Complex water area water pollution monitoring and early warning system based on image recognition

Abstract

The invention discloses a complex water area water pollution monitoring and early warning system based on image recognition, which relates to the field of water pollution monitoring and comprises an image acquisition module, an image preprocessing module, a data fusion module, a water pollution detection module and an early warning response module. A multispectral camera and an unmanned aerial vehicle are combined to carry out all-dimensional and multi-dimensional image acquisition of a complex water area; a multi-task visual representation learning framework is adopted to carry out processing and feature extraction on images, and image data from different devices are integrated through data fusion, so that the comprehensiveness and reliability of water pollution detection are improved; the early warning response module discovers the water pollution problem in time and quickly responds and treats the water pollution problem; the invention solves the problem that the existing complex water area water pollution monitoring and early warning technology cannot comprehensively acquire the water area pollution information in real time, and can efficiently, accurately and comprehensively identify and monitor the water area pollution condition and timely send out the early warning information.

Description

Technical Field

[0001] The present invention relates to the field of water pollution monitoring, and particularly to a water pollution monitoring and early warning system for complex waters based on image recognition. Background Art

[0002] With the accelerating advancement of industrialization and urbanization, the water ecological environment is facing increasingly severe challenges. The poor water ecological environment will affect the distribution of water resources. Therefore, water pollution prevention and control has become an important task in current water resource management and environmental protection. In terms of water pollution prevention and control, monitoring and early warning are very crucial links. Especially for the water pollution monitoring and early warning of complex waters such as rivers, lakes, and coastal waters, due to the influence of various factors such as water flow, terrain, and climate, the water quality conditions often show a high degree of spatio-temporal heterogeneity, and it is difficult to monitor and give early warnings. Therefore, the monitoring and early warning of water pollution in complex waters have become a research hotspot.

[0003] The existing water pollution monitoring and early warning technologies for complex waters adopt the method of sampling and monitoring local areas or specific points, which can only cover local areas. However, the water surface area of complex waters is large and has strong fluidity. The existing water pollution monitoring and early warning technologies for complex waters cannot obtain the pollution information of the waters in real time and comprehensively, thus affecting the timeliness and accuracy of early warnings.

[0004] Therefore, a water pollution monitoring and early warning system for complex waters based on image recognition is needed to solve the above problems. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention discloses a water pollution monitoring and early warning system for complex waters based on image recognition, which can monitor the pollution status of complex waters in real time and accurately, and give early warnings in a timely manner, helping to improve the efficiency and response speed of water environmental protection. Through the cooperation of four major modules: image acquisition, data fusion, pollution detection, and early warning response, it can comprehensively and accurately identify water pollution problems, providing reliable data support and decision-making basis for relevant decision-making departments.

[0006] The present invention adopts the following technical solutions:

[0007] A water pollution monitoring and early warning system for complex waters based on image recognition, comprising:

[0008] An image acquisition module, which performs real-time acquisition of fixed-point images by deploying multiple multi-spectral cameras at intervals around the waters to be monitored, and uses an unmanned aerial vehicle equipped with a high-definition camera to conduct mobile cruising of the waters to be monitored to fill the blind spots of fixed-point image acquisition, and transmits the acquired water area images to the cloud platform in real time for storage and processing;

[0009] An image preprocessing module, which performs different preprocessing operations on the collected images based on the environmental conditions of water area image collection;

[0010] A data fusion module, which extracts color, texture, shape and edge features from the preprocessed water area images through a multi-task visual representation learning framework, and uses a secondary data information fusion model to fuse the water area image data collected by multiple multi-spectral cameras and drones to form a feature distribution map of the water area to be monitored;

[0011] A water pollution detection module, which first adaptively identifies and samples the features of potential water pollution areas in the feature distribution map of the water area to be monitored, and quantitatively evaluates the water pollution degree of the water area to be monitored according to the sampling results;

[0012] An early warning response module, which conducts early warning response according to the quantitative results of the water pollution degree of the water area to be monitored. When the water pollution degree of the water area to be monitored exceeds the preset threshold, the early warning mechanism is automatically triggered.

[0013] Furthermore, the multiple multi-spectral cameras achieve data intercommunication and sharing through a wireless local area network. The drone equipped with a high-definition camera adjusts the direction and height of mobile cruising through a preset path planning algorithm and a remote controller. The image acquisition module marks the acquisition time, location and weather information of each water area image with a timestamp, and converts the collected analog image signal into a digital image signal through an analog-to-digital converter.

[0014] Furthermore, the image preprocessing module obtains the environmental conditions of water area image collection according to the marked timestamp, and performs different preprocessing operations on the water area images collected under different environmental conditions in a distributed edge computing manner. The environmental conditions at least include daytime lighting conditions, nighttime lighting conditions and rain, snow and fog meteorological conditions.

[0015] Furthermore, the multi-task visual representation learning framework encodes the input water area image into a low-dimensional feature vector representation through an autoencoder, and maps the low-dimensional feature vector to the original image space to realize the reconstruction of the water area image. The autoencoder uses a pre-trained convolutional neural network to automatically extract color, texture, shape and edge features from the input water area image, and uses a shared feature extraction layer and a task-specific layer for multi-task feature representation learning.

[0016] Furthermore, in the secondary data information fusion model, first, a BP neural network is used for the first fusion of local features and global features. The extracted local features and global features are combined into a high-dimensional feature vector and input into the BP neural network to represent the spatial and temporal dimension features of the water area image. The set of combined high-dimensional feature vectors is X = {x1,..., x g ,..., x n}, where g represents the ordinal number of the high-dimensional feature vector, 1 ≤ g ≤ n, and n represents the total number of high-dimensional feature vectors. The BP neural network adjusts the neuron weights through the hidden layer activation function to learn the relationship between local features and global features. The neuron weight coefficient iteration equation is expressed as:

[0017]

[0018] In formula (1), W ij (t + 1) represents the weight coefficient from the i-th input node to the j-th hidden node at the (t + 1)-th iteration, t + 1 represents the number of iterations, i represents the ordinal number of the input node, 1 ≤ i, j represents the ordinal number of the hidden node, 1 ≤ j, and W ij (t) represents the weight coefficient from the i-th input node to the j-th hidden node at the t-th iteration. v is a weighting factor used to adjust the speed of weight update, τ is the step size of neuron iteration, L is the loss function of the BP neural network, represents the partial derivative of the loss function L with respect to the weight coefficient W ij (t), θ represents the partial derivative. The BP neural network reduces the difference between the actual output and the target output by minimizing the loss function. Then, the loss function of the BP neural network is expressed as:

[0019]

[0020] In formula (2), L represents the loss function of the BP neural network, y g represents the given target output, represents the actual output of the BP neural network;

[0021] Then, an interpolation method is used to synchronize the time of the water area image features after the first fusion, and the water area image features after the first fusion are spatially registered by solving the transformation matrix;

[0022] Then, based on the D-S evidence model, the secondary fusion of the feature information after the first fusion at the same time, different spaces, and different modalities is performed. The D-S evidence model takes the image features after the first fusion as evidence elements and quantifies the relationship between the evidence elements through the belief function to obtain the final fusion result. The calculation formula is:

[0023]

[0024] In formula (3), m(A) is the evidence support degree after fusion, and m g (A) represents the evidence support degree of the g-th image.

[0025] Furthermore, the water pollution detection module uses an adaptive threshold segmentation method to identify the water quality characteristics of potential water pollution areas in the distribution map of water area characteristics to be monitored, and uses an automatic sampler to sample the water quality characteristics of potential water pollution areas. The adaptive threshold segmentation method performs adaptive threshold segmentation of water pollution areas based on pixel density and feature similarity; the water pollution detection module uses a multi-dimensional depth evaluation model to quantitatively analyze the sampled water quality characteristics to judge the concentration, type, and distribution characteristics of pollutants.

[0026] Furthermore, the multi-dimensional depth evaluation model includes an input layer, a data layer, an adaptive weight layer, a model layer, an optimization layer, and an output layer. The operation of the multi-dimensional depth evaluation model includes the following steps:

[0027] Step 1: Receive the water quality characteristics sampled by the automatic sampler through the input layer, and perform standardization processing on the input water quality characteristics;

[0028] Step 2: Obtain calculation parameters, calculation objectives, and constraint conditions through the data layer. The calculation parameters and constraint conditions include calculation scale, objective function, constraint conditions, and variable range. The calculation objectives include the concentration, type, and distribution characteristics of pollutants;

[0029] Step 3: The model layer extracts pollutant morphology, color distribution, and texture features from the standardized water quality characteristics through a deep convolutional neural network and a recurrent neural network, and establishes a weight relationship between the primary and secondary dimensions through the adaptive weight layer. The adaptive weight layer adaptively adjusts the weights of multi-dimensional feature vectors according to the calculation objectives;

[0030] Step 4: The model layer establishes a multi-dimensional evaluation mathematical model based on the extracted multi-dimensional features, calculation scale, objective function, constraint conditions, and variable range, and performs iterative calculation, parameter correction, and comparison between the calculation result and the true value on the multi-dimensional evaluation mathematical model through historical data. The model layer uses a parallel calculation method to allocate calculation tasks to multiple processors or calculation nodes, and obtains the neighbor list of calculation nodes according to the objective function and the distribution of calculation nodes;

[0031] Step 5: The optimization layer performs feedback training on the multi-dimensional evaluation mathematical model through the reconstruction error backpropagation algorithm, and sets the threshold and the number of iterations through the adaptive parameter selection method;

[0032] Step 6: Output the calculation result through the output layer.

[0033] Furthermore, the early warning mechanism automatically generates a pollution monitoring report through the GIS visualization platform and uploads it to the management platform. The management platform sends early warning information to the management personnel through SMS, email, or mobile application push.

[0034] The beneficial effects of the present invention are as follows:

[0035] 1. The present invention deploys multiple multispectral cameras around the water area to be monitored for fixed-point real-time acquisition, and at the same time uses drones for mobile cruising, effectively filling the blind spots of fixed-point monitoring, performing all-round and three-dimensional image acquisition, ensuring the integrity of data acquisition, and transmitting the acquired images to the cloud platform for storage and processing in real time, avoiding the problems of data loss and information lag in traditional monitoring methods, and also facilitating remote management and real-time monitoring, realizing comprehensive and efficient monitoring of complex water areas; the image preprocessing module can perform different preprocessing operations according to the environmental conditions (daytime lighting, nighttime lighting, and rain, snow, and fog) of water area image acquisition, reducing the impact of environmental factors on data quality, and improving the image quality and the speed and accuracy of subsequent processing.

[0036] 2. The present invention uses a multi-task visual representation learning framework to perform in-depth analysis on the preprocessed water area images, extracts multi-dimensional features such as color, texture, shape, and edges, which helps to improve the recognition accuracy of water pollution characteristics. Especially in the case of complex and changeable water area conditions, more detailed and comprehensive features can be extracted, and the image data collected from multispectral cameras and drones are comprehensively processed through a secondary data fusion model, and a more complete and accurate water area feature distribution map can be obtained, thereby improving the accuracy of water pollution monitoring.

[0037] 3. The present invention adopts a water pollution detection module to adaptively identify potential pollution areas in the water area feature distribution map of the water area to be monitored, and further perform feature sampling on these areas. This adaptive recognition can flexibly adjust the monitoring strategy according to the specific conditions of different water areas, enhance the universality of the system, and through quantitative analysis of the sampling results, the severity of water pollution can be accurately evaluated, which helps to judge the change trend of water body pollution and provides a scientific basis for relevant departments to make timely treatment decisions.

[0038] 4. When the pollution degree of the water area to be monitored exceeds the preset threshold, the present invention automatically triggers the early warning mechanism, quickly responds to potential water pollution problems, helps to detect water quality anomalies as early as possible, takes timely countermeasures, reduces the impact of pollution on the ecological environment, effectively improves the monitoring efficiency, and avoids human omissions. Description of the Drawings

[0039] Figure 1 It is a schematic diagram of the overall system architecture of the present invention;

[0040] Figure 2 It is the architecture diagram of the multi-dimensional depth evaluation model in the present invention;

[0041] Figure 3 It is the schematic diagram of the working process of the multi-dimensional depth evaluation model in the present invention. Specific embodiments

[0042] Next, in combination with the accompanying drawings in the embodiments of the present invention Figure 1 to Figure 3 , the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0043] The embodiments of the present invention disclose a complex water area water pollution monitoring and early warning system based on image recognition, which is an environmental monitoring solution integrating a variety of advanced technologies, aiming to realize real-time, efficient and comprehensive monitoring of the water area environment. As shown in the accompanying drawings Figure 1 as follows:

[0044] An image acquisition module, and the image acquisition module includes a fixed-point image acquisition method and a mobile cruise image acquisition method.

[0045] Fixed-point image acquisition method: Around the water area to be monitored, a plurality of multi-spectral cameras are deployed at a predetermined interval distance to ensure that the entire water area can be covered. These cameras should have high resolution and wide viewing angle, and be able to capture spectral information in different bands, such as visible light, near-infrared, short-wave infrared, etc., and can capture different reflection characteristics of the water body to obtain more detailed information, so as to provide more accurate water pollution detection data. This information is crucial for analyzing water quality, detecting pollution and identifying biological species, etc. Through a wireless local area network (such as Wi-Fi or a dedicated wireless transmission system), data can be interconnected and shared among multiple multi-spectral cameras. In this way, even if an abnormal image is captured by a certain camera, it can be quickly transmitted to other cameras and the cloud platform for further analysis.

[0046] The deployment positions of multiple multi-spectral cameras need to be determined according to the actual water area situation and monitoring purpose. Consider the following factors:

[0047] 1. Water area type and size: For different types and sizes of water areas, different camera devices and camera deployment schemes are required to achieve real-time monitoring. For small water areas, a small number of cameras can be used for panoramic coverage; for large water areas, multiple cameras are required for sub-region real-time monitoring.

[0048] 2. Monitoring Purposes and Requirements: Determine the deployment locations of cameras based on the monitoring purposes and requirements. For example, for water pollution monitoring, cameras can be directly deployed at nearby monitoring stations; for monitoring water body oil spills or leaks, cameras need to be deployed near the problem areas.

[0049] 3. Water Depth and Water Flow Velocity in the Water Area: For water areas with different water depths and water flow velocities, different types of cameras need to be selected, and the stability and safety of the cameras need to be considered. For example, high-definition cameras can be used in water areas with relatively shallow water depths, while more professional equipment or divers need to be used for monitoring in water areas with relatively deep water depths.

[0050] 4. Natural Environment and Meteorological Factors: The natural environment and meteorological factors will also affect the deployment locations of cameras. For example, in water areas with strong winds and waves, cameras need to be deployed in safe areas and the stability of the cameras needs to be ensured; for complex surrounding environments of water areas, a large amount of data analysis is required to determine the optimal deployment locations.

[0051] Mobile Cruising Image Acquisition Method: Use drones equipped with high-definition cameras for mobile cruising to fill the blind spots that may exist in fixed-point image acquisition. The drones adopt a preset path planning algorithm to plan the flight path and altitude of the drones according to the characteristics of the water area to ensure the best image acquisition effect and avoid collisions with other obstacles (such as trees, buildings, etc.). The drones should have strong stability and intelligent control systems, and be able to autonomously complete the preset cruising route and capture high-resolution images. Operators can adjust the flight direction and altitude of the drones in real time through the remote controller to deal with emergencies or obtain more detailed image information.

[0052] The water area images collected are transmitted to the cloud platform in real time through the wireless network for storage and processing. The cloud platform should have the characteristics of high throughput and low latency to ensure the real-time and integrity of the images. The cloud platform provides a large amount of storage space for storing the image data collected from multi-spectral cameras and drones. These data can be classified and indexed according to information such as time, location, and weather, which is convenient for subsequent retrieval and analysis. The image acquisition module marks the collection time, location, and weather information of each water area image with a timestamp. These information are crucial for subsequent data analysis and anomaly detection. Through the analog-to-digital converter (ADC), the collected analog image signals are converted into digital image signals. The digital image signals have higher resolution and richer color information, which is convenient for subsequent image processing and analysis. The image acquisition module should be integrated with other relevant systems (such as water quality monitoring systems, meteorological systems, etc.) to achieve data sharing and collaborative analysis. Ensure the compatibility between the image acquisition module, the cloud platform, the drones, and the multi-spectral cameras to avoid problems such as data format mismatch or transmission errors.

[0053] Through the above specific implementation manners, the image acquisition module can achieve comprehensive, efficient, and real-time monitoring of the water area to be monitored, providing strong data support for water quality assessment, ecological protection, disaster warning, etc. The acquired water area images are transmitted to the cloud platform in real time for storage and processing. The cloud platform should have powerful data processing capabilities and storage space to handle a large amount of image data.

[0054] Image preprocessing module. Each time the image acquisition device captures an image, it can record an accurate timestamp. Based on this timestamp, the system can obtain the environmental conditions at the acquisition moment from environmental monitoring devices (such as weather sensors, light sensors, etc.) or historical data. The environmental conditions include: Daylight illumination conditions: strong light, soft light, cloudy day, etc. Nighttime illumination conditions: complete darkness, weak night light, moonlight, etc. Rain, snow, and fog meteorological conditions: rainy day, snowy day, etc. This association can be achieved by embedding a GPS module and sensors in the image acquisition system to obtain meteorological and illumination data at the acquisition location in real time.

[0055] According to different environmental conditions, the preprocessing operations should be different. A conditional judgment module can be used to select different image processing strategies based on the environmental conditions. The main preprocessing operations may include:

[0056] Daylight illumination conditions: Contrast enhancement (enhancing the brightness difference in the image), white balance adjustment (ensuring natural image colors and avoiding overly cold or hot tones), noise removal (using filtering algorithms to remove image noise that may appear under strong light); In the case of sufficient daylight, the color information in the image may be more abundant. Therefore, color correction methods (such as white balance adjustment, color saturation adjustment, etc.) can be used to optimize the color performance of the image. For geometric distortion problems (such as perspective distortion, rotational distortion, etc.) that may occur during image acquisition, geometric transformation methods (such as affine transformation, projective transformation, etc.) can be used to correct these distortions.

[0057] Nighttime illumination conditions: Image brightness enhancement (using methods such as gamma correction, histogram equalization, etc. to enhance the image brightness), noise reduction processing (nighttime images usually have more noise, and median filtering, mean filtering, etc. can be used), edge enhancement (using methods such as Canny edge detection to highlight the object contours in the image); Since the nighttime light is relatively dim, the image may exhibit characteristics such as low brightness and low contrast. Therefore, the color image can be converted to a grayscale image through grayscale processing to reduce the amount of data to be processed. Further, a binarization method can be used to convert the grayscale image to a black-and-white image to more clearly present the targets in the image. Image enhancement techniques (such as histogram equalization, contrast stretching, etc.) are used to improve the brightness and contrast of the image, making the targets more prominent.

[0058] Rain, snow, and fog meteorological conditions: Removal of rain and snow (removing the interference of rain and snow through image dehazing technology and motion compensation), enhancement of image clarity (using image dehazing technology such as dark channel prior method, Retinex algorithm, etc.), image enhancement (contrast enhancement or brightness adjustment to help improve image visibility); Images collected under adverse weather conditions (such as rain and snow) may contain a large amount of noise. Therefore, denoising methods (such as mean filtering, median filtering, Gaussian filtering, etc.) need to be adopted to eliminate this noise and improve image quality. For the problems of image blurring or distortion caused by weather conditions, image restoration techniques (such as inverse filtering, Wiener filtering, etc.) can be used to restore the original information of the image.

[0059] These preprocessing steps will be dynamically selected based on different environmental conditions to ensure that the image quality adapts to subsequent analysis tasks as well as possible.

[0060] To improve system efficiency and meet real-time requirements, a distributed edge computing architecture is adopted. Under this architecture, the preprocessing of water area images will be carried out in parallel on multiple edge devices instead of being centralized on a single central server. Each edge device preprocesses the received image according to the environmental conditions where it is located. Each acquisition device (such as a camera) is equipped with sensors and obtains the current environmental conditions based on the image acquisition timestamp and environmental data. At the edge computing node, for different environmental conditions, the system will automatically allocate appropriate preprocessing tasks. For example, during the day with good lighting, the processing tasks may focus on image contrast and color balance; while at night, they may focus on brightness enhancement and noise removal. The edge computing device itself has data processing, computing, and storage capabilities, enabling efficient parallel processing and reducing the burden on the central server. The system will distribute the image data to different computing nodes (such as different edge computing units or local devices), and each node preprocesses the image according to the preconfigured environmental conditions and returns the processed image to the central system for further analysis or storage. On the edge computing node, hardware acceleration of the preprocessing algorithm (such as using accelerators like GPUs, FPGAs, etc.) can be carried out to improve the image processing speed and efficiency.

[0061] Through the combination of the image preprocessing module and the distributed edge computing architecture, water area images under different environmental conditions can be preprocessed efficiently. By obtaining environmental data and timestamps in real time, the system can select appropriate processing strategies to maximize image quality and provide clearer and more accurate image data for subsequent analysis tasks. The advantage of this method is that it can reduce data transmission latency, improve processing speed, and achieve more efficient resource utilization.

[0062] The data fusion module extracts color, texture, shape, and edge features from the preprocessed water area images through a multi-task visual representation learning framework, and uses a quadratic data information fusion model to fuse the water area image data collected by multiple multi-spectral cameras and drones to form a feature distribution map of the water area to be monitored.

[0063] The data fusion module uses a multi-task visual representation learning framework to extract features from the preprocessed water area images. This framework can simultaneously process multiple visual tasks, such as color recognition, texture analysis, shape detection, and edge detection. Color information in the image is extracted using methods such as color space conversion and color histogram. This information helps to identify different objects and regions in the water area. Texture analysis algorithms, such as gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), etc., are used to extract texture information in the image. Texture features are very useful for identifying changes and anomalies on the water surface. Shape information in the image is extracted using methods such as shape description and contour matching. This information helps to identify the object contours and edges in the water area. Edge detection algorithms, such as Canny operator, Sobel operator, etc., are used to extract edge information in the image. Edge features are very important for identifying object boundaries and details in the water area.

[0064] After extracting color, texture, shape, and edge features, a quadratic data information fusion model is used to fuse the water area image data collected by multiple multi-spectral cameras and drones. The data fusion process includes: fusing features from different data sources to form a more comprehensive feature representation. This can be achieved through methods such as feature splicing, feature mapping, or feature alignment. Based on the feature-level fusion, the features are further analyzed and processed to form a decision result. This can be achieved through methods such as classifiers, clustering algorithms, or regression models.

[0065] After data fusion, the obtained feature information is integrated into the feature distribution map of the water area to be monitored. This distribution map can intuitively display the different feature distributions and change situations in the water area, providing strong support for subsequent monitoring and analysis.

[0066] In summary, the data fusion module extracts color, texture, shape, and edge features from the preprocessed water area images through a multi-task visual representation learning framework and a quadratic data information fusion model, and fuses the water area image data collected by multiple multi-spectral cameras and drones to form a feature distribution map of the water area to be monitored. This process provides strong support for water area monitoring and analysis.

[0067] The multi-task visual representation learning framework combines the characteristics of pre-trained convolutional neural network (CNN) and multi-task learning, and is used to extract color, texture, shape, and edge features from the input water area images, and encode and reconstruct low-dimensional feature vectors.

[0068] An autoencoder is a neural network model composed of an encoder and a decoder. The encoder is responsible for compressing the input data into a low-dimensional feature vector, while the decoder is responsible for reconstructing these low-dimensional feature vectors into an approximate representation of the original data. In this solution, the autoencoder uses a pre-trained convolutional neural network (CNN) as a feature extractor and combines a shared feature extraction layer and task-specific layers for multi-task feature representation learning.

[0069] Use a pre-trained CNN model (such as VGG, ResNet, etc.) as a feature extractor. Input the input water area image into the pre-trained CNN to extract features such as color, texture, shape, and edges. Select appropriate feature layers from multiple convolutional layers of the pre-trained CNN for feature extraction. These feature layers are usually located in the middle and deep layers of the network and can capture the high-level semantic information of the image. The shared feature extraction layer consists of the feature extraction part of the pre-trained CNN and is used to extract the general features of the input image. Perform feature extraction on the input water area image to generate shared feature vectors. These shared feature vectors will be used as the input for the subsequent task-specific layers. The task-specific layers include multiple parallel fully connected layers or convolutional layers, and each layer corresponds to a specific task (such as color feature extraction, texture feature extraction, shape feature extraction, and edge feature extraction). Extract the features related to the specific task from the shared feature vectors. These features will be used for subsequent multi-task feature representation learning.

[0070] Input the input water area image into the shared feature extraction layer to generate shared feature vectors. Input the shared feature vectors into the fully connected layer or convolutional layer of the encoder to generate low-dimensional feature vectors. Input the low-dimensional feature vectors into the fully connected layer or convolutional layer of the decoder to generate the feature vectors of the reconstructed image. Input the feature vectors of the reconstructed image into the output layer of the decoder to generate the reconstructed image. Define a multi-task loss function, including reconstruction loss and multiple task-specific losses (such as color feature loss, texture feature loss, shape feature loss, and edge feature loss). By optimizing the multi-task loss function, the goal of multi-task feature representation learning is achieved. Use a large amount of labeled water area image data to train the autoencoder. During the training process, optimize the multi-task loss function simultaneously, so that the encoder can extract feature vectors related to multiple tasks, and at the same time the decoder can accurately reconstruct the original image.

[0071] In summary, by combining the characteristics of pre-trained convolutional neural networks and multi-task learning, an efficient multi-task visual representation learning framework is constructed. This framework can automatically extract the color, texture, shape, and edge features of the input water area image, and perform encoding and reconstruction of low-dimensional feature vectors, providing strong support for the subsequent water pollution detection module.

[0072] In the secondary data information fusion model, first, a BP neural network is used for the first fusion of local features and global features. The extracted local features and global features are combined into a high-dimensional feature vector and input into the BP neural network to represent the spatial and temporal dimensional features of the water area image. The set of combined high-dimensional feature vectors is X = {x1,..., x g ,..., x n}, where g represents the ordinal number of the high-dimensional feature vector, 1 ≤ g ≤ n, and n represents the total number of high-dimensional feature vectors. The BP neural network adjusts the neuron weights through the hidden layer activation function to learn the relationship between local features and global features. The neuron weight coefficient iteration equation is expressed as:

[0073]

[0074] In formula (1), W ij (t + 1) represents the weight coefficient from the i-th input node to the j-th hidden node at the (t + 1)-th iteration. t + 1 represents the number of iterations, i represents the ordinal number of the input node, 1 ≤ i, and j represents the ordinal number of the hidden node, 1 ≤ j. W ij (t) represents the weight coefficient from the i-th input node to the j-th hidden node at the t-th iteration. v is a weighting factor used to adjust the speed of weight update. τ is the step size of neuron iteration, and L is the loss function of the BP neural network. represents the partial derivative of the loss function L with respect to the weight coefficient W ij (t). θ represents the partial derivative. The BP network reduces the output error by updating the weight coefficient. The iteration equation given in the formula describes the update method of the weight coefficient, and the amplitude of each weight adjustment is controlled by the step size (learning rate). The iteration process includes two steps: forward propagation and backward propagation. Forward propagation passes the input data (i.e., local and global features) to the output layer of the network to obtain the predicted value of the network. Backward propagation calculates the partial derivative of the loss function with respect to the weights and then adjusts the weights through the gradient descent method.

[0075] The BP neural network continuously adjusts the weights of neurons through the backpropagation algorithm so that the output of the network approaches the target output, thereby minimizing the loss function. The loss function represents the difference between the network output and the target output. The BP neural network reduces the difference between the actual output and the target output by minimizing the loss function. Then, the loss function of the BP neural network is expressed as:

[0076]

[0077] In formula (2), L represents the loss function of the BP neural network, y g represents the given target output. represents the actual output of the said BP neural network; the loss function measures the gap between the actual output and the target output, and the goal of the neural network is to continuously adjust the network parameters (i.e., weights) through backpropagation to minimize this loss. By continuously calculating the partial derivatives of the loss function and updating the weights, the network can gradually learn how to extract useful information from the input features and finally achieve the fusion of local features and global features.

[0078] Then, an interpolation method is used to synchronize the time of the water area image features after the first fusion, so that the features of the image at different time points can be matched or aligned. This usually involves image resampling to align the resolution or position of the image at different time points. And the water area image features after the first fusion are spatially registered by solving the transformation matrix, with the aim of aligning the water area image features at different spatial positions. This step can adopt methods such as affine transformation and non-linear transformation, and select the appropriate registration technology according to actual needs. The key to registration is to make the spatial correspondence of the image features from different perspectives or different modalities, so that data from different sources can be merged into a unified spatial framework.

[0079] Then, based on the D-S evidence model, the feature information after the first fusion at the same time, different spaces and different modalities is secondarily fused. The D-S evidence theory is an information fusion method for dealing with uncertainty and ambiguity. Its main idea is to quantify the relationship between them by combining information from different sources, so as to obtain a more reliable result. In this model, the image features are input into the D-S evidence theory as "evidence elements". These evidence elements can represent image features from different sources, different modalities or different times. The D-S evidence theory quantifies the mutual relationship between evidence elements through the belief function. The belief function reflects the degree of support of an evidence for a certain hypothesis. Through the fusion rule of the D-S theory, different evidences are comprehensively processed to obtain the final fusion result. This usually includes weighting or summing the degrees of support to calculate the final decision or inference result. The D-S evidence model takes the image features after the first fusion as evidence elements and quantifies the relationship between evidence elements through the belief function to obtain the final fusion result. The calculation formula is:

[0080]

[0081] In formula (3), m(A) is the evidence support degree after fusion, and m g (A) represents the evidence support degree of the g-th image.

[0082] Through the above process, on the one hand, local and global features are fused once through a BP neural network, and on the other hand, time synchronization and spatial registration are performed through an interpolation method. Finally, a D-S evidence model is used for secondary fusion. This model obtains comprehensive information from features of different sources (space, time, modality) by quantifying the relationship between evidence elements to achieve a more accurate and robust feature fusion result. The overall process combines the advantages of the BP neural network and the D-S evidence theory by gradually fusing information of multiple modalities (local features and global features, time and spatial features, etc.), which can effectively improve the representation ability and fusion effect of image features. This multi-level and gradually optimized fusion strategy can demonstrate good performance in multi-modal analysis and processing tasks of water area images.

[0083] The hardware working environment of the secondary data information fusion model usually includes the following aspects:

[0084] Server: A high-performance server is a key device for running the secondary data information fusion model, usually equipped with a powerful CPU (such as a multi-core processor), a large amount of memory (such as DDR4 or higher version memory), and a high-speed storage device (such as an SSD solid-state drive). These configurations can ensure the efficiency and stability of the model when processing large-scale data sets.

[0085] GPU / TPU acceleration card: For deep learning models, especially complex secondary data information fusion models, a GPU (graphics processing unit) or TPU (tensor processing unit) acceleration card can significantly improve the calculation speed. These acceleration cards accelerate matrix operations and the execution of deep learning algorithms through parallel computing technology.

[0086] Storage device: Including hard disk arrays, network attached storage (NAS), or storage area network (SAN), etc., which are used to store a large amount of raw data, preprocessed data, model parameters, and fusion results. These storage devices should have high-speed read and write capabilities and high reliability to ensure the integrity and availability of data.

[0087] Database system: The database system is used to manage and organize data, providing efficient data access and query functions. For the secondary data information fusion model, a relational database or a non-relational database may be used to store and manage data.

[0088] Network devices: Including switches, routers, firewalls, etc., which are used to build a stable and efficient network environment to ensure the security and real-time nature of data during acquisition, transmission, and processing.

[0089] Network bandwidth: Sufficient network bandwidth can support the rapid transmission of a large amount of data, ensuring that the secondary data information fusion model can obtain and process the required data in a timely manner.

[0090] Power supply equipment: including uninterruptible power supply (UPS), generators, etc., which are used to ensure the power supply of the hardware working environment and prevent system downtime or data loss caused by power failures.

[0091] In summary, the hardware working environment of the secondary data information fusion model needs to comprehensively consider multiple aspects such as high-performance computing equipment, data storage and access equipment, network communication equipment, and other auxiliary equipment to ensure that the model can operate efficiently and stably and meet the needs of actual applications.

[0092] Comparative experiments were carried out using the secondary data information fusion model (Group A) and the weighted average image fusion method (Group B) respectively. Watershed image data from multiple multispectral cameras and drones were collected, and the data were preprocessed, including denoising, enhancing contrast, etc., to ensure that the data had a certain degree of difference and diversity. The data set was randomly divided into two groups, with 30 data in each group, ensuring that the data characteristics of the two experimental groups were similar. Using a set of data, according to the secondary data information fusion model described above, first use the BP neural network for the first fusion of local features and global features, then perform time synchronization and spatial registration, and finally perform secondary fusion on the feature information after the first fusion of the same time, different spaces, and different modalities based on the D-S evidence model to obtain the final fusion result. For the other set of data, directly perform weighted average for fusion to obtain the final fusion result. According to the different processing methods of Group A and Group B, the experiment was carried out within the same time period. Each experimental group was compared with the real watershed feature map, and the accuracy rate, recall rate, and F1 score indicators of the results were calculated. The above experiment was repeated five times with different data, and the comparison results were recorded in Table 1.

[0093] Table 1 Result statistical table

[0094]

[0095] According to the experimental result table, it can be seen that the secondary data information fusion model of Group A performs better than the weighted average image fusion method of Group B in terms of accuracy rate, recall rate, and F1 score indicators, and has higher stability and reliability. At the same time, the standard deviation of the experimental results of Group A is smaller in the repeated ten experiments, showing more stability. Therefore, it can be concluded that using the method of Group A for the secondary fusion of multispectral camera and drone image data can more effectively improve the accuracy and reliability of the watershed feature distribution map and achieve more accurate water pollution monitoring and early warning.

[0096] Water pollution detection module, which is used to adaptively identify potential water pollution areas in the characteristic distribution map of the water area to be monitored and perform feature sampling, and quantitatively evaluate the degree of water pollution in the water area to be monitored according to the sampling results. This water pollution detection module mainly includes components such as an adaptive threshold segmentation method, an automatic sampler, and a multi-dimensional depth evaluation model. According to the pixel density and feature similarity in the characteristic distribution map of the water area to be monitored, the adaptive threshold segmentation method is used to segment potential water pollution areas to distinguish potential pollution areas from non-pollution areas. Specifically, based on pixel density and feature similarity, this adaptive threshold segmentation method divides the image into multiple sub-regions, then uses the Otsu algorithm to determine the threshold in each sub-region, and finally forms a binary mask of the pollution area for water quality feature recognition. According to the obtained binary mask of the pollution area, an automatic sampler is used to sample the water quality features of potential water pollution areas. The automatic sampler adopts an automatic sampling scheme based on a mobile robot, which can move autonomously on the water surface and transmit the sampling data to the computer for analysis. A multi-dimensional depth evaluation model is used to quantitatively analyze the sampled water quality features to judge the concentration, type, and distribution characteristics of pollutants. This multi-dimensional depth evaluation model combines various technical means such as machine learning, deep learning, computer vision, and statistics, and can comprehensively and efficiently evaluate and analyze problems such as the complexity, diversity, and time-variability of water area pollutants. Specifically, this model can comprehensively evaluate multiple pollutant parameters, including water temperature, pH value, dissolved oxygen, chemical oxygen demand, total phosphorus, total nitrogen, copper, zinc, lead and other indicators, and predict the concentration, distribution, and change trend of pollutants through multi-dimensional data analysis and modeling methods.

[0097] Combining the above steps, this water pollution detection module can efficiently perform adaptive identification and feature sampling on water pollution areas.

[0098] The multi-dimensional depth evaluation model, as shown in the appendix Figure 2 shown, includes an input layer, a data layer, an adaptive weight layer, a model layer, an optimization layer, and an output layer. The working method of the multi-dimensional depth evaluation model, as shown in the appendix Figure 3 shown, includes the following steps:

[0099] Step 1: The input layer receives and standardizes the processing

[0100] Responsible for receiving the water quality feature data collected by the automatic sampler. These data may include various water quality indicators, such as pH value, dissolved oxygen, turbidity, heavy metal content, etc. Since the dimensions and value ranges of different water quality indicators may be different, in order to unify processing and analysis, it is necessary to standardize the input water quality feature data. Standardization processing usually includes steps such as data cleaning (removing outliers, missing values, etc.), data conversion (such as normalization, standardization, etc.), and data scaling (such as min-max scaling, Z-score scaling, etc.).

[0101] Step 2: The data layer obtains calculation parameters and targets

[0102] It includes calculation scale (such as the size and complexity of the model, etc.), objective function (such as minimizing error, maximizing accuracy, etc.), constraint conditions (such as the value range of variables, the physical meaning of the model, etc.) and variable range (such as the value range of water quality indicators). Clearly define the targets that the model needs to predict or evaluate, including the concentration, types and distribution characteristics of pollutants.

[0103] Step 3: Feature extraction and weight adjustment in the model layer

[0104] Use deep convolutional neural network (CNN) and recurrent neural network (RNN) to further extract features from the standardized water quality features. CNN can extract local features of water quality features, such as the morphology and color distribution of pollutants, etc.; RNN can capture the time series information of water quality features, such as the change trend of pollutants, etc. Adaptively adjust the weights of the multi-dimensional feature vectors according to the calculation objectives. The adaptive weight layer can dynamically adjust the importance of different features according to the learning situation of the model and the characteristics of the data, so as to improve the prediction accuracy and robustness of the model.

[0105] Step 4: Establish a multi-dimensional evaluation mathematical model

[0106] Establish a multi-dimensional evaluation mathematical model according to the extracted multi-dimensional features, calculation scale, objective function, constraint conditions and variable range. This model can be a complex non-linear function used to describe the relationship between water quality features and target variables. Adopt a parallel computing method to distribute the computing tasks to multiple processors or computing nodes to improve the computing efficiency. According to the objective function and the distribution of computing nodes, obtain the neighbor list of computing nodes for subsequent computing and communication.

[0107] Step 5: Feedback training and parameter selection in the optimization layer

[0108] Use the reconstruction error backpropagation algorithm to perform feedback training on the multi-dimensional evaluation mathematical model. This algorithm can adjust the parameters of the model according to the prediction error and gradient information of the model, so as to continuously improve the prediction accuracy of the model. Set reasonable thresholds and iteration times to control the convergence speed and performance of the model during the training process. The adaptive parameter selection method can dynamically adjust the thresholds and iteration times according to the training situation of the model and the characteristics of the data to obtain better training effects.

[0109] Step 6: The output layer outputs the results

[0110] The calculation results are output through the output layer. These results may include the concentration, type, and distribution characteristics of pollutants, etc., which can be used for subsequent decision-making support and environmental governance work. Interpret and analyze the output results to understand the pollution situation of the water area and the distribution characteristics of pollutants. Develop corresponding treatment plans based on the output results, such as reducing pollutant emissions and strengthening water quality monitoring, etc., to improve the environmental quality of the water area.

[0111] In summary, through the collaborative work of multiple levels such as the input layer, data layer, adaptive weight layer, model layer, optimization layer, and output layer, the multi-dimensional depth evaluation model realizes the accurate extraction and quantitative evaluation of water quality characteristics. The application of this model will provide strong support for water environmental protection and governance work.

[0112] The multi-dimensional depth evaluation model is a data model used to evaluate and analyze data in multiple dimensions, usually used in the fields of machine learning, data science, and artificial intelligence, and requires a certain hardware environment to support its efficient computing and storage requirements. The following are the key components of the hardware working environment of the multi-dimensional depth evaluation model:

[0113] Central Processing Unit (CPU): Used to execute the main computing tasks of the model. Especially when there is no dedicated Graphics Processing Unit (GPU), the CPU has a heavy burden. Usually, a multi-core, high-frequency processor is required to ensure efficient data processing and model training.

[0114] Graphics Processing Unit (GPU): In deep learning models, the GPU is used to accelerate a large number of matrix operations and parallel computing tasks, so it is crucial for the multi-dimensional depth evaluation model. Modern deep learning models, especially when dealing with evaluation tasks involving a large amount of data, usually rely on NVIDIA's CUDA technology for acceleration.

[0115] Tensor Processing Unit (TPU): The TPU developed by Google can also be used to accelerate deep learning tasks, especially optimized for frameworks such as TensorFlow, and is suitable for tasks that require large-scale data and high-frequency training.

[0116] Solid State Drive (SSD): Used to store a large amount of training data, model parameters, and intermediate calculation results. The high-speed read and write capabilities of the SSD can significantly speed up the data loading and storage process and reduce the training time.

[0117] Distributed storage system: When conducting large-scale training, a single SSD may not be sufficient to meet the storage requirements. Distributed storage systems such as HDFS (Hadoop Distributed File System) or Ceph, etc., can provide efficient and scalable storage solutions to support data parallel processing of multiple nodes.

[0118] Memory (RAM): Deep learning and data analysis tasks usually require a large amount of memory to cache data, store the intermediate states of the model, and accelerate the calculation process. For multi-dimensional depth evaluation models, memory with a capacity of 64GB or higher is usually required.

[0119] High-speed network connection: Especially in a distributed computing environment, the communication and data transfer speed between nodes have a crucial impact on the training efficiency of the model. Gigabit Ethernet (10GbE) or a higher-speed network is usually adopted to connect computing nodes.

[0120] Data center network: When deployed in a cloud computing platform or within an enterprise, the network architecture within the data center also needs to have high bandwidth and low latency to support large-scale parallel computing and the fast transfer of large amounts of data.

[0121] FPGA (Field Programmable Gate Array): Some deep learning applications and evaluation models may use FPGA for hardware acceleration. FPGA can be customized and optimized according to application requirements, and can provide lower latency and higher efficiency in certain tasks.

[0122] ASIC (Application-Specific Integrated Circuit): Some dedicated hardware platforms (such as Google's TPU) can also be used as hardware accelerators to accelerate specific computing tasks.

[0123] Cluster computing: In some extremely large multi-dimensional depth evaluation model applications, computing resources may be distributed across multiple computing nodes. These nodes are interconnected by a high-speed network to form a distributed computing cluster, providing stronger computing power and higher fault tolerance.

[0124] Cloud computing platform: Many companies and research institutions choose to use cloud computing platforms (such as AWS, Google Cloud, Microsoft Azure, etc.) for large-scale computing tasks. These platforms provide flexible resource scheduling, elastic scaling, and pay-as-you-go services, and can dynamically allocate computing resources and storage resources according to needs.

[0125] Efficient power supply: Deep learning model training and evaluation usually consume a large amount of power, especially GPU clusters and large servers. Therefore, it is very important to ensure that the system has an efficient power management and redundant backup mechanism.

[0126] Cooling system: In a high-load computing environment, the heat dissipation problem of hardware is a key issue. Especially when GPUs and CPUs are in a high-load state for a long time, effective heat dissipation solutions such as liquid cooling systems and air cooling systems are required.

[0127] Data backup system: To ensure data security and fault tolerance during the computing process, a real-time backup mechanism is usually required to avoid data loss caused by hardware failures.

[0128] Fault tolerance mechanism: In a large-scale distributed computing environment, hardware failures or network interruptions may cause computing tasks to be interrupted. Therefore, a fault tolerance mechanism needs to be designed, such as task rescheduling, redundant computing nodes, etc., to ensure the stability and reliability of the system.

[0129] The hardware working environment of the multi-dimensional depth evaluation model usually involves multiple computing, storage, network, and power management components. The selection and configuration of these components need to be customized according to the complexity of the model, the data scale, and the application scenario. With the development of technology, the acceleration of modern hardware and distributed computing enable the depth evaluation model to process larger-scale data and provide more efficient computing power.

[0130] Comparative experiments were conducted using the multi-dimensional depth evaluation model (Group A) and statistical analysis (Group B). Water quality data for a certain period were collected through an automatic water quality sampler, including characteristic information such as pollutant concentration, type, and color distribution. Each set of collected data included at least 30 water quality characteristic indicators, such as temperature, pH value, dissolved oxygen concentration, heavy metal concentration, pollutant type, and distribution characteristics, etc., to ensure that the data had a certain degree of difference and diversity. The data set was randomly divided into two groups to ensure that the data characteristics of the two experimental groups were similar. One set of data was used to evaluate the degree of water pollution according to the multi-dimensional depth evaluation model described above. For the other set of data, statistical analysis was used to evaluate the degree of water pollution. According to the different treatment methods of Group A and Group B, the experiment was carried out within the same time period. The number of water quality samples exceeding the threshold in each set of data was counted to evaluate the sensitivity of the two methods in water pollution degree evaluation. The above experiment was repeated ten times with different data, and the comparison results were recorded in Table 2.

[0131] Table 2 Statistical results table

[0132]

[0133]

[0134] By comparing the number of water quality samples exceeding the threshold statistically in each experiment for Group A and Group B, it can be found that the multi-dimensional depth assessment model (Group A) shows higher sensitivity in identifying samples with high pollution levels. This is because the multi-dimensional depth assessment model can handle complex non-linear relationships and high-dimensional data, thus more accurately evaluating the degree of water pollution. Although the statistical analysis method (Group B) can identify the relationship between water quality characteristics and the degree of water pollution, it may be affected by factors such as data distribution and feature correlation when dealing with complex water quality data, resulting in lower sensitivity. Therefore, the multi-dimensional depth assessment model is suitable for processing complex and high-dimensional water quality data, can accurately evaluate the degree of water pollution and identify samples with high pollution levels, providing strong support for water quality monitoring and management.

[0135] The early warning response module is used to carry out early warning response according to the quantification result of the water pollution degree of the water area to be monitored. When the water pollution degree of the water area to be monitored exceeds the preset threshold, the early warning mechanism is automatically triggered. First, based on the pollution degree quantification result provided by the water pollution detection module, the pollution degree of the water area to be monitored is evaluated and analyzed to determine whether to trigger the early warning mechanism. This pollution degree quantification result is based on the multi-dimensional depth assessment model, which can automatically identify and quantify the concentration, types and distribution of various types of pollutants in the water area, and make judgments and alarms according to the preset pollution degree threshold. When the pollution degree of the water area to be monitored exceeds the preset threshold, the early warning mechanism is automatically triggered and enters the early warning response stage. The early warning mechanism includes sub-modules such as alarm content generation, alarm information transmission and alarm record preservation. In the alarm content generation sub-module, according to the pollution degree assessment result, a pollution monitoring report is automatically generated using the GIS visualization platform and uploaded to the management platform. In the alarm information transmission sub-module, when the pollution monitoring report is uploaded to the management platform, early warning information is sent to the management personnel by means of SMS, email or mobile application push. In the alarm record preservation sub-module, the alarm information is saved to the database for convenient subsequent query and maintenance.

[0136] Combining the above steps, the early warning response module can efficiently monitor and give early warning responses to water area pollution situations in real time, ensuring the safety of water source areas. At the same time, through various effective alarm methods, pollution monitoring information can be transmitted to the management personnel in a timely manner, strengthening the management and control of water area pollution incidents.

[0137] The system includes an image acquisition module, an image preprocessing module, a data fusion module, a water pollution detection module, and an early warning response module. Data exchange and communication are carried out between the modules through a data transmission interface. The water area image collected by the image acquisition module is transmitted to the image preprocessing module for processing. The processed image is input into the data fusion module. The data fusion module forms a characteristic distribution map of the water area to be monitored through feature extraction and information fusion. The characteristic distribution map is input into the water pollution detection module for analysis, identification, and quantitative evaluation. After the detection is completed, the result is input into the early warning response module. Whether to trigger the early warning mechanism is judged according to the set early warning threshold, and an early warning message is sent to the management personnel.

[0138] So far, the technical solution of the present invention has been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or replacements to the relevant technical features, and the technical solutions after these changes or replacements will fall within the protection scope of the present invention.

Claims

1. A complex water area water pollution monitoring and early warning system based on image recognition, characterized in that, Including: An image acquisition module, which performs real-time acquisition of fixed-point images by deploying multiple multispectral cameras at intervals around the water area to be monitored, and uses an unmanned aerial vehicle equipped with a high-definition camera to conduct mobile cruising over the water area to be monitored to fill the blind spots of fixed-point image acquisition, and transmits the acquired water area images to the cloud platform in real time for storage and processing; An image preprocessing module, which performs different preprocessing operations on the acquired images based on the environmental conditions of water area image acquisition; A data fusion module, which extracts color, texture, shape and edge features from the preprocessed water area images through a multi-task visual representation learning framework, and uses a secondary data information fusion model to fuse the water area image data acquired by multiple multispectral cameras and unmanned aerial vehicles to form a feature distribution map of the water area to be monitored; A water pollution detection module, which first adaptively identifies and samples the potential water pollution areas in the feature distribution map of the water area to be monitored, and quantitatively evaluates the water pollution degree of the water area to be monitored according to the sampling results; An early warning response module, which conducts early warning response according to the quantitative results of the water pollution degree of the water area to be monitored. When the water pollution degree of the water area to be monitored exceeds the preset threshold, the early warning mechanism is automatically triggered.

2. The complex water pollution monitoring and early warning system based on image recognition according to claim 1, wherein, The multiple multispectral cameras achieve data intercommunication and sharing through a wireless local area network. The unmanned aerial vehicle equipped with a high-definition camera adjusts the direction and altitude of mobile cruising through a preset path planning algorithm and a remote controller. The image acquisition module marks the acquisition time, location and weather information of each water area image with a timestamp, and converts the acquired analog image signal into a digital image signal through an analog-to-digital converter.

3. The complex water pollution monitoring and early warning system based on image recognition according to claim 2, characterized in that, The image preprocessing module obtains the environmental conditions of water area image acquisition according to the marked timestamp, and performs different preprocessing operations on the water area images acquired under different environmental conditions in parallel by using a distributed edge computing method. The environmental conditions at least include daytime lighting conditions, nighttime lighting conditions and rain, snow, fog meteorological conditions.

4. The complex water pollution monitoring and early warning system based on image recognition according to claim 1, characterized in that The multi-task visual representation learning framework encodes the input water area image through an autoencoder into a low-dimensional feature vector representation, and maps the low-dimensional feature vector to the original image space to realize the reconstruction of the water area image. The autoencoder uses a pre-trained convolutional neural network to automatically extract color, texture, shape and edge features from the input water area image, and uses a shared feature extraction layer and a task-specific layer for multi-task feature representation learning.

5. The complex water area water pollution monitoring and early warning system based on image recognition according to claim 1, characterized in that, In the secondary data information fusion model, first, a BP neural network is used for the first fusion of local features and global features. The extracted local features and global features are combined into a high-dimensional feature vector and input into the BP neural network to represent the spatial and temporal dimension features of the water area image. The set of combined high-dimensional feature vectors is X = {x1,..., x g ,..., x n}, where g represents the ordinal number of the high-dimensional feature vector, 1 ≤ g ≤ n, and n represents the total number of high-dimensional feature vectors. The BP neural network adjusts the neuron weights through the hidden layer activation function to learn the relationship between local features and global features. The neuron weight coefficient iteration equation is expressed as: In formula (1), W ij (t + 1) represents the weight coefficient from the i-th input node to the j-th hidden node at the (t + 1)-th iteration, t + 1 represents the number of iterations, i represents the ordinal number of the input node, 1 ≤ i, j represents the ordinal number of the hidden node, 1 ≤ j, W ij (t) represents the weight coefficient from the i-th input node to the j-th hidden node at the t-th iteration, v is a weighting factor used to adjust the speed of weight update, τ is the step size of neuron iteration, L is the loss function of the BP neural network, represents the partial derivative of the loss function L with respect to the weight coefficient W ij (t), θ represents the partial derivative. The BP neural network reduces the difference between the actual output and the target output by minimizing the loss function. Then the loss function of the BP neural network is expressed as: In formula (2), L represents the loss function of the BP neural network, and y g represents the given target output, and represents the actual output of the BP neural network; Then, an interpolation method is used to synchronize the time of the water area image features after the first fusion, and the water area image features after the first fusion are spatially registered by solving the transformation matrix; Then, based on the D-S evidence model, the feature information after the first fusion at the same time, different spaces and different modalities is secondarily fused. The D-S evidence model takes the image features after the first fusion as evidence elements, and quantifies the relationship between the evidence elements through a belief function to obtain the final fusion result. The calculation formula is: In formula (3), m(A) is the evidence support degree after fusion, and m g (A) represents the evidence support degree of the g-th image.

6. The complex water area water pollution monitoring and early warning system based on image recognition according to claim 1, characterized in that, The water pollution detection module uses an adaptive threshold segmentation method to identify water quality characteristics in potential water pollution areas in the characteristic distribution map of the water area to be monitored, and uses an automatic sampler to sample water quality characteristics in potential water pollution areas. The adaptive threshold segmentation method performs adaptive threshold segmentation of water pollution areas based on pixel density and feature similarity; the water pollution detection module uses a multi-dimensional depth evaluation model to quantitatively analyze the sampled water quality characteristics to judge the concentration, type, and distribution characteristics of pollutants.

7. An image recognition-based complex water area water pollution monitoring and early warning system according to claim 6, characterized in that, The multi-dimensional depth evaluation model includes an input layer, a data layer, an adaptive weight layer, a model layer, an optimization layer, and an output layer. The operation of the multi-dimensional depth evaluation model includes the following steps: Step 1: Receive the water quality characteristics sampled by the automatic sampler through the input layer, and perform standardization processing on the input water quality characteristics; Step 2: Obtain calculation parameters, calculation objectives, and constraint conditions through the data layer. The calculation parameters and constraint conditions include calculation scale, objective function, constraint conditions, and variable range. The calculation objectives include the concentration, type, and distribution characteristics of pollutants; Step 3: The model layer extracts pollutant morphology, color distribution, and texture characteristics from the standardized water quality characteristics through a deep convolutional neural network and a recurrent neural network, and establishes a weight relationship between the primary and secondary dimensions through the adaptive weight layer. The adaptive weight layer adaptively adjusts the weights of multi-dimensional feature vectors according to the calculation objectives; Step 4: The model layer establishes a multi-dimensional evaluation mathematical model based on the extracted multi-dimensional features, calculation scale, objective function, constraint conditions, and variable range, and performs iterative calculation, parameter correction, and comparison between the calculation result and the true value on the multi-dimensional evaluation mathematical model through historical data. The model layer uses a parallel computing method to allocate calculation tasks to multiple processors or computing nodes, and obtains a neighbor list of computing nodes according to the objective function and the distribution of computing nodes; Step 5: The optimization layer performs feedback training on the multi-dimensional evaluation mathematical model through the reconstruction error backpropagation algorithm, and sets the threshold and the number of iterations through an adaptive parameter selection method; Step 6: Output the calculation result through the output layer.

8. The complex water pollution monitoring and early warning system based on image recognition according to claim 1, characterized in that, The early warning mechanism automatically generates a pollution monitoring report through the GIS visualization platform and uploads it to the management platform. The management platform sends early warning information to management personnel through SMS, email, or mobile application push.

Images