter temporal smoothing, and spatiotemporal benchmark unified model construction, providing a high-quality data foundation for subsequent model construction. This fundamentally solves the problems of low model inversion accuracy and poor robustness caused by insufficient data preprocessing in existing technologies.

[0021] (2) This invention achieves deep integration of the entire water quality monitoring technology system with the domestic software and hardware ecosystem, filling the application gap of the domestic AI remote sensing cloud platform in the field of multi-source heterogeneous water quality data fusion and dynamic monitoring. This invention constructs exclusive water quality sensitive feature factors based on the spectral characteristics of the domestic Gaofen series satellites, and designs a refined flight and data processing flow to adapt to the technical characteristics of domestic UAV remote sensing payloads. At the same time, the entire process algorithm and model have been adapted and optimized to the domestic Ascend AI acceleration chip, Kylin operating system, and PaddlePaddle deep learning framework, realizing the domestic independent control of the entire chain from data acquisition, data processing, feature fusion to model inference, forming a water quality monitoring technology solution with complete independent intellectual property rights, which greatly improves the safety and compliance of the technology solution in key fields such as water conservancy and environmental protection in China.

[0022] (3) This invention constructs a monitoring and prediction model system adapted to the dynamic changes in water quality. Simultaneously, through cloud-edge collaborative computing architecture optimization, it significantly improves the real-time performance and scenario adaptability of water quality monitoring, solving the technical defects of insufficient real-time performance and weak emergency response capabilities in existing static monitoring models. This invention achieves high-precision inversion of water quality parameters through a hybrid kernel function support vector machine and advance prediction of water quality change trends through a hidden Markov model. Compared to existing technologies that can only achieve post-pollution source tracing, this invention can achieve early warning of water quality anomalies and dynamic tracking of pollution development trends. Furthermore, the cloud-edge collaborative computing architecture adopted in this invention decentralizes data preprocessing and local feature extraction to edge nodes, effectively reducing end-to-end data processing latency. It is adaptable to various application scenarios such as routine watershed monitoring, lake eutrophication early warning, and emergency monitoring of sudden water pollution events. The technical solution possesses good feasibility and engineering implementation capabilities, providing stable and reliable technical support for refined water environment management and pollution emergency response. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0024] Figure 1 This is a flowchart illustrating the overall process of the method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestic AI remote sensing cloud platform, as described in this invention. Figure 2 This is a block diagram of the multi-source heterogeneous water quality data acquisition method of the present invention; Figure 3 This is a flowchart of the multi-source heterogeneous data standardization preprocessing method of the present invention; Figure 4 This is a diagram illustrating the real-time fusion and dynamic monitoring of multi-source data in this invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] like Figures 1-4 As shown, this invention presents a method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform. The entire process is deployed on this platform, adapting to domestic hardware architecture, operating system, and deep learning framework, achieving domestically developed and independently controllable water quality monitoring throughout the entire process. The specific implementation process consists of four core steps, the details of which are as follows: Multi-source heterogeneous water quality data acquisition: Three types of core data sources are collected simultaneously, and all collected data are connected to the domestic AI remote sensing cloud platform in real time through an encrypted transmission channel to achieve unified data aggregation and management.

[0027] The first category is satellite remote sensing data. Multispectral and hyperspectral imagery data from the domestically produced Gaofen-6 and Gaofen-7 satellites were selected, covering the entire target water area, with a revisit period of no more than two days. The acquired raw satellite images underwent radiometric calibration, atmospheric correction, and water masking processing sequentially, removing non-water areas such as land and vegetation, and extracting the remote sensing reflectance of each sensitive band within the water body area. Based on the extracted spectral data, spectral characteristic factors for water quality parameter inversion were constructed, calculated using the following formula: ; ; In the formula, It is a chlorophyll a concentration-sensitive characteristic factor. As a sensitive characteristic factor of suspended solids concentration, wavelength The water body's remote sensing reflectance at a given location. Using these characteristic factors, preliminary inversion of core water quality parameters at the satellite level is achieved, revealing the spatial distribution characteristics of water quality over a large area of ​​the target water body.

[0028] The second category is UAV remote sensing data. Fixed-wing UAVs equipped with multispectral imaging payloads are used to conduct refined data collection on key areas within the target water area. The collection process consists of three sub-steps. First, based on the vector boundary of the target water area and the key areas of interest initially identified by satellite remote sensing, an adaptive flight path is generated, setting a forward overlap of 80% and a lateral overlap of 70%, with additional flight segments set for shoreline inflection points, sewage outfalls, and algae accumulation areas. Second, the UAV is controlled to complete the flight operation according to the planned path, simultaneously collecting multispectral image data, POS positioning and attitude data, and flight attitude data of the target water area. Finally, the collected raw images undergo radiometric calibration, orthorectification, image stitching, band registration, and water masking processing to remove land and non-water areas, resulting in a sub-meter spatial resolution spectral feature dataset of key areas in the target water area, supplementing the local detailed features of the satellite remote sensing data.

[0029] The third category is time-series data from ground-based sensors. Ground-based sensor arrays are deployed in the target water area, covering the inlet, outlet, key pollution control areas, sensitive shoreline areas, and background control areas. Each location is equipped with an integrated multi-parameter water quality monitoring buoy. Before data acquisition, all sensors undergo laboratory calibration and on-site calibration. Calibration parameters are simultaneously uploaded to the domestic AI remote sensing cloud platform, and a sensor calibration log is established. On-site time-series data for pH, dissolved oxygen, conductivity, turbidity, chemical oxygen demand, total phosphorus, and total nitrogen are simultaneously collected via the sensor array, with a collection frequency of no less than once per hour. The collected on-site data is uploaded in real-time to the domestic AI remote sensing cloud platform via a 5G / NB-IoT wireless network and an encrypted transmission channel, while data caching and backup are performed at local edge nodes.

[0030] Multi-source heterogeneous data standardization preprocessing: Data cleaning, spatiotemporal registration, format normalization, and noise reduction and smoothing are performed sequentially on the aggregated multi-source heterogeneous data to obtain a standardized multi-source dataset with spatiotemporal dimension matching, providing a high-quality data foundation for subsequent data fusion.

[0031] First, data cleaning is performed to detect and correct outliers, which consists of three sub-steps. The first step is to remove outliers from a single data source. Based on the 3σ criterion, the mean and standard deviation of the time series data from a single data source are calculated, and invalid outlier data exceeding the mean ± 3 times the standard deviation are removed. The formula for calculating the 3σ criterion is as follows: ; ; In the formula, The mean of the time series data. The standard deviation of the time series data. This represents the number of samples in the time series data. For the first A time-series data sample. When the data sample satisfies... If the discrepancy is too high, it is considered an invalid gross error and is discarded. The second step is to perform multi-source data cross-validation, comparing multi-source data from the same spatiotemporal location and marking abnormal data whose deviation from the measured values ​​of the ground sensor exceeds a preset threshold. The third step is to perform outlier correction, using linear interpolation of valid data from adjacent time periods at the same location to correct the marked abnormal data. The corrected data is then incorporated into the subsequent processing flow.

[0032] Next, spatiotemporal registration is performed, using a unified spatiotemporal reference model to map multi-source data to the same spatiotemporal coordinate system. The calculation formula is as follows: ; In the formula, To standardize water quality data in a spatiotemporal coordinate system, This is the original data. The time-dimensional linear registration transformation matrix is... This is the affine registration transformation matrix for spatial dimensions. Through this model, data from different data sources, with different acquisition frequencies, and different spatial resolutions are unified to the same time step and spatial grid coordinate system, achieving precise matching of spatiotemporal dimensions.

[0033] Subsequently, format normalization processing is performed to unify the encoding format, data structure and dimensions of multi-source data. Min-max normalization processing is performed on different water quality parameters to map all parameter values ​​to the range of 0 to 1, eliminating the impact of dimensional differences on subsequent fusion and modeling.

[0034] Finally, denoising and time-series smoothing are performed. Wavelet transform denoising is used to denoise the spatial spectral data, and Kalman filtering is used to smooth the time-series data. Wavelet transform denoising uses a hard threshold function, as shown in the following formula: ; In the formula, These are the denoised wavelet coefficients. These are the original wavelet coefficients. An adaptive threshold calculated based on noise variance. To decompose the scale, The translation coefficient is used. Through multi-level wavelet decomposition and thresholding, noise interference in the image data is effectively removed, improving the signal-to-noise ratio. Kalman filtering, through recursive estimation, smooths the time-series data of ground sensors, eliminating random noise interference with the time-series trend and obtaining stable time-series variation characteristics.

[0035] Real-time fusion of multi-source data based on a domestic AI framework: This is executed on a domestic AI remote sensing cloud platform, using a cloud-edge collaborative computing architecture. By combining convolutional neural networks, long short-term memory networks, and attention mechanisms, real-time fusion of multi-source heterogeneous water quality data is achieved, resulting in a fused feature dataset.

[0036] The execution process of the cloud-edge collaborative computing architecture consists of four sub-steps. The first step involves deploying domestically produced edge computing nodes at the monitoring site to perform preliminary cleaning, outlier removal, local feature extraction, and data compression of multi-source heterogeneous data, reducing data transmission bandwidth consumption. The second step involves the edge nodes encrypting the pre-processed feature data using national cryptographic algorithms before uploading it to the cloud-based central node of the domestic AI remote sensing cloud platform. The third step involves the cloud-based central node receiving the feature data uploaded from multiple nodes, performing global registration in the spatiotemporal dimensions, and executing global fusion of multi-source features to generate a fused feature dataset. The fourth step involves the cloud synchronously transmitting the fusion processing results and model inference instructions back to the corresponding edge nodes, ensuring that end-to-end data processing latency does not exceed 15 minutes.

[0037] The specific implementation process of the multi-source data fusion algorithm is as follows. First, a convolutional neural network (CNN) is used to extract features from satellite and UAV remote sensing image data. The CNN adopts a 5-layer convolutional structure, extracting spatial distribution features of water bodies at different scales through convolutional kernels, and outputting a spatial feature vector with uniform dimensions. Then, a long short-term memory (LSTM) network is used to model the time-series data of ground sensors to explore the dynamic changes of water quality parameters. The LSM network, through its input gate, forget gate, and output gate structure design, effectively captures long-term dependency features in the time-series data and outputs a time feature vector with uniform dimensions.

[0038] Based on this, an attention mechanism is introduced to dynamically adjust the feature weights of different types of data, achieving efficient fusion of multi-source data. The dynamic weight allocation calculation formula of the attention mechanism is as follows: ; ; In the formula, For the first The fusion weights of features from similar data sources For the first Salience score of data source features. For the trainable weight matrix of the attention mechanism, This represents the total number of data source categories. For the first Feature vectors extracted from class data sources, This is the final output fused feature dataset. Through an attention mechanism, the algorithm can dynamically adjust the fusion weights of different data sources according to data quality and scenario requirements, fully leveraging the complementary advantages of multi-source data and improving the reliability and accuracy of the fused data.

[0039] Water quality dynamic monitoring model construction and execution: Based on the fused feature dataset, a water quality dynamic monitoring model is constructed. Water quality parameter regression and inversion are completed through support vector machine. A water quality time series change prediction model is constructed through hidden Markov model. Real-time dynamic monitoring and trend prediction of water quality in the target water area are realized in the domestic AI remote sensing cloud platform.

[0040] The water quality parameter regression and inversion adopted a support vector machine model. The model uses a hybrid kernel function to balance local fitting ability and global generalization ability. The calculation formula of the hybrid kernel function is as follows: ; In the formula, For radial basis kernel functions, For polynomial kernel functions, These are the kernel function weighting coefficients, with values ​​ranging from 1 to 2. , The input is the fused feature vector. The model uses ground-based sensor data as labels and a fused feature dataset as input to complete model training and optimization. It achieves high-precision inversion of core water quality parameters such as pH, dissolved oxygen, chemical oxygen demand, total phosphorus, total nitrogen, chlorophyll a, and suspended solids across the entire water area, and outputs spatial distribution raster data of water quality parameters across the entire water area.

[0041] A Hidden Markov Model (HMM) is used to predict water quality changes over time. The model divides water quality into five discrete states: excellent, good, moderate, poor, and very poor. The model uses the inversion results of water quality parameters as the observation sequence to construct the time-series prediction model. The state transition probability matrix of the model is solved iteratively using the Baum-Welch algorithm. The optimization objective function is: ; In the formula, For the triplet parameters of the Hidden Markov Model, Let be the initial state probability vector. Here is the state transition probability matrix. For the observation probability matrix, This is a sequence of water quality parameter observations. After model optimization, it can output the probability of water quality status changes and predicted values ​​of water quality parameters for the next 24 to 72 hours, enabling early prediction and warning of water quality change trends.

[0042] Example 1: Routine dynamic monitoring of water quality in rivers within the basin The application scenario of this embodiment is the routine dynamic water quality monitoring of a section of the main stream of a river basin. The monitored section is 42 kilometers long and has a basin area of ​​186 square kilometers. There are 3 industrial park sewage outlets and 5 agricultural non-point source pollution inflow areas along the river. The water quality control requirement is the Class III surface water standard.

[0043] The specific implementation process of this embodiment is as follows.

[0044] The first step involves multi-source heterogeneous data acquisition. Satellite remote sensing data is selected from multispectral images of the domestic Gaofen-6 satellite, with a spatial resolution of 16 meters, a revisit period of 2 days, covering the entire monitored river section, and acquisition over a continuous period of 12 months, acquiring 2 to 3 valid images per month. After performing radiometric calibration, atmospheric correction, and water masking on the satellite images, the remote sensing reflectance of the water's sensitive bands is extracted, and spectral characteristic factors are calculated. Taking measured data from a specific image as an example, the remote sensing reflectance at a wavelength of 670 nm is 0.023, and the remote sensing reflectance at a wavelength of 700 nm is 0.058. Substituting these values ​​into the formula for the chlorophyll a concentration-sensitive characteristic factor, the following calculation is performed: ; The remote sensing reflectance at a wavelength of 550 nm is 0.065, and the remote sensing reflectance at a wavelength of 850 nm is 0.012. Substituting these values ​​into the formula for the suspended matter concentration sensitive characteristic factor, the following calculations are performed: ; The UAV remote sensing data was collected using a fixed-wing UAV equipped with a 6-channel multispectral camera. This UAV conducted detailed monitoring of sewage outlets in three industrial parks and five agricultural non-point source pollution inflow areas. The flight altitude was 200 meters, with a spatial resolution of 0.2 meters, a forward overlap of 80%, and a lateral overlap of 70%. Two flight operations were conducted monthly to acquire high-resolution spectral data for key areas. A ground-based sensor array deployed eight monitoring buoys along the monitored river section, located at the river inlet, outlet, downstream of each sewage outlet, and downstream of the agricultural inflow areas. These buoys simultaneously collected data on pH, dissolved oxygen, conductivity, turbidity, chemical oxygen demand, total phosphorus, and total nitrogen at a frequency of once per hour. The data was uploaded in real-time to a domestically developed AI remote sensing cloud platform.

[0045] The second step involves standardization preprocessing of the multi-source heterogeneous data. First, data cleaning is performed. Taking dissolved oxygen time-series data from a specific monitoring point as an example, 24 consecutive hours of monitoring data are selected, and the data mean is calculated. mg / L, standard deviation mg / L, according to the 3σ criterion, the effective data range is [missing value]. The water quality parameters were initially set at 6.77 mg / L to 8.87 mg / L. Two outlier data points outside this range were removed, and the missing values ​​were corrected using linear interpolation of data from adjacent time periods. Spatiotemporal registration was then performed, unifying satellite, UAV, and ground sensor data into a standardized spatiotemporal coordinate system with a 1-hour time step and 10-meter spatial resolution using a unified spatiotemporal benchmark model. Format normalization was then performed, mapping all water quality parameters to the 0-1 range. Finally, the remote sensing image data underwent 5-layer wavelet decomposition and hard thresholding for denoising, and the ground sensor time-series data was smoothed using Kalman filtering to obtain a standardized multi-source dataset.

[0046] The third step involves real-time fusion of multi-source data. A cloud-edge collaborative computing architecture is adopted, where data preprocessing and local feature extraction are completed at the edge nodes of the monitoring site, and then encrypted and uploaded to the central node of the domestic AI remote sensing cloud platform. In the cloud, spatial features of the remote sensing data are extracted using a convolutional neural network, and temporal features of the ground time-series data are extracted using a long short-term memory network. An attention mechanism is introduced to dynamically allocate fusion weights. In this embodiment, the fusion weights for satellite remote sensing data, UAV remote sensing data, and ground sensor data are 0.25, 0.35, and 0.40, respectively. The weighted fusion yields a fused feature dataset, with an end-to-end processing latency of 12 minutes, meeting the requirements for real-time monitoring.

[0047] The fourth step is to run the water quality dynamic monitoring model. Based on the fused feature dataset, a hybrid kernel function support vector machine is used to complete the water quality parameter inversion, with the kernel function weight coefficients... The coefficient of performance (COP) was set to 0.6 to balance local fitting and global generalization capabilities. The model's inversion error for core indicators such as chemical oxygen demand (COD) and total phosphorus was controlled within 4.2%, which is superior to traditional single-data source inversion methods. A water quality time-series prediction model was constructed using a Hidden Markov Model (HMM), dividing water quality into five states. The model parameters were optimized using the Baum-Welch algorithm to output the water quality change trend for the next 72 hours. During the operation of this embodiment, three instances of water quality anomalies caused by agricultural non-point source pollution were successfully detected, and early warning signals were issued 18 hours in advance, providing accurate data support for watershed water quality management.

[0048] Example 2: Monitoring and Early Warning of Eutrophication in Lakes and Reservoirs The application scenario of this embodiment is the monitoring and early warning of eutrophication in a medium-sized reservoir. The reservoir has a water area of ​​12.8 square kilometers and a total storage capacity of 120 million cubic meters. It is a centralized drinking water source for the local area. The water quality control requirement is the Class II standard for surface water. The core control target is the eutrophication risk caused by algal blooms.

[0049] The specific implementation process of this embodiment is as follows.

[0050] The first step involved multi-source heterogeneous data acquisition. Satellite remote sensing data was selected from hyperspectral images of the domestic Gaofen-7 satellite, with a spatial resolution of 6 meters, a revisit period of 2 days, covering the entire reservoir area. Acquisition occurred from April to October, the peak algae bloom period, with one valid image acquired every two days. After preprocessing the satellite images, the remote sensing reflectance of the water's sensitive bands was extracted, and the spectral characteristic factors of chlorophyll a and suspended matter concentration were calculated. Taking an image from the peak algae bloom period as an example, the remote sensing reflectance at 670nm was 0.031, and the remote sensing reflectance at 700nm was 0.076. Substituting these values ​​into the formula, the chlorophyll a sensitive characteristic factor was calculated. ; The UAV remote sensing data was collected using a vertical take-off and landing fixed-wing UAV equipped with a hyperspectral analyzer. Intensive monitoring was conducted in areas prone to algae accumulation, such as reservoir bays and inlets. The UAV flew at an altitude of 300 meters, with a spatial resolution of 0.5 meters, a forward overlap of 80%, and a lateral overlap of 70%. Flights were conducted twice a week, increasing to once daily during peak algae bloom periods, to acquire detailed spatial data on algae distribution. A ground-based sensor array deployed six monitoring buoys in the reservoir, located at the center, main inlets, intake, and bay areas. These buoys simultaneously collected data on pH, dissolved oxygen, turbidity, chlorophyll a, total phosphorus, total nitrogen, and cyanobacteria density at a frequency of once every 30 minutes. The data was then transmitted in real-time to a domestically developed AI remote sensing cloud platform.

[0051] The second step involves standardization preprocessing of multi-source heterogeneous data. First, data cleaning is performed. Taking chlorophyll a time-series data from a specific monitoring point as an example, 48 consecutive hours of monitoring data are selected, and the mean is calculated. μg / L, standard deviation μg / L, according to the 3σ criterion, the effective data range is [missing value]. The concentration ranged from 6.3 μg / L to 18.9 μg / L. One outlier data point outside this range was removed, and linear interpolation was used for correction. Subsequently, spatiotemporal registration was performed to unify the multi-source data into a standardized spatiotemporal coordinate system with a 30-minute time step and a 5-meter spatial resolution. Format normalization, wavelet denoising, and Kalman filtering smoothing were then performed to obtain a standardized multi-source dataset.

[0052] The third step involves real-time fusion of multi-source data. A cloud-edge collaborative computing architecture is adopted, with edge nodes performing data preprocessing and feature extraction, followed by encrypted uploading to the cloud. The cloud uses a convolutional neural network to extract the spatial distribution features of algae from remote sensing images, and a long short-term memory network to extract the temporal features of algae growth from ground sensors. An attention mechanism is used to dynamically adjust the fusion weights. In this embodiment, to meet the algae monitoring requirements, the fusion weight of UAV remote sensing data is increased to 0.45, satellite remote sensing data to 0.30, and ground sensor data to 0.25. This fully leverages the advantages of high-resolution UAV data in identifying local algae clusters, resulting in a fused feature dataset with an end-to-end processing latency of 8 minutes.

[0053] The fourth step is to run the water quality dynamic monitoring model. Based on the fused feature dataset, a hybrid kernel function support vector machine is used to invert eutrophication indicators such as chlorophyll a and cyanobacteria density. The kernel function weight coefficients... The parameter was set to 0.55, and the model inversion error was controlled within 3.8%. A hidden Markov model was used to construct an algal growth trend prediction model. The model's triplet parameters were optimized, and the probability of algal density changes and eutrophication risk level for the next 72 hours were output. During the operation of this embodiment, two small-scale algal aggregation events were successfully predicted, and an eutrophication risk warning was issued 24 hours in advance, providing technical support for the safety of the reservoir's drinking water source.

[0054] Example 3: Emergency Dynamic Monitoring of Sudden Water Pollution Incidents The application scenario of this embodiment is the emergency dynamic monitoring of a sudden industrial wastewater leakage incident in a river. The incident occurred in an industrial park section of a river. The leaked wastewater contained high concentrations of organic pollutants, causing abnormal water quality in the downstream river section. The emergency monitoring objective is to quickly grasp the spatial distribution, migration path, and concentration change trend of the pollution plume, and provide real-time data support for emergency response.

[0055] The specific implementation process of this embodiment is as follows.

[0056] The first step was to perform multi-source heterogeneous emergency data acquisition. Satellite remote sensing data was immediately deployed to the domestically produced Gaofen-6 satellite for emergency imaging, acquiring effective images within 24 hours of the incident, covering a 30-kilometer stretch of the river upstream and downstream of the leak point. Remote sensing reflectance in sensitive wavelength bands of the water was extracted, and spectral characteristic factors were calculated. Taking image data 5 kilometers downstream of the leak point as an example, the remote sensing reflectance at a wavelength of 670nm is 0.018, and the remote sensing reflectance at a wavelength of 700nm is 0.032. Substituting these values ​​into the formula, the following was calculated: ; The remote sensing reflectance at a wavelength of 550 nm is 0.042, and the remote sensing reflectance at a wavelength of 850 nm is 0.008. Substituting these values ​​into the formula, the characteristic factor of suspended matter is calculated as follows: ; The drone remote sensing data was collected using a formation of multi-rotor drones equipped with multispectral cameras. Emergency aerial flights were conducted at an altitude of 150 meters, with a spatial resolution of 0.1 meters, a forward overlap of 80%, and a lateral overlap of 70%. Four flights were performed daily to acquire high-resolution spatial distribution data of the pollution plume in real time. Ground-based sensors employed emergency-deployed multi-parameter monitoring buoys, with 12 emergency monitoring points deployed at the leak point and key upstream and downstream sections. Data on pH, dissolved oxygen, chemical oxygen demand, and conductivity were collected once every 10 minutes. The data was uploaded in real time to a domestically developed AI remote sensing cloud platform via a 5G network.

[0057] The second step involves standardizing and preprocessing the multi-source heterogeneous data. The emergency rapid processing procedure is initiated, starting with data cleaning. Taking dissolved oxygen data from emergency monitoring points as an example, six sets of monitoring data from one consecutive hour are selected, and the mean value is calculated. mg / L, standard deviation mg / L, according to the 3σ criterion, the effective data range is [missing value]. The concentration ranged from 1.99 mg / L to 4.51 mg / L. After removing outliers, interpolation correction was performed. Subsequently, emergency rapid spatiotemporal registration was performed to unify the multi-source data into an emergency standardized coordinate system with a 10-minute time step and a 2-meter spatial resolution. Format normalization, rapid wavelet denoising, and Kalman filtering smoothing were completed, and the preprocessing process was completed within 5 minutes, outputting a standardized multi-source dataset.

[0058] The third step involves real-time fusion of multi-source data. A cloud-edge collaborative emergency computing model is adopted, with on-site edge computing nodes performing real-time data preprocessing and feature extraction, followed by encrypted uploading to the cloud central node. In the cloud, a convolutional neural network is used to quickly extract the spatial distribution features of the pollution plume, a long short-term memory network is used to extract real-time changes in water quality parameters, and an attention mechanism is used to dynamically adjust the fusion weights. In emergency scenarios, the fusion weight of real-time ground sensor data is increased to 0.55, UAV remote sensing data to 0.35, and satellite remote sensing data to 0.10, maximizing the weight ratio of real-time monitoring data to obtain a fused feature dataset. End-to-end processing latency is controlled within 10 minutes.

[0059] The fourth step involves implementing emergency operation of the water quality dynamic monitoring model. Based on the fused feature dataset, a hybrid kernel function support vector machine is used to complete the spatial inversion of core pollution indicators such as chemical oxygen demand (COD). The kernel function weight coefficients... Setting the coefficient to 0.65 enhances local fitting capabilities, keeping the model inversion error within 4.5%, and enabling rapid mapping of the spatial distribution of pollutant concentration. A pollutant migration prediction model is constructed using a Hidden Markov Model (HMM), and the model parameters are optimized to output the migration path, arrival time, and peak concentration prediction of the pollutant plume within the next 24 hours. In this embodiment, the model detects water quality anomalies and issues an early warning within 2 hours of the leak, continuously tracking the migration process of the pollutant plume. This provides minute-level real-time monitoring data support for emergency interception and water purification, effectively controlling the spread of pollution.

[0060] This invention addresses the technical pain points of insufficient depth in the fusion of multi-source heterogeneous data and inadequate utilization of feature complementarity in existing water quality monitoring technologies. It constructs a multi-source data deep fusion technology system that integrates spatial feature extraction, temporal feature mining, and dynamic weight fusion. Compared to the existing fusion method combining principal component analysis and linear weighting, this invention uses convolutional neural networks to specifically extract the spatial distribution features of water bodies from satellite and UAV remote sensing data, and uses long short-term memory networks to mine the dynamic changes in water quality from ground sensor temporal data. Simultaneously, it introduces an attention mechanism to achieve dynamic adaptive allocation of multi-source feature weights, effectively overcoming the technical bottleneck of poor adaptability to heterogeneous data features in traditional linear fusion methods, and significantly improving the accuracy and reliability of multi-source data fusion. Furthermore, this invention employs a standardized preprocessing process—wavelet transform denoising, Kalman filtering temporal smoothing, and spatiotemporal benchmark model construction—to effectively eliminate the spatiotemporal heterogeneity and noise interference of multi-source data, providing a high-quality data foundation for subsequent model construction. This fundamentally solves the problems of low model inversion accuracy and poor robustness caused by insufficient data preprocessing in existing technologies.

[0061] This invention achieves deep integration of the entire water quality monitoring technology system with the domestic software and hardware ecosystem, filling the application gap of domestic AI remote sensing cloud platforms in the field of multi-source heterogeneous water quality data fusion and dynamic monitoring. This invention constructs exclusive water quality sensitive feature factors based on the spectral characteristics of domestic Gaofen series satellites, and designs a refined flight and data processing flow adapted to the technical characteristics of domestic UAV remote sensing payloads. Simultaneously, the entire process algorithm and model have been adapted and optimized for domestic Ascend AI acceleration chips, Kylin operating systems, and PaddlePaddle deep learning frameworks. This achieves full-chain domestic independent control from data acquisition, data processing, feature fusion to model inference, forming a water quality monitoring technology solution with complete independent intellectual property rights, significantly improving the safety and compliance of the technology solution in key areas such as water conservancy and environmental protection in China.

[0062] This invention constructs a monitoring and prediction model system adapted to the dynamic changes in water quality. Simultaneously, through optimization of the cloud-edge collaborative computing architecture, it significantly improves the real-time performance and scenario adaptability of water quality monitoring, overcoming the technical shortcomings of existing static monitoring models, such as insufficient real-time performance and weak emergency response capabilities. This invention achieves high-precision inversion of water quality parameters through a hybrid kernel function support vector machine and advance prediction of water quality change trends through a hidden Markov model. Compared to existing solutions that only enable post-event pollution source tracing, this invention can achieve early warning of water quality anomalies and dynamic tracking of pollution development trends. Furthermore, the cloud-edge collaborative computing architecture adopted in this invention decentralizes data preprocessing and local feature extraction to edge nodes, effectively reducing end-to-end data processing latency. It is adaptable to various application scenarios, including routine watershed monitoring, lake eutrophication early warning, and emergency monitoring of sudden water pollution events. The technical solution possesses good feasibility and engineering implementation capabilities, providing stable and reliable technical support for refined water environment management and pollution emergency response.

[0063] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform, characterized by: Includes the following steps: S1 Multi-Source Heterogeneous Water Quality Data Acquisition: Satellite remote sensing multispectral / hyperspectral data of the target water area is acquired through domestic Gaofen series satellites, UAV remote sensing data of key areas of the target water area is acquired through UAVs equipped with multispectral imaging payloads, and in-situ water quality time series data is acquired through ground sensor arrays deployed in the target water area. All acquired data is transmitted to the domestic AI remote sensing cloud platform in real time. S2 Multi-Source Heterogeneous Data Standardization Preprocessing: Data cleaning, spatiotemporal registration, and format normalization are performed sequentially on the collected multi-source data. Wavelet transform denoising and Kalman filter temporal smoothing are then applied to obtain a standardized multi-source dataset with matching spatiotemporal dimensions. S3 Real-time fusion of multi-source data based on domestic AI framework: In the domestic AI remote sensing cloud platform, spatial features of remote sensing images are extracted by convolutional neural network (CNN), temporal features of ground time series data are extracted by long short-term memory network (LSTM), and attention mechanism is introduced to dynamically allocate the fusion weights of multi-source features to complete the real-time fusion of multi-source heterogeneous water quality data and obtain a fused feature dataset. S4 Water Quality Dynamic Monitoring Model Construction and Execution: Based on the fused feature dataset, water quality parameters are regressed and inverted using Support Vector Machine (SVM), and a water quality time series change prediction model is constructed using Hidden Markov Model (HMM). Real-time dynamic monitoring and trend prediction of water quality in the target water area are realized in the domestic AI remote sensing cloud platform.

2. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S2, the spatiotemporal registration process uses a unified spatiotemporal reference model to map multi-source data to the same spatiotemporal coordinate system. The calculation formula for the unified spatiotemporal reference model is as follows: in, For water quality data in a standardized spatiotemporal coordinate system, The original data, The time-dimensional linear registration transformation matrix is... The affine registration transformation matrix is ​​the spatial dimension. The wavelet transform denoising uses a hard threshold function, the formula of which is: in, These are the denoised wavelet coefficients. These are the original wavelet coefficients. An adaptive threshold calculated based on noise variance. To decompose the scale, This is the translation coefficient.

3. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S1, the satellite remote sensing data consists of multispectral and hyperspectral imagery data from the domestically produced Gaofen-6 and Gaofen-7 satellites. After radiometric calibration, atmospheric correction, and water body masking, the spectral reflectance of the sensitive bands of the water body is extracted to construct spectral feature factors for water quality parameter inversion. The calculation formula for the spectral feature factors is as follows: in, It is a chlorophyll a concentration-sensitive characteristic factor. As a sensitive characteristic factor of suspended solids concentration, wavelength Remote sensing reflectance of water bodies at a given location.

4. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S1, the acquisition and processing of UAV remote sensing data specifically includes the following sub-steps: S121 Track Planning: Based on the vector boundary of the target water area and the key areas of interest identified by satellite remote sensing, an adaptive track is generated, with a heading overlap of 80% and a lateral overlap of 70%. At the same time, denser sections are set for shoreline inflection points, sewage outlets into rivers, and algae accumulation areas. S122 Data Acquisition: Controls a fixed-wing UAV carrying a multispectral imaging payload to complete flight operations according to a planned trajectory, and simultaneously acquires multispectral image data, POS positioning and attitude data, and flight attitude data of the target water area; S123 Image Preprocessing: Radiometric calibration, orthorectification, image stitching, band registration, and water masking are performed sequentially on the acquired raw images to remove land and non-water areas, resulting in a sub-meter spatial resolution spectral feature dataset of key areas of the target water body, supplementing the local detail features of the satellite remote sensing data.

5. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S1, the deployment and data acquisition of the ground sensor array specifically includes the following sub-steps: S131 Monitoring Point Deployment: Monitoring points are deployed at the inlet, outlet, key pollution control area, shoreline sensitive area and background control area of ​​the target water area. Each point is equipped with an integrated multi-parameter water quality monitoring buoy to form a ground sensor array covering the entire water area. S132 Sensor Calibration: Before data acquisition, all sensors are calibrated in the laboratory and on-site in-situ. The calibration parameters are uploaded to the domestic AI remote sensing cloud platform and a sensor calibration log is established. S133 Time Series Data Acquisition: In-situ time series data of pH, dissolved oxygen, conductivity, turbidity, chemical oxygen demand, total phosphorus, and total nitrogen are synchronously acquired through a sensor array, with an acquisition frequency of no less than once per hour. S134 Real-time Data Transmission: The collected in-situ data is uploaded to the domestic AI remote sensing cloud platform in real time via a 5G / NB-IoT wireless network and an encrypted transmission channel, while data caching and backup are completed at the local edge node.

6. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S3, the dynamic weight allocation calculation formula for the attention mechanism is as follows: in, For the first The fusion weights of features from similar data sources For the first Salience score of data source features. For the trainable weight matrix of the attention mechanism, This represents the total number of data source categories. For the first Feature vectors extracted from class data sources This is the final output fused feature dataset.

7. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, Step S3, real-time fusion of multi-source data, adopts a cloud-edge collaborative computing architecture and specifically executes the following sub-steps: S31 edge node preprocessing: On domestically produced edge computing nodes deployed at the monitoring site, preliminary cleaning, outlier removal, local feature extraction and data compression of multi-source heterogeneous data are completed to reduce data transmission bandwidth usage; S32 encrypted data transmission: The edge nodes encrypt the preprocessed feature data using the national cryptographic algorithm and then upload it to the cloud center node of the domestic AI remote sensing cloud platform; S33 Cloud Global Fusion Processing: The cloud central node receives feature data uploaded by multiple nodes, completes global registration in the spatiotemporal dimension, performs global fusion of multi-source features based on CNN-LSTM-attention mechanism, and generates a fused feature dataset; S34 Result Synchronous Backhaul: The cloud synchronously backhauls the fusion processing results and model inference instructions to the corresponding edge nodes, ensuring that the end-to-end data processing delay does not exceed 15 minutes.

8. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S4, the Support Vector Machine (SVM) uses a hybrid kernel function to perform water quality parameter regression and inversion. The calculation formula for the hybrid kernel function is as follows: in, For radial basis kernel functions, For polynomial kernel functions, These are the kernel function weighting coefficients, with values ​​ranging from 1 to 2. , The input is the fused feature vector. These are support vectors.

9. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S4, the state transition probability matrix of the Hidden Markov Model (HMM) is iteratively optimized using the Baum-Welch algorithm, and the objective function is: in, These are the triplet parameters of the HMM model. Let be the initial state probability vector. The state transition probability matrix is... For the observation probability matrix, The model is a sequence of water quality parameter observations. It divides the water quality status into five discrete states: excellent, good, medium, poor, and bad, and outputs the probability of water quality status changes and the predicted values ​​of water quality parameters for the next 24 to 72 hours.

10. The method for real-time fusion and dynamic monitoring of multi-source heterogeneous water quality data based on a domestically developed AI remote sensing cloud platform according to claim 1, characterized in that, In step S2, the outlier detection and correction in the data cleaning process specifically includes the following sub-steps: S21 Single Data Source Gross Error Removal: Based on the 3σ criterion, the mean and standard deviation of time series data from a single data source are calculated, and invalid gross error data exceeding the range of mean ± 3 times the standard deviation are removed; S22 Multi-source data cross-validation: compares multi-source data from the same spatiotemporal location and marks abnormal data whose deviation from the measured values ​​of ground sensors exceeds a preset threshold; S23 Outlier Correction: For marked outlier data, linear interpolation of valid data from adjacent time periods at the same location is used to correct the outlier data. The corrected data is then incorporated into a standardized multi-source dataset.

Images