A water quality monitoring method and system based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based water quality monitoring method and system. It acquires a comprehensive dataset composed of sensor data, satellite images, and meteorological parameters. 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, outputting adjusted data. The adjusted data undergoes spatiotemporal interpolation to obtain a preprocessed dataset with high-density spatiotemporal coverage. Principal component analysis is used to extract key features from the preprocessed dataset, including sensor values, image textures, and meteorological parameters, constructing a weighted feature matrix to obtain a fused feature set. The dimensionality of the fused feature set is determined to be within a preset threshold. If the dimensionality exceeds the preset threshold, a random forest algorithm is used to reduce the dimensionality and classify the fused features. Cross-validation is then used to optimize the model parameters, yielding a pollution concentration prediction result. This invention provides effective technical support for water environment protection and has significant ecological and social benefits.

Description

Technical Field

[0001] This invention relates to the field of water quality monitoring technology, and in particular discloses a water quality monitoring method and system based on artificial intelligence. Background Technology

[0002] Traditional water quality monitoring methods rely on single-point sampling and manual analysis, resulting in narrow data coverage, poor timeliness, and difficulty in adapting to complex and ever-changing aquatic environments. This is particularly pronounced in watersheds or groundwater systems where pollution incidents are frequent, where response lag and insufficient accuracy are especially prominent. These limitations prevent existing methods from meeting the demands for real-time, accurate, and comprehensive monitoring.

[0003] In the field of water quality monitoring, the core challenges stem from the inefficiency and lack of intelligence in data acquisition and analysis. Static sampling methods limit the spatiotemporal coverage density of data, making it difficult to capture early signals of pollution events. Consequently, the high-dimensional heterogeneity of the data further exacerbates the processing difficulties, and traditional methods struggle to effectively integrate multi-source data, such as sensor data, image data, and meteorological information, resulting in inefficient feature extraction and analysis.

[0004] Furthermore, the predictive power of the model is limited. Faced with nonlinear water quality changes and complex pollution diffusion patterns, existing algorithms struggle to achieve accurate classification, trend prediction, and pollution source tracing.

[0005] Ultimately, the lack of dynamic early warning and decision support mechanisms prevents the monitoring system from shifting from passive response to proactive prevention and control.

[0006] Therefore, how to build a water quality monitoring system that integrates dynamic data acquisition, multi-source feature fusion, intelligent prediction and real-time early warning to overcome the bottlenecks of insufficient data coverage, inefficient feature processing and limited prediction accuracy has become a key issue in promoting the upgrading of water quality monitoring technology paradigm. Summary of the Invention

[0007] This invention provides a water quality monitoring method and system based on artificial intelligence, aiming to solve at least one of the defects of the prior art.

[0008] Acquire a comprehensive dataset consisting of sensor data, satellite imagery, and meteorological parameters;

[0009] 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, and the adjusted data is output.

[0010] Spatiotemporal interpolation is performed on the adjusted data to obtain a preprocessed dataset with high-density spatiotemporal coverage;

[0011] Principal component analysis was used to extract key features of sensor values, image textures and meteorological parameters from the preprocessed dataset, and a weighted feature matrix was constructed to obtain a fused feature set.

[0012] Determine whether the dimension of the fused feature set exceeds a preset threshold. If the dimension of the fused feature set exceeds the preset threshold, use the random forest algorithm to reduce the dimensionality and classify the fused features. Optimize the model parameters through cross-validation to obtain the pollution concentration prediction result.

[0013] The spatial pattern of the pollution concentration prediction results is analyzed by using convolutional neural networks, and the pollution migration path is calculated by combining water flow field data. The location of the pollution source is determined through iterative optimization, and the pollution source tracing results are obtained.

[0014] Determine whether the confidence level of the pollution source tracing results is higher than a preset threshold. If the confidence level of the pollution source tracing results is higher than the preset threshold, a real-time early warning model is constructed using a Bayesian network, and dynamic early warning signals are generated by integrating the pollution concentration prediction results and the pollution source tracing results.

[0015] The dynamic early warning signal is input into the push algorithm to transmit the early warning data to the decision support system.

[0016] Furthermore, the steps for acquiring a comprehensive dataset consisting of sensor data, satellite imagery, and meteorological parameters include:

[0017] Collect and preprocess multi-source data: Obtain environmental parameters, satellite images and meteorological parameters through sensor equipment, satellite remote sensing platform and meteorological station database, and use format conversion tools to standardize the multi-source data to obtain a standardized dataset;

[0018] Perform time and space alignment: synchronize the standardized dataset in time based on timestamps. If timestamps are missing, they are filled in by linear interpolation. Spatial alignment of the standardized dataset is performed using a geographic coordinate matching method to obtain a time-space synchronized dataset.

[0019] Handling missing and anomalies: For time-space synchronized datasets, if sensor data is missing, it is filled by interpolation using the K-nearest neighbor algorithm. If meteorological parameters exceed a preset threshold, they are marked as anomalies and removed to obtain a cleaned dataset.

[0020] Data fusion and correlation calculation: The cleaned datasets are fused using a weighted average method, and the correlation of the fused data is calculated using the Pearson correlation coefficient to obtain a comprehensive dataset.

[0021] Furthermore, 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, and the steps for outputting the adjusted data include:

[0022] Water flow velocity and pollution concentration data are obtained from a comprehensive dataset. A two-dimensional gradient operator is used to calculate the pollution concentration gradient, resulting in a concentration gradient distribution dataset. The gradient operator is defined as follows: C represents the pollution concentration, and x and y represent spatial coordinates;

[0023] Based on the concentration gradient distribution dataset, if the concentration gradient value is greater than the preset threshold, the high gradient region is divided into sampling locations using a grid partitioning method, and the sampling frequency is calculated using a weighted average method to obtain a preliminary adjusted sampling scheme.

[0024] For the initially adjusted sampling scheme, the K-means clustering algorithm was used to optimize the sampling locations by grouping them, and the sampling frequency was synchronized by timestamps to obtain the optimized sampling scheme.

[0025] The adjusted sampling frequency and location data are obtained from the optimized sampling scheme. The adjusted sampling frequency and location data are integrated using a data fusion method to output the adjusted data.

[0026] Furthermore, the steps of performing spatiotemporal interpolation on the adjusted data to obtain a preprocessed dataset with high-density spatiotemporal coverage include:

[0027] The sampling frequency, sampling location, and timestamp are obtained from the adjusted data. The timestamps are sorted using time series analysis methods. If the timestamp interval is greater than a preset threshold, the gaps in the time series are filled by linear interpolation to obtain the time-synchronized sampling data.

[0028] Based on the time-synchronized sampling data, the pollution concentration and water flow velocity are obtained. The sampling location is mapped to a two-dimensional coordinate system using a spatial grid division method. If the pollution concentration gradient within the grid is greater than a preset threshold, spatial interpolation is performed using the Kriging interpolation method to obtain high-density spatial distribution data.

[0029] Data gaps are identified in high-density spatially distributed data. Pollution concentration and water flow velocity are integrated using a data fusion method. If the gaps are located at the spatial grid boundary, they are filled using a weighted average method to obtain a spatiotemporally continuous dataset.

[0030] For spatiotemporally continuous datasets, interpolation accuracy metrics are obtained, and interpolation parameters are adjusted using an iterative interpolation method. If the interpolation accuracy is lower than a preset threshold, the iterative adjustment is repeated to obtain a preprocessed dataset with high-density spatiotemporal coverage.

[0031] Furthermore, the steps of extracting key features of sensor values, image textures, and meteorological parameters from the preprocessed dataset using principal component analysis to construct a weighted feature matrix and obtain the fused feature set include:

[0032] Sensor values, image textures, and meteorological parameters are obtained from the preprocessed dataset. Principal component analysis is used to standardize the data dimensions, calculate the covariance matrix, and obtain eigenvalues ​​and eigenvectors. If the cumulative contribution rate of the eigenvalues ​​is greater than a preset threshold, the corresponding eigenvector is selected to obtain the dimensionality-reduced key feature set.

[0033] Based on the key feature set, the variance contribution rate of each feature is obtained, and the weight is calculated using the entropy method. If the weight is lower than the preset threshold, the weight is reset to zero. The weight and feature are integrated by matrix multiplication to obtain the weighted feature matrix.

[0034] Feature subsets of sensor values, image textures, and meteorological parameters are obtained from the weighted feature matrix and stitched together using a data fusion method. If the dimension after stitching exceeds a preset threshold, it is adjusted by a linear dimensionality reduction method to obtain a preliminary fused feature set.

[0035] Based on the preliminary fusion feature set, the correlation coefficient matrix between features is obtained. Clustering method is used to group features with correlation coefficients higher than a preset threshold. The features are then integrated using a weighted average method within the group to obtain the final fusion feature set.

[0036] Further, the steps include determining whether the dimensionality of the fused feature set exceeds a preset threshold. If the dimensionality exceeds the preset threshold, a random forest algorithm is used to reduce the dimensionality and classify the fused features. The model parameters are then optimized through cross-validation to obtain the pollution concentration prediction results.

[0037] Feature dimension data is obtained from the fused feature set, and the distribution characteristics of the feature dimension data are calculated using statistical methods. If the feature dimension exceeds a preset threshold, the feature dimension data is sorted by importance using the random forest algorithm to obtain a feature subset after dimensionality reduction.

[0038] Based on the feature subset after dimensionality reduction, the training set and validation set are obtained. The cross-validation method is used to optimize the tree depth and number of trees of the random forest, and the optimized model parameter set is obtained.

[0039] The classification labels are obtained from the optimized model parameter set, the random forest algorithm is used to train the classification on the training set, the classification performance is evaluated through the validation set, and the predicted probability distribution of the classification labels is obtained.

[0040] Based on the predicted probability distribution of the classification labels, the mapping function of pollution concentration is obtained, and the predicted probability distribution is fitted using the linear regression method to obtain the predicted pollution concentration.

[0041] Furthermore, the steps of analyzing the spatial pattern of pollution concentration prediction results using convolutional neural networks, calculating pollution migration paths by combining water flow field data, and determining the pollution source location through iterative optimization to obtain pollution source tracing results include:

[0042] Concentration distribution data is obtained from the pollution concentration prediction results, and spatial pattern extraction is performed on the concentration distribution data using a convolutional neural network to obtain a spatial pattern feature set.

[0043] Based on the spatial pattern feature set and combined with water flow field data, a fluid dynamics algorithm is used to calculate the pollution migration path and obtain the pollution migration path set.

[0044] Determine whether the number of paths in the pollution migration path set exceeds a preset threshold. If the number of paths in the pollution migration path set exceeds the preset threshold, then use an iterative optimization algorithm to filter and cluster the path set to determine the candidate set of pollution source locations.

[0045] For the candidate set of pollution source locations, the probability distribution of each candidate location is calculated using the Bayesian inference method to obtain the source tracing results of the pollution source locations.

[0046] Furthermore, the confidence level of the pollution source tracing results is determined to be higher than a preset threshold. If the confidence level of the pollution source tracing results is higher than the preset threshold, a real-time early warning model is constructed using a Bayesian network. The steps of fusing the pollution concentration prediction results and the pollution source tracing results to generate a dynamic early warning signal include:

[0047] Concentration distribution data is obtained from the pollution concentration prediction results. Bayesian network is used to perform probabilistic modeling of the concentration distribution data. Combined with flow field data, the dynamic probability distribution of pollution migration paths is calculated to obtain the pollution migration path set.

[0048] If the number of paths in the pollution migration path set exceeds a preset threshold, the pollution migration path set is filtered by a clustering algorithm, similar migration paths are merged, and a candidate location set is determined.

[0049] For the candidate location set, the confidence level of each candidate location is calculated using the Bayesian inference method. The confidence level distribution is then generated by combining the source tracing results to obtain the probability distribution of the pollution source location.

[0050] Based on the probability distribution, a dynamic signal generation algorithm is used to fuse the probability distribution and concentration prediction results to generate a dynamic signal for real-time early warning.

[0051] Furthermore, the steps of inputting dynamic early warning signals into the push algorithm and transmitting early warning data to the decision support system include:

[0052] Signal feature data is obtained from dynamic early warning signals, and a classification algorithm is used to prioritize the signal feature data. A priority queue is generated by combining the data with preset push rules to obtain a push priority sequence.

[0053] Based on the push priority sequence, the dynamic early warning signal is converted into standardized early warning data using a data encapsulation method. Combined with the preset data format specifications, a formatted data packet is generated to determine the content of the transmitted data.

[0054] For formatted data packets, a transmission protocol is used to fragment the formatted data packets. During fragmentation, a checksum algorithm is combined to generate a checksum, resulting in a transmission data unit with checksum.

[0055] Data units are transmitted to the decision support system via a data transmission channel. The received data units are reassembled and verified using an asynchronous processing method to determine data integrity and generate an acknowledgment signal.

[0056] Another aspect of the present invention relates to an artificial intelligence-based water quality monitoring system for implementing the aforementioned artificial intelligence-based water quality monitoring method. The artificial intelligence-based water quality monitoring system includes:

[0057] The first acquisition module is used to acquire a comprehensive dataset consisting of sensor data, satellite images, and meteorological parameters;

[0058] The adjustment module is used to adjust the sampling frequency and location based on the water flow velocity and pollution concentration gradient in the comprehensive dataset using a dynamic sampling algorithm, and output the adjusted data.

[0059] The second acquisition module is used to perform spatiotemporal interpolation on the adjusted data to obtain a preprocessed dataset with high-density spatiotemporal coverage.

[0060] The third acquisition module is used to extract key features of sensor values, image textures and meteorological parameters from the preprocessed dataset using principal component analysis, construct a weighted feature matrix and obtain a fused feature set.

[0061] The fourth acquisition module is used to determine whether the dimension of the fused feature set exceeds a preset threshold. If the dimension of the fused feature set exceeds the preset threshold, the random forest algorithm is used to reduce the dimensionality and classify the fused features. The model parameters are optimized through cross-validation to obtain the pollution concentration prediction result.

[0062] The fifth acquisition module is used to analyze the spatial pattern of the pollution concentration prediction results using a convolutional neural network, calculate the pollution migration path by combining water flow field data, determine the location of the pollution source through iterative optimization, and obtain the pollution source tracing results.

[0063] The generation module is used to determine whether the confidence level of the pollution source tracing result is higher than the preset threshold. If the confidence level of the pollution source tracing result is higher than the preset threshold, a real-time early warning model is constructed using a Bayesian network, and the pollution concentration prediction result and the pollution source tracing result are integrated to generate a dynamic early warning signal.

[0064] The transmission module is used to input dynamic early warning signals into the push algorithm and transmit early warning data to the decision support system.

[0065] The beneficial effects achieved by this invention are as follows:

[0066] This invention provides an artificial intelligence-based water quality monitoring method and system. By acquiring multi-source data, a dynamic sampling algorithm is used to optimize data collection. The data undergoes spatiotemporal interpolation and feature extraction to construct a fused feature set. A random forest algorithm is used to reduce the dimensionality of the features and classify them to predict pollution concentrations. A convolutional neural network is combined to analyze pollution spatial patterns, calculate pollution migration paths, and determine the location of pollution sources. Finally, a real-time early warning model is built based on a Bayesian network to generate dynamic early warning signals and push them to a decision support system. This invention achieves intelligent monitoring, precise source tracing, and timely early warning of water pollution, providing effective technical support for water environment protection and possessing significant ecological and social benefits. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating an embodiment of an artificial intelligence-based water quality monitoring method according to the present invention. Detailed Implementation

[0068] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0069] like Figure 1 As shown, the first embodiment of the present invention proposes a water quality monitoring method based on artificial intelligence, including the following steps:

[0070] Step S100: Acquire a comprehensive dataset consisting of sensor data, satellite images, and meteorological parameters.

[0071] A comprehensive dataset is a standardized collection of structured data, including sensor data, satellite imagery, and meteorological parameters, which is typically used for analysis, modeling, or processing.

[0072] Step S200: 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, and the adjusted data is output.

[0073] Water flow velocity refers to the speed at which water flows in a specific direction, that is, the distance the water travels per unit time.

[0074] Pollution concentration gradient refers to the rate of change of pollutant concentration in spatial distribution, characterizing the magnitude and direction of increase or decrease in pollutant concentration per unit distance.

[0075] Dynamic sampling algorithms are sampling techniques that adjust in real time based on data streams or dynamic datasets. They aim to approximate global features using local data, thereby reducing resource consumption and improving efficiency in large-scale data processing.

[0076] Adjusting data refers to changing the distribution, quantity, or structure of original data through specific technical means to adapt to model requirements or optimize system performance.

[0077] Step S300: Perform spatiotemporal interpolation on the adjusted data to obtain a preprocessed dataset with high-density spatiotemporal coverage.

[0078] Spatiotemporal interpolation is a data reconstruction technique that combines spatial and temporal dimensions. It aims to infer values ​​for unobserved areas or time points from data at known, discrete spatiotemporal observation points, generating a continuous and spatiotemporally correlated data distribution. Its core objective is to address the uneven spatiotemporal distribution of monitoring data caused by acquisition limitations, providing a complete dataset for subsequent analyses (such as environmental modeling and trend prediction).

[0079] High-density spatiotemporal coverage achieves refined observation capabilities of environmental parameters (such as soil moisture, groundwater level, and pollutant concentration) in a target area through dense monitoring point layout and high-frequency data acquisition, ensuring both temporal and spatial continuity. The core objective of high-density spatiotemporal coverage is to overcome the limitations of insufficient spatiotemporal resolution in traditional monitoring and establish a dynamic monitoring network that is "blind-spot-free and operates around the clock."

[0080] Preprocessed datasets refer to structured data sets that have undergone systematic technical processes to clean, transform, and reconstruct raw data, eliminating noise, standardizing formats, and adapting them to specific analytical needs. The core objective of preprocessed datasets is to improve data quality and usability, providing reliable input for subsequent modeling or data mining tasks.

[0081] Step S400: Use principal component analysis to extract key features of sensor values, image textures and meteorological parameters from the preprocessed dataset, construct a weighted feature matrix, and obtain a fused feature set.

[0082] Principal Component Analysis (PCA) is a linear dimensionality reduction method that projects high-dimensional data into a low-dimensional space through orthogonal transformations. Its core objective is to extract the directions with the largest variance in the data (principal components), preserving the main information of the original data in its simplest form while eliminating redundancy and noise. PCA achieves data reconstruction through eigenvalue decomposition of the covariance matrix and is widely used in data compression, feature selection, and visualization.

[0083] A weighted feature matrix typically refers to a matrix representation method that introduces weights in feature analysis to adjust the importance of different features.

[0084] Feature fusion refers to the process of systematically integrating multiple feature sets or data sources using specific methods. The aim is to leverage the complementarity between different features to generate a composite feature set with stronger representational capabilities. This fusion can overcome the limitations of single features and enhance the model's adaptability to complex data. For example, in computer vision, multi-scale feature fusion can combine low-resolution location information with high-resolution detail semantics to form a more complete visual representation.

[0085] Step S500: Determine whether the dimension of the fused feature set exceeds the preset threshold. If the dimension of the fused feature set exceeds the preset threshold, use the random forest algorithm to reduce the dimensionality and classify the fused features, and optimize the model parameters through cross-validation to obtain the pollution concentration prediction result.

[0086] Random forests integrate multiple decision trees, combining the Bagging (Bootstrap Aggregating) method with a random feature selection mechanism, and ultimately output the results through voting (classification tasks) or averaging (regression tasks). Their core objective is to reduce the overfitting risk of a single decision tree and improve the model's generalization ability.

[0087] Cross-validation is a statistical technique that evaluates the predictive power of a model by splitting the dataset and repeating the validation process. The core objective of cross-validation is to make full use of limited data, reduce performance evaluation bias caused by a single data split (such as training / test set split), and effectively detect overfitting or underfitting problems.

[0088] Pollution concentration prediction results refer to the numerical output that quantitatively estimates the unit volume concentration (e.g., micrograms per cubic meter) of pollutants (such as PM2.5, SO2, etc.) in a specific future period and region by using mathematical models, statistical methods, or machine learning techniques, combined with meteorological conditions, pollution source emission data, and historical observation information. The core objective of pollution concentration prediction results is to predict environmental quality change trends in advance and provide a scientific basis for pollution prevention and control decisions.

[0089] Step S600: Use a convolutional neural network to analyze the spatial pattern of the pollution concentration prediction results, combine it with water flow field data to calculate the pollution migration path, determine the location of the pollution source through iterative optimization, and obtain the pollution source tracing results.

[0090] Convolutional neural networks (CNNs) automatically extract multi-level features from input data through mechanisms such as local receptive fields, weight sharing, and spatial downsampling, ultimately achieving tasks such as classification, detection, or generation. Their core advantage lies in their robustness to geometric transformations such as translation, scaling, and rotation, making them particularly suitable for image processing.

[0091] Spatial patterns refer to local or global feature structures in data (such as images, videos, or geographic information) that exhibit specific spatial distribution patterns, manifested as ordered arrangements of pixels, voxels, or geometric elements. In convolutional neural networks (CNNs), spatial patterns are extracted layer by layer through the sliding scan of convolutional kernels, forming a hierarchical representation from low-level edges to high-level semantic features.

[0092] Water flow field data refers to a set of multi-dimensional parameters describing the flow state of water bodies, obtained through observation or numerical simulation. Core elements include flow velocity, flow direction, flow regime, and their spatiotemporal distribution characteristics. This type of data can characterize the motion patterns of water bodies, such as the stratified characteristics of laminar and turbulent flow, or the dynamic evolution of vortex structures, and serves as the scientific basis for hydrological analysis, environmental assessment, and engineering design.

[0093] Pollution migration pathways refer to the specific channels or trajectories driven by physical, chemical, or biological processes that pollutants follow when they undergo spatial transfer within the environment. This concept encompasses the transport routes of pollutants across different media (such as the atmosphere, water bodies, and soil), as well as the direction and extent of diffusion within the same medium. Its core focus is on how pollutants change location through the dynamic movement of environmental media, accompanied by changes in concentration distribution (such as enrichment or dilution).

[0094] Iterative optimization refers to a dynamic improvement process that gradually approaches the optimal solution through repeated adjustments and feedback mechanisms. Its core lies in using the calculation results of the initial scheme to modify parameters or strategies to improve performance indicators (such as efficiency, cost, and accuracy) until preset convergence conditions or constraints are met. This process is widely used in path planning, model calibration, and strategy decision-making.

[0095] Pollution source tracing results refer to the conclusive data set obtained through systematic analysis and verification of the source, diffusion path, emission intensity, and impact range of pollutants using scientific methods and technical means. Its core objective is to identify the responsible parties for pollution (such as industrial enterprises, agricultural activities, etc.) and to reveal the migration patterns of pollutants in the spatiotemporal dimensions (such as transmission through water bodies, the atmosphere, or the food chain), thereby providing quantifiable evidence for environmental governance and legal accountability.

[0096] Step S700: Determine whether the confidence level of the pollution source tracing result is higher than the preset threshold. If the confidence level of the pollution source tracing result is higher than the preset threshold, a real-time early warning model is constructed using a Bayesian network, and a dynamic early warning signal is generated by integrating the pollution concentration prediction result and the pollution source tracing result.

[0097] A Bayesian Network is a statistical model based on probability theory and graph theory. It uses a directed acyclic graph (DAG) to represent the dependencies between random variables and combines conditional probability tables (CPTs) to quantify the strength of the associations between variables. Its core function is to process uncertain information through probabilistic reasoning (such as prediction, diagnosis, or causal inference), and it is widely used in fields such as environmental risk assessment, pollution source apportionment, and complex system modeling.

[0098] Real-time early warning models are computational frameworks based on dynamic data acquisition and real-time analysis. They aim to rapidly identify abnormal emissions or pollution spread risks by continuously monitoring environmental parameters (such as pollutant concentrations, water flow rates, and meteorological conditions) and combining them with predictive algorithms (such as AI models and diffusion simulations). Upon triggering a threshold, they generate tiered alarm signals. Their core function is to provide decision support for emergency response to pollution incidents through real-time feedback and adaptive adjustments.

[0099] Dynamic early warning signals are multi-level risk alerts based on real-time monitoring data and predictive models. Through dynamic threshold determination and spatiotemporal correlation analysis, they reflect the real-time evolution and potential severity of environmental risks (such as pollution spread, meteorological disasters, and ecological anomalies). Their core feature is automatic escalation / degradation as risk changes (e.g., transitioning from a "yellow" to a "red" alert) and support for multi-modal output (such as map heatmaps, SMS push notifications, and audible and visual alarms) to adapt to different emergency scenarios.

[0100] Step S800: Input the dynamic early warning signal into the push algorithm and transmit the early warning data to the decision support system.

[0101] Push algorithms are intelligent decision-making models based on multimodal data integration and real-time analysis. They aim to dynamically collect, fuse, and analyze multi-source information such as user behavior, environmental parameters, and device status, and combine this with preset rules or machine learning models to generate personalized instruction distribution strategies. This enables targeted, tiered, and multi-channel (such as SMS, app push, and audio-visual signals) precise information delivery. Its core value lies in optimizing the timeliness, accuracy, and user response efficiency of information delivery through data-driven approaches and scenario adaptation.

[0102] A Decision Support System (DSS) is an intelligent information system that integrates data analysis, model inference, and human-computer interaction. It aims to provide dynamic risk assessment, solution simulation, and strategy recommendation for complex decision-making scenarios by fusing multi-source data (such as historical data, real-time sensor signals, and expert knowledge bases) and combining probabilistic models (such as Bayesian networks) with optimization algorithms. Its core function is to enhance decision-makers' judgment capabilities in situations with incomplete information or dynamic environments through uncertainty reasoning and interpretable output.

[0103] Furthermore, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S100 includes:

[0104] Step S110: Collect and preprocess multi-source data: Obtain environmental parameters, satellite images, and meteorological parameters through sensor devices, satellite remote sensing platforms, and meteorological station databases. Use format conversion tools to standardize the multi-source data to obtain a standardized dataset.

[0105] Collecting and preprocessing multi-source data involves acquiring environmental data from different sources and standardizing the format. For example, air quality is monitored using sensor devices to obtain PM2.5 and temperature data; vegetation cover images of urban areas are obtained through satellite remote sensing platforms; and rainfall and wind speed data are obtained from meteorological station databases.

[0106] These data may be in different formats, such as CSV for sensor data, GeoTIFF for satellite imagery, and JSON for meteorological data. Format conversion tools, such as GDAL, are used to process satellite imagery, converting it to a uniform grid resolution; the Pandas library is used to convert CSV and JSON into a uniform table structure, generating a standardized dataset. This preprocessing ensures data compatibility and facilitates subsequent analysis.

[0107] Step S120: Perform time and space alignment: Synchronize the standardized dataset in time based on the timestamps. If the timestamps are missing, they are filled in by linear interpolation. The standardized dataset is spatially aligned using a geographic coordinate matching method to obtain a time-space synchronized dataset.

[0108] In one possible implementation, execution time and space alignment needs to address the temporal and spatial inconsistencies of multi-source data.

[0109] Time synchronization is based on timestamps; for example, sensors record data every 10 minutes, and weather stations record data every hour. If meteorological data for a certain point in time is missing, it can be estimated through linear interpolation, such as calculating the intermediate time value based on rainfall in the two hours before and after.

[0110] Spatial alignment is based on geographic coordinates, such as aligning sensor locations with grid points in satellite imagery using nearest neighbor matching, and unifying coordinates from different data sources to the same projection system, such as WGS84.

[0111] The time-space synchronized dataset provides a unified time-space framework for subsequent analysis, improving data consistency.

[0112] Step S130: Handling missing and abnormal data: For the time-space synchronization dataset, if there are missing sensor data, interpolation is performed using the K-nearest neighbor algorithm to fill the gaps. If the meteorological parameters exceed the preset threshold, they are marked as abnormal and removed to obtain a cleaned dataset.

[0113] Specifically, handling missing and anomalies involves further cleaning the time-space synchronized dataset.

[0114] If PM2.5 data is missing due to sensor malfunction, the K-nearest neighbor algorithm can be used to interpolate and fill the missing data based on values ​​from nearby time points or nearby sensors. For example, if a sensor is missing data at a certain moment, a weighted average can be calculated by referring to data from three nearby sensors at the same time to fill the missing data.

[0115] If meteorological parameters are abnormal, such as wind speed exceeding 50 m / s, which is outside the physically reasonable range, the data is marked as abnormal and removed. This cleaning process ensures data quality and reduces noise interference with the analysis.

[0116] Step S140: Merge data and calculate correlation: Merge the cleaned dataset using a weighted average method, and calculate the correlation of the merged data using the Pearson correlation coefficient to obtain a comprehensive dataset.

[0117] Preferably, fusing data and calculating correlations is a step of integrating multi-source data and mining relationships.

[0118] Data is fused using a weighted average method. For example, for PM2.5 data, the sensor weight is set to 0.6 and the satellite estimation weight is set to 0.4. The fused data then generates a comprehensive PM2.5 distribution.

[0119] Pearson correlation coefficient analysis was used to analyze the relationships between fused data. For example, the correlation between PM2.5 and rainfall was calculated, revealing a negative correlation, indicating that rainfall may reduce particulate matter concentration. This fusion and analysis reveals the interaction of environmental factors, providing data support for pollution control.

[0120] Understandably, the implementation of the above methods significantly improves the accuracy and reliability of environmental monitoring. For example, standardization facilitates the integration of multi-source data, temporal and spatial alignment eliminates data bias, the cleaning process improves data quality, and fusion and correlation analysis uncover hidden patterns. These technical effects support precise environmental management, such as optimizing air quality early warning systems or guiding urban greening planning.

[0121] Preferably, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S200 includes:

[0122] Step S210: Obtain water flow velocity and pollution concentration data from the comprehensive dataset, calculate the pollution concentration gradient using a two-dimensional gradient operator, and obtain the concentration gradient distribution dataset; where the gradient operator is defined as... C represents the pollution concentration, and x and y represent spatial coordinates.

[0123] For example, when obtaining water flow velocity and pollution concentration data from a comprehensive dataset, relevant fields can be filtered through database queries. For instance, water flow velocity data might originate from river sensors, recording a flow velocity of, say, 0.5 m / s.

[0124] Pollution concentration data comes from water quality monitoring stations, recording chemical oxygen demand (COD) concentrations such as 20 mg / L.

[0125] It should be noted that data extraction must ensure consistent timestamps, such as recording once per hour, to avoid affecting subsequent analysis due to time discrepancies.

[0126] In one possible implementation, the data can be stored in a relational database, and SQL queries can be used to extract flow rate and concentration fields by time and location, generating a data table containing time, coordinates, flow rate, and concentration. This extraction method facilitates rapid location of target data.

[0127] Specifically, when using a two-dimensional gradient operator to calculate the pollution concentration gradient, the gradient reflects the rate of change of concentration in space. For example, in a certain river section, the pollution concentration changes from 20 mg / L to 25 mg / L in the x-direction and from 20 mg / L to 22 mg / L in the y-direction. The concentration gradient distribution map can be obtained using the gradient operator.

[0128] Preferably, numerical methods such as finite difference approximation can be used to calculate the gradient, generating a gradient vector field based on gridded concentration data. This method intuitively reflects the pollution diffusion trend and facilitates the identification of areas with high variability.

[0129] Step S220: Based on the concentration gradient distribution dataset, if the concentration gradient value is greater than the preset threshold, the sampling location of the high gradient region is divided by the grid division method, and the sampling frequency is calculated by the weighted average method to obtain the preliminary adjusted sampling scheme.

[0130] In one embodiment, when identifying high gradient regions based on a concentration gradient distribution dataset, a threshold such as 0.5 mg / L / m can be set.

[0131] If a gradient value in a certain area reaches 0.7 mg / L / m, exceeding a threshold, then the area is divided into grids. For example, the area might be divided into a 5×5 grid, with each grid cell having a side length of 10 m. The sampling frequency is calculated using a weighted average method. The weights can be based on the magnitude of the gradient value; for example, higher gradient grids are assigned higher weights, resulting in a sampling frequency of once every 2 hours per grid cell. This division and frequency adjustment ensures that sampling points cover areas with drastic changes in pollution levels.

[0132] Step S230: For the initially adjusted sampling scheme, the K-means clustering algorithm is used to group and optimize the sampling locations, and the sampling frequency is synchronized by timestamps to obtain the optimized sampling scheme.

[0133] For example, for the initially adjusted sampling scheme, the K-means clustering algorithm is used to optimize the grouping of sampling locations. Assuming there are 100 sampling points, they are clustered into 5 groups, each representing areas with similar pollution characteristics. The clustering is based on location coordinates and gradient values; after optimization, the sampling points within each group are more concentrated, reducing redundancy.

[0134] Time synchronization is based on timestamps, for example, unifying the sampling frequency to the hour, such as 8:00 or 10:00. This optimization makes the sampling scheme more efficient, covering key areas while reducing costs.

[0135] Step S240: Obtain the adjusted sampling frequency and location data from the optimized sampling scheme, integrate the adjusted sampling frequency and location data using a data fusion method, and output the adjusted data.

[0136] Understandably, after obtaining sampling frequency and location data from the optimized sampling scheme, a data fusion method is used to integrate them. For example, multi-source data such as sensor concentration data and model-predicted concentration data are fused with weights of 0.7 and 0.3, respectively, to generate a comprehensive concentration distribution. The fused data can be output as a table containing time, location, frequency, and concentration fields. This integration improves data consistency and facilitates subsequent dynamic pollution analysis.

[0137] Furthermore, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S300 includes:

[0138] Step S310: Obtain the sampling frequency, sampling location and timestamp from the adjusted data, sort the timestamps using time series analysis methods, and if the timestamp interval is greater than a preset threshold, fill the time series gaps using linear interpolation methods to obtain the time-synchronized sampling data.

[0139] For example, when obtaining the sampling frequency, sampling location, and timestamp from the adjusted data, the relevant fields can be extracted through a database query. For instance, the sampling frequency is recorded as once every 2 hours, the sampling location is expressed in latitude and longitude, such as (120.5, 30.2), and the timestamp is 2025-04-30 08:00:00.

[0140] It should be noted that timestamp sorting must ensure that the data is arranged in chronological order to avoid errors in subsequent analysis.

[0141] In one possible implementation, time series analysis is used to examine the timestamp intervals. If the preset threshold is 1 hour and a certain timestamp interval is 2 hours, then a linear interpolation method is used to fill in the gaps. For example, if the concentrations at 8:00 and 10:00 are 20 mg / L and 24 mg / L respectively, the interpolated concentration at 9:00 is approximately 22 mg / L. This method ensures the continuity of the time series, facilitating subsequent analysis.

[0142] Step S320: Based on the time-synchronized sampling data, obtain the pollution concentration and water flow velocity. Use the spatial grid division method to map the sampling location to a two-dimensional coordinate system. If the pollution concentration gradient within the grid is greater than a preset threshold, then use the Kriging interpolation method to perform spatial interpolation to obtain high-density spatial distribution data.

[0143] Specifically, when acquiring pollution concentration and water flow velocity based on time-synchronized data and mapping them to a two-dimensional coordinate system, a spatial grid partitioning method can be used. For example, the sampling area can be divided into a 10×10 grid, with each grid having a side length of 100m.

[0144] Assuming the pollution concentration gradient within a certain grid is 0.6 mg / L / m³, exceeding a preset threshold of 0.5 mg / L / m³, high-density distribution data is generated using Kriging interpolation. Kriging interpolation is based on spatial autocorrelation and preferably uses the concentration values ​​of surrounding sampling points, such as 20 mg / L and 25 mg / L, to estimate the concentration values ​​of unsampled points within the grid. This method improves spatial resolution and is suitable for analyzing pollution distribution trends.

[0145] Step S330: Obtain missing data points from high-density spatial distribution data, and use data fusion methods to integrate pollution concentration and water flow velocity. If the missing points are located at the spatial grid boundary, fill the missing points using a weighted average method to obtain a spatiotemporally continuous dataset.

[0146] In one embodiment, missing data points are identified from high-density spatially distributed data, and a data fusion method is used to integrate pollution concentration and water flow velocity. For example, a missing point may be located at a grid boundary, where concentration data is missing but flow velocity data is 0.4 m / s. A weighted averaging method can be used, combining concentration values ​​from neighboring grids, such as 22 mg / L and 23 mg / L, with weights of 0.6 and 0.4 respectively, to fill the missing point with a concentration of 22.4 mg / L. This method ensures the spatiotemporal continuity of the dataset and is suitable for dynamic pollution analysis.

[0147] Step S340: For a spatiotemporally continuous dataset, obtain the interpolation accuracy index, and adjust the interpolation parameters using an iterative interpolation method. If the interpolation accuracy is lower than a preset threshold, repeat the iterative adjustment to obtain a preprocessed dataset with high-density spatiotemporal coverage.

[0148] Understandably, when obtaining interpolation accuracy metrics for spatiotemporally continuous datasets, the root mean square error (RMSE) can be calculated. For example, if the interpolated concentration deviates from the actual concentration by 0.3 mg / L, which is lower than the preset threshold of 0.5 mg / L, it indicates high accuracy. If the accuracy is insufficient, iterative interpolation methods are used to adjust the parameters, such as increasing the search radius of the Kriging interpolation from 100 m to 150 m, and repeating the calculation until the threshold is met.

[0149] Preferably, this iterative approach improves data reliability, and the generated preprocessed dataset covers the entire spatiotemporal domain, making it suitable for dynamic pollution monitoring.

[0150] Furthermore, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S400 includes:

[0151] Step S410: Obtain sensor values, image textures, and meteorological parameters from the preprocessed dataset. Use principal component analysis to standardize the data dimensions, calculate the covariance matrix, and obtain eigenvalues ​​and eigenvectors. If the cumulative contribution rate of the eigenvalues ​​is greater than a preset threshold, select the corresponding eigenvector to obtain the dimensionality-reduced key feature set.

[0152] For example, when obtaining sensor values, image textures, and meteorological parameters from a preprocessed dataset, relevant fields can be extracted through database queries. For instance, the sensor value record might be a pollution concentration of 10 mg / L, the image texture might be a water surface reflectance of 0.8, and the meteorological parameter might be a wind speed of 2 m / s.

[0153] It should be noted that the data must be in a consistent format, such as standardizing all values ​​to the same unit, to avoid bias in subsequent analysis.

[0154] One possible implementation involves using principal component analysis (PCA) to reduce the dimensionality of the data. PCA extracts the principal component directions of the data by calculating the covariance matrix. For example, the covariance matrices of sensor values, image texture, and meteorological parameters show a high correlation between pollution concentration and wind speed, generating eigenvalues ​​of 3.2, 1.5, and 0.8, respectively. If a preset cumulative contribution rate threshold of 85% is set, the eigenvectors corresponding to the first two eigenvalues ​​are selected, retaining the key feature set. This method effectively reduces data dimensionality while preserving key information.

[0155] Step S420: Based on the key feature set, obtain the variance contribution rate of each feature, calculate the weight using the entropy method, and if the weight is lower than the preset threshold, reset the weight to zero. Integrate the weight and features through matrix multiplication to obtain the weighted feature matrix.

[0156] Specifically, the variance contribution rate of each feature is calculated based on the key feature set, and the weights are determined using the entropy method. For example, the variance contribution rate of pollution concentration is 0.6, image texture is 0.3, and wind speed is 0.1. If the preset weight threshold is 0.2, the wind speed weight is set to zero because it is below the threshold. Finally, the weights of pollution concentration and image texture (0.6 and 0.3, respectively) are multiplied by the feature values ​​using matrix multiplication to generate a weighted feature matrix. This weight allocation highlights the main features and improves the targeting of subsequent analyses.

[0157] Step S430: Obtain feature subsets of sensor values, image textures, and meteorological parameters from the weighted feature matrix, and stitch them together using a data fusion method. If the dimension after stitching exceeds a preset threshold, adjust it using a linear dimensionality reduction method to obtain a preliminary fused feature set.

[0158] In one embodiment, a subset of features is extracted from the weighted feature matrix and concatenated using a data fusion method. For example, pollution concentration and image texture are concatenated into a feature vector with a dimension of 10. If a preset dimension threshold of 8 is set, the feature vector is further compressed using a linear dimensionality reduction method, such as principal component analysis, to obtain a preliminary fused feature set. This concatenation and fusion enhances the integrity of the features.

[0159] Step S440: Based on the preliminary fused feature set, obtain the correlation coefficient matrix between features, use clustering method to group features with correlation coefficients higher than a preset threshold, and integrate them through the weighted average method within the group to obtain the final fused feature set.

[0160] Preferably, a correlation coefficient matrix is ​​calculated based on the preliminary fused feature set to analyze the correlation between features. For example, the correlation coefficient between pollution concentration and image texture is 0.7, which is higher than the preset threshold of 0.6, so clustering is required.

[0161] In one embodiment, K-means clustering is used to group features with high correlation coefficients together, and then the features within each group are integrated using a weighted average method. For example, pollution concentration and image texture have weights of 0.5 and 0.5 respectively, resulting in a final fused feature value of 0.75. This clustering integration reduces redundant features and improves the representativeness of the feature set.

[0162] Understandably, correlation coefficient analysis and clustering ensure the independence of features and avoid information overlap. For example, the correlation coefficient between wind speed and pollution concentration is 0.3, which is below the threshold and is retained as an independent feature. This method generates efficient feature sets through multi-dimensional analysis, making it suitable for dynamic analysis in pollution monitoring scenarios.

[0163] Furthermore, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S500 includes:

[0164] Step S510: Obtain feature dimension data from the fused feature set, calculate the distribution characteristics of the feature dimension data using statistical methods, and if the feature dimension exceeds a preset threshold, sort the feature dimension data by importance using the random forest algorithm to obtain the feature subset after dimensionality reduction.

[0165] When obtaining feature dimension data from the fused feature set, feature vectors can be extracted through database queries. For example, the fused feature set contains pollution concentration, image texture, and humidity, with a dimension of 15.

[0166] It should be noted that the feature dimension data must be formatted consistently, such as normalizing all values ​​to the range of 0 to 1, to avoid distribution bias.

[0167] In one possible implementation, statistical methods are used to calculate the distribution characteristics of the feature dimensions. Specifically, the mean and standard deviation of each feature are calculated. For example, the mean of pollution concentration is 0.5 and the standard deviation is 0.2; the mean of image texture is 0.6 and the standard deviation is 0.15. This distribution characteristic analysis helps to understand the range of data fluctuations.

[0168] If the feature dimension exceeds a preset threshold, such as 10, then the importance is ranked using a random forest algorithm. For example, the random forest calculates the Gini index for each feature, with pollution concentration scoring 0.4, image texture 0.3, and humidity 0.1. The top 10 features by score are selected to obtain the dimensionality-reduced feature subset.

[0169] Step S520: Based on the feature subset after dimensionality reduction, obtain the training set and validation set, and use the cross-validation method to optimize the tree depth and number of trees of the random forest to obtain the optimized model parameter set.

[0170] Based on the dimensionality-reduced feature subset, obtain the training and validation sets. For example, the feature subset contains 8 dimensions, and the dataset is divided into 80% training set and 20% validation set.

[0171] Preferably, the tree depth and number of trees in the random forest are optimized using a 5-fold cross-validation method.

[0172] In one embodiment, the test tree depth ranged from 5 to 15, and the number of trees ranged from 50 to 200, ultimately resulting in an optimized tree depth of 10 and a number of trees of 100. This parameter optimization ensures the model's generalization ability.

[0173] Step S530: Obtain classification labels from the optimized model parameter set, use the random forest algorithm to perform classification training on the training set, evaluate the classification performance through the validation set, and obtain the predicted probability distribution of the classification labels.

[0174] Based on the optimized model parameter set, classification labels are obtained. For example, the classification labels are high, medium, and low pollution levels.

[0175] The random forest algorithm is used to train the classification on the training set, and the classification performance is evaluated using the validation set to obtain the predicted probability distribution of the classification labels. For example, the probability of a sample being predicted as highly contaminated is 0.7, medium contaminated is 0.2, and low contaminated is 0.1.

[0176] Step S540: Based on the predicted probability distribution of the classification labels, obtain the mapping function of pollution concentration, and use the linear regression method to fit the predicted probability distribution to obtain the predicted result of pollution concentration.

[0177] Understandably, the predicted probability distribution reflects the model's confidence in the pollution level.

[0178] Based on this distribution, a mapping function for pollution concentration is obtained. For example, the probability of high pollution is mapped to a concentration range of 8-10 mg / L, and that of medium pollution is 4-8 mg / L.

[0179] Linear regression was used to fit the predicted probability distribution to the actual concentration. For example, the linear regression result showed that a probability of 0.7 corresponded to a concentration of 9 mg / L. This fitting method transforms the classification results into continuous concentration predictions, facilitating dynamic analysis in pollution monitoring.

[0180] It should be noted that the mapping function needs to be updated periodically to adapt to changes in data distribution.

[0181] In one embodiment, if humidity features contribute little to concentration prediction, the model can be further simplified through feature selection to improve prediction efficiency.

[0182] This method generates efficient prediction results through multi-dimensional analysis, making it suitable for real-time applications in pollution monitoring scenarios.

[0183] Furthermore, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S600 includes:

[0184] Step S610: Obtain concentration distribution data from the pollution concentration prediction results, and use a convolutional neural network to extract spatial patterns from the concentration distribution data to obtain a spatial pattern feature set.

[0185] For example, when obtaining concentration distribution data from pollution concentration prediction results, the monitoring area can be divided into multiple sub-regions using a gridding method, with each sub-region corresponding to a concentration value.

[0186] Assuming the monitoring area is a lake covering 100 square kilometers, it is divided into 1000 grids of 1 square kilometer each, with each grid recording the pollution concentration. For example, if the concentration in one grid is 5 mg / L, the concentration in a neighboring grid is 4.8 mg / L. This gridding method facilitates the capture of spatial variations in concentration distribution, providing a data foundation for subsequent analysis.

[0187] In one possible implementation, a convolutional neural network is used to extract spatial pattern features of the concentration distribution. The convolutional neural network identifies local correlations in the concentration distribution through multiple convolutional operations.

[0188] Assuming the input is a 100×100 concentration grid, the network contains three convolutional layers, each using a 3×3 kernel. The first layer extracts edge features, such as concentration gradient changes; the second layer captures regional patterns, such as high-concentration clusters; and the third layer integrates global features to form a spatial pattern feature set. For example, the output feature set includes the geometry and diffusion trends of high-concentration regions.

[0189] Step S620: Based on the spatial pattern feature set and combined with the water flow field data, the pollution migration path is calculated using a fluid dynamics algorithm to obtain the pollution migration path set.

[0190] It should be noted that when calculating the pollution migration path by combining water flow field data, fluid dynamics algorithms can simulate the trajectory of pollutants moving with the water flow.

[0191] Flow field data includes velocity and direction. For example, the velocity in a certain area is 0.2 m / s and the direction is northeast.

[0192] Based on this, the algorithm calculates the migration path of pollutants from high-concentration grids to low-concentration grids. For example, the migration path of pollutants from a grid with a concentration of 5 mg / L to a grid with a concentration of 3 mg / L is 2 kilometers long and takes approximately 10,000 seconds. This path simulation reflects the dynamic propagation pattern of pollutants.

[0193] Step S630: Determine whether the number of paths in the pollution migration path set exceeds a preset threshold. If the number of paths in the pollution migration path set exceeds the preset threshold, then use an iterative optimization algorithm to filter and cluster the path set to determine the candidate set of pollution source locations.

[0194] Specifically, if the number of paths in the pollution migration path set exceeds a preset threshold, such as 100, then the paths are filtered and clustered through an iterative optimization algorithm.

[0195] Iterative optimization algorithms can sort paths based on their length and concentration gradient, retaining paths with high confidence. For example, after selecting 50 paths, K-means clustering can be used to divide the paths into three classes, each representing a possible pollution source area.

[0196] Clustering results show that a certain type of path is concentrated in the northwest corner of the lake, indicating the location of possible pollution sources.

[0197] Step S640: For the candidate set of pollution source locations, use Bayesian inference to calculate the probability distribution of each candidate location to obtain the source tracing results of the pollution source locations.

[0198] In one embodiment, for a candidate set of pollution source locations, a Bayesian inference method calculates the probability of each candidate location. Assume the candidate set contains three locations: the northwest corner, the central area, and the southeast corner. Bayesian inference combines path density and flow field data to calculate the probability that each location is a pollution source. For example, the probability is 0.6 for the northwest corner, 0.3 for the central area, and 0.1 for the southeast corner. The northwest corner, with the highest probability, is determined as the source. This method quantifies uncertainty through probability distribution, improving the reliability of source tracing.

[0199] Preferably, the above method, through multi-dimensional analysis, forms a complete pollution monitoring chain, from concentration distribution to source tracing results. For example, spatial pattern feature sets reveal the distribution patterns of high-concentration areas, migration path sets reflect the dynamics of pollutant propagation, and source tracing results accurately locate pollution sources. This information can provide data support for pollution control, such as prioritizing the treatment of pollution sources in the northwest corner.

[0200] Understandably, the probability distribution of Bayesian inference can be updated periodically to adapt to changes in flow field or concentration data. For example, if lake flow velocity adjusts due to seasonal variations, the path and probability need to be recalculated. This dynamic updating ensures the timeliness of source tracing results, making it suitable for real-time monitoring scenarios.

[0201] Furthermore, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S700 includes:

[0202] Step S710: Obtain concentration distribution data from the pollution concentration prediction results, use a Bayesian network to perform probabilistic modeling on the concentration distribution data, and combine the flow field data to calculate the dynamic probability distribution of pollution migration paths to obtain a set of pollution migration paths.

[0203] For example, when obtaining concentration distribution data from pollution concentration prediction results, discrete concentration values ​​of the monitoring area can be transformed into a continuous concentration distribution map through spatial interpolation methods.

[0204] Assuming the monitoring area is a river basin of 50 square kilometers, the monitoring station provides concentration data from 100 points. For example, the concentration at a certain point is 3.2 mg / L, and the concentration at a neighboring point is 3.0 mg / L.

[0205] Spatial interpolation methods generate a 1000×1000 concentration grid by weighted averaging, capturing the continuity of concentration changes and providing a data foundation for Bayesian network modeling.

[0206] In one possible implementation, Bayesian networks probabilistically model concentration distribution data to identify conditional dependencies in the concentration distribution.

[0207] Bayesian networks divide a concentration grid into nodes, each representing a concentration value in a sub-region, with edges representing probabilistic dependencies between nodes. For example, the concentration of a certain grid is influenced by neighboring grids and the flow field. By learning from historical data, the network outputs the probability distribution of the concentration for each grid, such as a probability of 0.7 for a concentration of 3.5 mg / L in a certain grid. This modeling approach reflects the uncertainty of the concentration distribution.

[0208] It should be noted that when calculating the dynamic probability distribution of pollution migration paths using flow field data, a particle tracking algorithm can be used to simulate the propagation of pollutants with the water flow. Flow field data includes flow velocity and flow direction; for example, the flow velocity in a certain area is 0.3 m / s, and the flow direction is southeast.

[0209] The algorithm generates multiple possible migration paths and calculates the probability of each path. For example, the path probability for a pollutant migrating from a grid with a concentration of 3.5 mg / L to a grid with a concentration of 2.8 mg / L is 0.6. This dynamic probability distribution reflects the randomness of pollutant propagation.

[0210] Step S720: If the number of paths in the pollution migration path set exceeds a preset threshold, the pollution migration path set is filtered by a clustering algorithm, similar migration paths are merged, and a candidate location set is determined.

[0211] Specifically, if the number of paths in the pollution migration path set exceeds a preset threshold, such as 200, then similar paths are filtered and merged using the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm.

[0212] DBSCAN clusters paths based on their start, end, and length, merging spatially adjacent paths. For example, after clustering, 100 paths are retained, generating three candidate location sets located in the upper, middle, and lower reaches of the river, respectively. This filtering method reduces redundant paths and improves the efficiency of subsequent analysis.

[0213] Step S730: For the candidate location set, the confidence level of each candidate location is calculated using the Bayesian inference method. The confidence level distribution is generated by combining the source tracing results to obtain the probability distribution of the pollution source location.

[0214] In one embodiment, for a set of candidate locations, a Bayesian inference method calculates the confidence level for each candidate location.

[0215] Assuming candidate locations are upstream, midstream, and downstream, Bayesian inference combines path probabilities and flow field data to calculate the confidence distribution. For example, the confidence level for the upstream location is 0.65, for the midstream it is 0.25, and for the downstream it is 0.1. The confidence distribution quantifies the probability that each location is a pollution source.

[0216] Step S740: Based on the probability distribution, a dynamic signal generation algorithm is used to fuse the probability distribution and concentration prediction results to generate a dynamic signal for real-time early warning.

[0217] Preferably, a dynamic signal generation algorithm is used to fuse probability distribution and concentration prediction results to generate a real-time early warning dynamic signal. The algorithm combines the confidence distribution with real-time concentration data through weighted fusion to generate an early warning signal. For example, when the confidence level at the upstream location is 0.65 and the concentration rises abnormally, the algorithm generates a high-priority early warning signal, indicating that the pollution source may be located upstream. This dynamic signal supports real-time monitoring and rapid response.

[0218] Understandably, dynamic signal generation algorithms can periodically update warning signals based on changes in the flow field. For example, if river flow velocity increases due to rainfall, the algorithm recalculates the path probability and confidence level, updating the warning signal accordingly. This dynamic update mechanism ensures the timeliness of warnings and is suitable for complex water monitoring scenarios.

[0219] Furthermore, in the water quality monitoring method based on artificial intelligence proposed in this embodiment, step S800 includes:

[0220] Step S810: Obtain signal feature data from dynamic early warning signals, sort the signal feature data by priority using a classification algorithm, generate a priority queue by combining the preset push rules, and obtain a push priority sequence.

[0221] For example, when obtaining signal feature data from dynamic early warning signals, key attributes of the signal can be identified through feature extraction methods. For instance, in a river pollution monitoring scenario, dynamic early warning signals may include features such as the magnitude of concentration anomalies, the time of occurrence, and geographical location.

[0222] Suppose a signal indicates a sudden increase in concentration from 2.5 mg / L to 4.0 mg / L in a certain area over a period of 2 hours, located in the middle reaches of a river. Feature extraction methods quantify these attributes into feature vectors. This feature extraction approach provides standardized input for subsequent classification, ensuring the comprehensiveness of the signal features.

[0223] In one possible implementation, a decision tree algorithm can be used when prioritizing signal feature data using a classification algorithm. The decision tree determines the urgency of the signal based on its feature vectors. For example, a signal with a concentration change greater than 1.0 mg / L and a duration exceeding 1 hour is marked as high priority.

[0224] Suppose a signal's feature vector is given. A decision tree, based on rules, classifies it as high priority, while another signal is marked as low priority. Combined with pre-defined push rules—such as high-priority signals requiring immediate push and low-priority signals being pushed hourly—a priority queue is generated. This method ensures that urgent signals are processed first.

[0225] Step S820: Based on the push priority sequence, the dynamic early warning signal is converted into standardized early warning data using a data encapsulation method, and a formatted data packet is generated in combination with the preset data format specifications to determine the content of the transmitted data.

[0226] It should be noted that after generating the push priority sequence, when converting dynamic early warning signals into standardized early warning data using data encapsulation methods, JSON format can be used for encapsulation. For example, high-priority signals are encapsulated into JSON objects containing fields such as timestamp, location, and concentration change, such as {"timestamp":"2025-04-30 10:00", "location":"midstream", "concentration_change":1.5}. Combined with preset data format specifications, such as a field length limit of 32 bytes, a formatted data packet is generated. This encapsulation method facilitates system parsing and transmission.

[0227] Step S830: For the formatted data packet, the transmission protocol is used to fragment the formatted data packet. During fragmentation, a checksum is generated by combining the checksum algorithm to obtain a transmission data unit with checksum.

[0228] Specifically, for the fragmentation of formatted data packets, a TCP protocol combined with a fragmentation algorithm can be used. If the data packet size is 2KB, exceeding the 1KB limit for single transmission, it is divided into two 1KB fragments. Each fragment is appended with a checksum, such as a 32-bit checksum generated using the CRC32 algorithm. For example, a fragment checksum might be 0xA1B2C3D4, ensuring error detection during transmission. This method improves transmission reliability.

[0229] Step S840: Send the transmission data unit to the decision support system through the data transmission channel, and use an asynchronous processing method to reassemble and verify the received transmission data unit, determine the data integrity and generate an acknowledgment signal.

[0230] In one embodiment, the MQTT protocol can be used when transmitting data units to the decision support system via a data transmission channel. MQTT supports a lightweight publish / subscribe model, suitable for real-time monitoring. For example, data units can be published with the topic "pollution_alert," and the decision support system can subscribe to this topic to receive data.

[0231] When reassembling received data units, the asynchronous processing method first sorts them by fragment number and then verifies their integrity using a checksum. If the verification fails, the system requests a retransmission. This asynchronous processing method reduces blocking and improves efficiency.

[0232] Preferably, the integrity of the reconstructed data is determined by comparing the hash values ​​of the received data packets. For example, the sending end generates a data packet hash value of 0xF1E2D3C4, the receiving end calculates and compares it, and if they match, an acknowledgment signal is generated. This mechanism ensures lossless data transmission and provides a reliable basis for decision support systems.

[0233] Understandably, once a confirmation signal is generated, the decision support system can trigger a response mechanism based on the early warning data. For example, a high-priority signal triggers an immediate investigation of the pollution source, while a low-priority signal is logged. This tiered response mechanism optimizes resource allocation and improves monitoring efficiency.

[0234] This invention relates to an artificial intelligence-based water quality monitoring system for implementing the aforementioned artificial intelligence-based water quality monitoring method. The artificial intelligence-based water quality monitoring system includes a first acquisition module, an adjustment module, a second acquisition module, a third acquisition module, a fourth acquisition module, a fifth acquisition module, a generation module, and a transmission module. The first acquisition module acquires a comprehensive dataset composed of sensor data, satellite images, and meteorological parameters. The adjustment module adjusts the sampling frequency and location based on the water flow velocity and pollution concentration gradient in the comprehensive dataset using a dynamic sampling algorithm, and outputs adjusted data. The second acquisition module performs spatiotemporal interpolation processing on the adjusted data to obtain a preprocessed dataset with high-density spatiotemporal coverage. The third acquisition module extracts key features of sensor values, image textures, and meteorological parameters from the preprocessed dataset using principal component analysis, and constructs a weighted feature matrix. The system first obtains a fused feature set. The fourth acquisition module determines whether the dimensionality of the fused feature set exceeds a preset threshold. If the dimensionality exceeds the preset threshold, 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. The fifth acquisition module uses a convolutional neural network to analyze the spatial pattern of the pollution concentration prediction result, calculates the pollution migration path based on water flow field data, and determines the pollution source location through iterative optimization to obtain the pollution source tracing result. The generation module determines whether the confidence level of the pollution source tracing result is higher than a preset threshold. If the confidence level is higher than the preset threshold, 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. The transmission module inputs the dynamic early warning signal into a push algorithm to transmit the early warning data to the decision support system.

[0235] The artificial intelligence-based water quality monitoring method and system provided in this embodiment, compared with existing technologies, acquires multi-source data, optimizes data acquisition using a dynamic sampling algorithm, performs spatiotemporal interpolation and feature extraction on the data, and constructs a fused feature set. A random forest algorithm is used to reduce the dimensionality of the features for classification and predict pollution concentration. A convolutional neural network is combined to analyze pollution spatial patterns, calculate pollution migration paths, and determine the location of pollution sources. Finally, a real-time early warning model is built based on a Bayesian network to generate dynamic early warning signals and push them to a decision support system. This embodiment achieves intelligent monitoring, precise source tracing, and timely early warning of water pollution, providing effective technical support for water environment protection and possessing significant ecological and social benefits.

[0236] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A water quality monitoring method based on artificial intelligence, characterized in that, Includes the following steps: Acquire a comprehensive dataset consisting of sensor data, satellite imagery, and meteorological parameters; Based on the pollution concentration gradient in the comprehensive dataset, a dynamic sampling algorithm is used to adjust the sampling frequency and location, and the adjusted data is output. The adjusted data is subjected to spatiotemporal interpolation to obtain a preprocessed dataset with high-density spatiotemporal coverage; Principal component analysis was used to extract key features of sensor values, image textures and meteorological parameters from the preprocessed dataset, and a weighted feature matrix was constructed to obtain a fused feature set. Determine whether the dimension of the fused feature set exceeds a preset threshold. If the dimension of the fused feature set exceeds the preset threshold, use the random forest algorithm to reduce the dimensionality and classify the fused features, optimize the model parameters through cross-validation, and obtain the pollution concentration prediction result. The spatial pattern of the pollution concentration prediction results is analyzed using a convolutional neural network, and the pollution migration path is calculated by combining water flow field data. The location of the pollution source is determined through iterative optimization, and the pollution source tracing results are obtained. The spatial pattern refers to the local or global feature structure with a specific spatial distribution pattern in the pollution concentration prediction data, which is manifested as an ordered arrangement relationship between pixels, voxels or geometric elements. Determine whether the confidence level of the pollution source tracing result is higher than a preset threshold. If the confidence level of the pollution source tracing result is higher than the preset threshold, then a real-time early warning model is constructed using a Bayesian network, and the pollution concentration prediction result and the pollution source tracing result are fused to generate a dynamic early warning signal. The dynamic early warning signal is input into the push algorithm to transmit the early warning data to the decision support system.

2. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, The steps for acquiring the comprehensive dataset consisting of sensor data, satellite images, and meteorological parameters include: Collect and preprocess multi-source data: Obtain environmental parameters, satellite images and meteorological parameters through sensor devices, satellite remote sensing platforms and meteorological station databases, and use format conversion tools to standardize the multi-source data to obtain a standardized dataset; Time and space alignment is performed: the standardized dataset is time-synchronized based on timestamps. If timestamps are missing, they are filled in by linear interpolation. The standardized dataset is spatially aligned using a geographic coordinate matching method to obtain a time-space synchronized dataset. Handling missing and anomalies: For the aforementioned time-space synchronized dataset, if sensor data is missing, it is filled by interpolation using the K-nearest neighbor algorithm. If meteorological parameters exceed a preset threshold, they are marked as anomalies and removed to obtain a cleaned dataset. Data fusion and correlation calculation: The cleaned datasets are fused using a weighted average method, and the correlation of the fused data is calculated using the Pearson correlation coefficient to obtain a comprehensive dataset.

3. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, Based on the pollution concentration gradient in the comprehensive dataset, the steps of adjusting the sampling frequency and location using a dynamic sampling algorithm and outputting the adjusted data include: Pollution concentration data is obtained from the comprehensive dataset, and the pollution concentration gradient is calculated using a two-dimensional gradient operator to obtain a concentration gradient distribution dataset; wherein the gradient operator is defined as follows: C represents the pollution concentration, and x and y represent spatial coordinates; Based on the concentration gradient distribution dataset, if the concentration gradient value is greater than a preset threshold, a grid division method is used to divide the sampling locations in the high gradient region, and a weighted average method is used to calculate the sampling frequency to obtain a preliminary adjusted sampling scheme. The weighted average method assigns higher weights to the high gradient grid to obtain the sampling frequency of each grid, so as to ensure that the sampling points cover areas with drastic pollution changes. For the initially adjusted sampling scheme, the K-means clustering algorithm is used to group and optimize the sampling locations, and the sampling frequency is synchronized by timestamps to obtain the optimized sampling scheme. The adjusted sampling frequency and location data are obtained from the optimized sampling scheme. The adjusted sampling frequency and location data are integrated using a data fusion method to output the adjusted data.

4. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, The steps of performing spatiotemporal interpolation on the adjusted data to obtain a preprocessed dataset with high-density spatiotemporal coverage include: The sampling frequency, sampling location, and timestamp are obtained from the adjusted data. The timestamps are sorted using a time series analysis method. If the timestamp interval is greater than a preset threshold, the time series gaps are filled by a linear interpolation method to obtain the time-synchronized sampling data. Based on the time-synchronized sampling data, the pollution concentration and water flow velocity are obtained. The sampling location is mapped to a two-dimensional coordinate system using a spatial grid division method. If the pollution concentration gradient within the grid is greater than a preset threshold, spatial interpolation is performed using the Kriging interpolation method to obtain high-density spatial distribution data. Data gaps are identified in the high-density spatial distribution data. Pollution concentration and water flow velocity are integrated using a data fusion method. If the gaps are located at the spatial grid boundary, they are filled using a weighted average method to obtain a spatiotemporally continuous dataset. For spatiotemporally continuous datasets, interpolation accuracy metrics are obtained, and interpolation parameters are adjusted using an iterative interpolation method. If the interpolation accuracy is lower than a preset threshold, the iterative adjustment is repeated to obtain a preprocessed dataset with high-density spatiotemporal coverage.

5. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, The steps of extracting key features of sensor values, image textures, and meteorological parameters from the preprocessed dataset using principal component analysis, constructing a weighted feature matrix, and obtaining a fused feature set include: Sensor values, image textures, and meteorological parameters are obtained from the preprocessed dataset. Principal component analysis is used to standardize the data dimensions, calculate the covariance matrix, and obtain eigenvalues ​​and eigenvectors. If the cumulative contribution rate of the eigenvalues ​​is greater than a preset threshold, the corresponding eigenvector is selected to obtain the dimensionality-reduced key feature set. Based on the key feature set, the variance contribution rate of each feature is obtained, and the weight is calculated using the entropy method. If the weight is lower than a preset threshold, the weight is reset to zero. The weight and feature are integrated by matrix multiplication to obtain a weighted feature matrix. The feature subsets of sensor values, image textures and meteorological parameters are obtained from the weighted feature matrix and stitched together using a data fusion method. If the dimension after stitching exceeds a preset threshold, it is adjusted by a linear dimensionality reduction method to obtain a preliminary fused feature set. Based on the preliminary fused feature set, the correlation coefficient matrix between features is obtained. Features with correlation coefficients higher than a preset threshold are grouped using a clustering method and then integrated using a weighted average method within the group to obtain the final fused feature set.

6. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, The steps of determining whether the dimension of the fused feature set exceeds a preset threshold, and if the dimension of the fused feature set exceeds the preset threshold, then using a random forest algorithm to reduce the dimensionality and classify the fused features, and optimizing the model parameters through cross-validation to obtain the pollution concentration prediction result include: Feature dimension data is obtained from the fused feature set, and the distribution characteristics of the feature dimension data are calculated using statistical methods. If the feature dimension exceeds a preset threshold, the feature dimension data is sorted by importance using the random forest algorithm to obtain a feature subset after dimensionality reduction. Based on the feature subset after dimensionality reduction, the training set and validation set are obtained, and the tree depth and number of trees of the random forest are optimized by cross-validation to obtain the optimized model parameter set. Classification labels are obtained from the optimized model parameter set, and the training set is trained using the random forest algorithm. The classification performance is evaluated using the validation set to obtain the predicted probability distribution of the classification labels. Based on the predicted probability distribution of the classification labels, the mapping function of pollution concentration is obtained, and the predicted probability distribution is fitted using a linear regression method to obtain the predicted pollution concentration.

7. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, The steps of analyzing the spatial pattern of the pollution concentration prediction results using a convolutional neural network, calculating the pollution migration path by combining water flow field data, determining the pollution source location through iterative optimization, and obtaining the pollution source tracing results include: Concentration distribution data is obtained from the pollution concentration prediction results, and spatial pattern extraction is performed on the concentration distribution data using a convolutional neural network to obtain a spatial pattern feature set; Based on the spatial pattern feature set and combined with water flow field data, a fluid dynamics algorithm is used to calculate the pollution migration path and obtain the pollution migration path set. Determine whether the number of paths in the pollution migration path set exceeds a preset threshold. If the number of paths in the pollution migration path set exceeds the preset threshold, then use an iterative optimization algorithm to filter and cluster the path set to determine a candidate set of pollution source locations. For the candidate set of pollution source locations, the probability distribution of each candidate location is calculated using Bayesian inference to obtain the source tracing results.

8. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, The steps of determining whether the confidence level of the pollution source tracing result is higher than a preset threshold, and if the confidence level of the pollution source tracing result is higher than the preset threshold, then using a Bayesian network to construct a real-time early warning model, and fusing the pollution concentration prediction result with the pollution source tracing result to generate a dynamic early warning signal, include: Concentration distribution data is obtained from the pollution concentration prediction results. Bayesian network is used to perform probabilistic modeling on the concentration distribution data. Combined with flow field data, the dynamic probability distribution of pollution migration paths is calculated to obtain a set of pollution migration paths. If the number of paths in the pollution migration path set exceeds a preset threshold, the pollution migration path set is filtered by a clustering algorithm to merge similar migration paths and determine a candidate location set. For the candidate location set, the confidence level of each candidate location is calculated using the Bayesian inference method. The confidence level distribution is then generated by combining the source tracing results to obtain the probability distribution of the pollution source location. Based on the probability distribution, a dynamic signal generation algorithm is used to fuse the probability distribution with the concentration prediction results to generate a dynamic signal for real-time early warning.

9. The water quality monitoring method based on artificial intelligence as described in claim 1, characterized in that, The steps of inputting the dynamic early warning signal into the push algorithm and transmitting the early warning data to the decision support system include: Signal feature data is obtained from dynamic early warning signals, and the signal feature data is prioritized using a classification algorithm. A priority queue is generated by combining the data with preset push rules to obtain a push priority sequence. According to the push priority sequence, the dynamic early warning signal is converted into standardized early warning data using a data encapsulation method, and a formatted data packet is generated in combination with the preset data format specification to determine the content of the transmitted data. For the formatted data packet, a transmission protocol is used to fragment the formatted data packet. During fragmentation, a checksum is generated by combining a checksum algorithm to obtain a transmission data unit with checksum. The data transmission unit is sent to the decision support system through the data transmission channel. The received data transmission unit is reassembled and verified using an asynchronous processing method to determine data integrity and generate an acknowledgment signal.

10. An artificial intelligence-based water quality monitoring system, used to implement the artificial intelligence-based water quality monitoring method as described in any one of claims 1 to 9, characterized in that, The artificial intelligence-based water quality monitoring system includes: The first acquisition module is used to acquire a comprehensive dataset consisting of sensor data, satellite images, and meteorological parameters; The adjustment module is used to adjust the sampling frequency and location based on the pollution concentration gradient in the comprehensive dataset using a dynamic sampling algorithm, and output the adjusted data. The second acquisition module is used to perform spatiotemporal interpolation processing on the adjusted data to obtain a preprocessed dataset with high-density spatiotemporal coverage. The third acquisition module is used to extract key features of sensor values, image textures and meteorological parameters from the preprocessed dataset using principal component analysis, construct a weighted feature matrix, and obtain a fused feature set. The fourth acquisition module is used to determine whether the dimension of the fused feature set exceeds a preset threshold. If the dimension of the fused feature set exceeds the preset threshold, the random forest algorithm is used to reduce the dimensionality and classify the fused features, and the model parameters are optimized through cross-validation to obtain the pollution concentration prediction result. The fifth acquisition module is used to analyze the spatial pattern of the pollution concentration prediction results using a convolutional neural network, calculate the pollution migration path by combining water flow field data, determine the location of the pollution source through iterative optimization, and obtain the pollution source tracing results. The generation module is used to determine whether the confidence level of the pollution source tracing result is higher than a preset threshold. If the confidence level of the pollution source tracing result is higher than the preset threshold, a real-time early warning model is constructed using a Bayesian network, and the pollution concentration prediction result and the pollution source tracing result are fused to generate a dynamic early warning signal. The transmission module is used to input the dynamic early warning signal into the push algorithm and transmit the early warning data to the decision support system.

Images