Deep learning-based pollution data analysis method and system

Abstract

The present application relates to the technical field of data analysis, and specifically to a pollution data analysis method and system based on deep learning, which comprises obtaining a historical pollution data set, determining pollution source coordinates through pollution concentration values, and establishing a pollution diffusion model using a convolutional neural network. The target water area is divided into grid cells, and a UAV cruising scheme carrying a hyperspectral camera is generated, each scheme containing the cruising state, cruising altitude value and pollution concentration prediction value of the grid cell. The UAV circulation path sequence is optimized by genetic algorithm, and the cruising path sequence is smoothed. The cruising altitude is adjusted according to the pollution concentration gradient. The present application adjusts the UAV cruising altitude by taking into account the pollution of water bodies of different types and concentrations under the hyperspectral camera, improving spectral resolution and reducing data processing volume.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a method and system for analyzing contaminated data based on deep learning. Background Technology

[0002] Water pollution is often insidious, making rapid analysis of water pollution using remote sensing technology essential. Water pollution is typically accompanied by changes in water color, composition, and temperature. Remote sensing essentially analyzes the absorption and backscattering characteristics of water components, establishes spectral indices through deep learning, and extracts water color anomalies by combining spatial adjacency relationships from remote sensing images. Currently, both satellites and drones equipped with hyperspectral cameras can perform remote sensing detection of water pollution. The former is limited by weather factors such as cloud cover and fixed scanning frequencies, while the latter offers higher resolution and greater flexibility. Drones used for water body inspections typically have a set cruising altitude; theoretically, lower altitudes result in clearer spectra, but excessively low altitudes lead to excessive data processing. Furthermore, with a fixed upper limit for spectral resolution, wastewater diffuses and mixes with normal water bodies; the further away from the pollution source, the lower the concentration of wastewater, and the less distinct the spectral characteristics per pixel. Therefore, lower cruising altitudes are necessary to ensure adequate resolution. During the analysis of pollution data, it is necessary to analyze historical pollution data for the local wastewater type in order to improve the accuracy and efficiency of pollution data collection and processing during drone patrols for the same type of pollution objects and pollution concentrations. Summary of the Invention

[0003] (1) Technical problems to be solved The purpose of this invention is to provide a pollution data analysis method and system based on deep learning to solve the problem of improving the accuracy and processing efficiency of water pollution spectral data acquisition.

[0004] (2) Technical solution To achieve the above objectives, in one aspect, the present invention provides a deep learning-based method for analyzing contamination data, the method comprising: Obtain historical pollution datasets for the target water area, including pollution type, pollution concentration values ​​recorded by geographic coordinates and timestamps, and corresponding drone cruising altitudes. Obtain pollution source coordinates from the pollution concentration values. Use a convolutional neural network to obtain a pollution diffusion model based on the pollution type, pollution source coordinates, and pollution concentration values ​​recorded by geographic coordinates and timestamps. Divide the target water area into grid cells. A drone cruise plan equipped with a hyperspectral camera is generated. The cruise plan consists of multiple individuals. Each individual includes the cruise status, cruise altitude value, and corresponding pollution concentration prediction value of each grid cell. The cruise status value is either 0 or 1. A cruise status value of 1 indicates that the drone passes through the corresponding grid cell. The cruise altitude value is randomly generated within the safe cruise altitude threshold of the drone. The pollution concentration prediction value is obtained by using a pollution diffusion model to obtain the pollution type, the geographical coordinates of the grid cell, the coordinates of the pollution source, and the pollution concentration of the pollution source. The cruise strategy is iteratively optimized, including calculating the information entropy of each hyperspectral camera unit pixel, selecting a preset number of winning individuals based on the information entropy using a roulette wheel selection method, cross-crossing and mutating the winning individuals and removing non-adjacent individuals from the grid cells to obtain offspring individuals, replacing the initial population with offspring individuals, and repeating the iterative evolution process until the fitness improvement values ​​of the offspring and parent individuals are lower than the set fitness threshold, and smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individuals after iterative evolution and using them as the UAV cruise path.

[0005] Furthermore, the method for obtaining the coordinates of the pollution source from the pollution concentration value includes: Pollution concentration values ​​from historical pollution datasets By geographical coordinates Spatial interpolation yields the concentration distribution function. ; Calculate the gradient field of the concentration distribution function The maximum point of the concentration distribution function is found using the gradient ascent method, which serves as the coordinate of the pollution source.

[0006] Furthermore, the method for calculating the information entropy of each individual hyperspectral camera unit pixel includes: Through concentration distribution function Spatial sampling is performed within the pixel coverage area to obtain the concentration values ​​of the sampling points. ,in For the first Coordinates of each sampling point; cruising altitude values ​​through each grid cell. Calculate ground pixel resolution ,in For sensor pixel size, The focal length of the lens; calculate the standard deviation of pollution concentration in the area covered by a unit pixel. ,in This refers to the number of sampling points within the coverage area of ​​a single pixel. Let K be the pollution concentration value at the kth sampling point. This represents the average pollution concentration value within the coverage area of ​​a single pixel. Establish the information entropy of a unit pixel ,in A preset stability constant is set when the pollution concentration values ​​of all sampling points within the unit pixel coverage area are equal. = 0.

[0007] Furthermore, the method for removing non-adjacent individuals from the grid cells includes: For each individual, an adjacency graph is created for the grid cells with a cruise state of 1, and the grid cells are... As grid nodes, when two grid nodes and satisfy and Edges are established between corresponding nodes; the adjacency graph is traversed using a depth-first search algorithm to count the number of connected components. ; When the number of connected components At that time, calculate the sum of the information entropy of each connected component. ,in For the first The set of grid cells contained in each connected component; retain the connected component with the largest sum of information entropy, and set the cruise state of the grid cells in other connected components to 0.

[0008] Furthermore, the method for smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individual after iterative evolution and using them as the drone's cruise path includes: Extract all grid cells with a cruise state of 1 from the winning individuals after the complete iterative evolution, and obtain the geographic coordinate center point and corresponding cruise altitude value of each grid cell. Calculate the pollution concentration gradient magnitude of neighboring grid cells surrounding the target grid cell. The partial derivatives pass through a radius of The concentration distribution within the sampling area was calculated; the distance between grid cells was calculated based on the coordinates of the center points of adjacent grid cells. ,in For adjacent grid cells, the minimum detectable pollution concentration difference is calculated. Divide by the distance between grid cells Obtain the gradient threshold ; When the pollution concentration gradient amplitude Exceeding the gradient threshold At that time, calculate the height adjustment amount. ,in The baseline cruising altitude value, For safety limit factors; the reference cruising altitude value With limited height adjustment The values ​​are added together to obtain the temporary altitude adjustment value; the temporary altitude adjustment value is then compared with the minimum safe flight altitude. The adjusted cruising altitude value is obtained by comparing and selecting the larger value. Arranged in order of geographic coordinates to form a sequence of waypoints containing the adjusted cruise altitude values; Calculate the adjusted cruise altitude values ​​of adjacent waypoints The absolute value of the difference between them will determine the drone's maximum climb speed. Distance between grid cells The product divided by the drone's flight speed Obtain highly smooth constraint values When the absolute value of the difference between the cruise altitude values ​​of adjacent grid cells after adjustment exceeds the altitude smoothing constraint value, at the adjacent path point that violates the constraint... Calculate the number of interpolation path points Insertion between adjacent path points through equally spaced linear interpolation An intermediate path point is selected, ensuring that the absolute value of the height difference between all adjacent path points does not exceed the height smoothness constraint value. The constraint check and path point insertion process is repeated until the entire path sequence meets the smoothness requirement. The final path point sequence is then arranged in chronological order to generate a drone cruise path containing geographic coordinates and cruise altitude values.

[0009] Based on the same inventive concept, this invention also provides a pollution data analysis system based on deep learning, the system comprising: The pollution diffusion analysis module is used to acquire historical pollution datasets of the target water area, including pollution type, pollution concentration values ​​recorded by geographic coordinates and timestamps, and the corresponding drone cruising altitude. The pollution source coordinates are obtained through the pollution concentration values. The pollution type, pollution source coordinates, and pollution concentration values ​​recorded by geographic coordinates and timestamps are used to obtain a pollution diffusion model through a convolutional neural network. The target water area is divided into grid cells. The data initialization module is used to generate a drone cruise plan equipped with a hyperspectral camera. The cruise plan consists of multiple individuals, each of which includes the cruise status, cruise altitude value, and corresponding pollution concentration prediction value of each grid cell. The cruise status value is either 0 or 1, with a cruise status value of 1 indicating that the drone passes through the corresponding grid cell. The cruise altitude value is randomly generated within the drone's safe cruise altitude threshold. The pollution concentration prediction value is obtained by using a pollution diffusion model based on the pollution type, the geographic coordinates of the grid cell, the coordinates of the pollution source, and the pollution concentration of the pollution source. The data analysis module is used to iteratively optimize the cruise plan. This includes calculating the information entropy of each hyperspectral camera unit pixel, selecting a preset number of winning individuals based on the information entropy using a roulette wheel selection method, cross-crossing and mutating the winning individuals, and removing non-adjacent individuals from the grid cells to obtain offspring individuals; replacing the initial population with offspring individuals, and repeating the iterative evolution process until the fitness improvement values ​​of the offspring and parent individuals are lower than a set fitness threshold; and smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individuals after iterative evolution and using them as the drone's cruise path.

[0010] Furthermore, the system also includes: The pollution source location module is used to locate pollution concentration values ​​from historical pollution datasets. By geographical coordinates Spatial interpolation yields the concentration distribution function. ; Calculate the gradient field of the concentration distribution function The maximum point of the concentration distribution function is found using the gradient ascent method, which serves as the coordinate of the pollution source.

[0011] Furthermore, the system also includes: The fitness calculation module is used to calculate the fitness using the concentration distribution function. Spatial sampling is performed within the pixel coverage area to obtain the concentration values ​​of the sampling points. ,in For the first Coordinates of each sampling point; cruising altitude values ​​through each grid cell. Calculate ground pixel resolution ,in For sensor pixel size, The focal length of the lens; calculate the standard deviation of pollution concentration in the area covered by a unit pixel. ,in This refers to the number of sampling points within the coverage area of ​​a single pixel. Let K be the pollution concentration value at the kth sampling point. This represents the average pollution concentration value within the coverage area of ​​a single pixel. Establish the information entropy of a unit pixel ,in A preset stability constant is set when the pollution concentration values ​​of all sampling points within the unit pixel coverage area are equal. = 0.

[0012] Furthermore, the system also includes: The mesh connectivity check module is used to build an adjacency graph for mesh cells with a cruise state of 1 in each individual, and to connect the mesh cells... As grid nodes, when two grid nodes and satisfy and Edges are established between corresponding nodes; the adjacency graph is traversed using a depth-first search algorithm to count the number of connected components. ; When the number of connected components At that time, calculate the sum of the information entropy of each connected component. ,in For the first The set of grid cells contained in each connected component; retain the connected component with the largest sum of information entropy, and set the cruise state of the grid cells in other connected components to 0.

[0013] Furthermore, the system also includes: The path optimization generation module is used to extract all grid cells with a cruising state of 1 from the winning individuals after iterative evolution, and obtain the geographic coordinate center point and corresponding cruising altitude value of each grid cell. Calculate the pollution concentration gradient magnitude of neighboring grid cells surrounding the target grid cell. The partial derivatives pass through a radius of The concentration distribution within the sampling area was calculated; the distance between grid cells was calculated based on the coordinates of the center points of adjacent grid cells. ,in For adjacent grid cells, the minimum detectable pollution concentration difference is calculated. Divide by the distance between grid cells Obtain the gradient threshold ; When the pollution concentration gradient amplitude Exceeding the gradient threshold At that time, calculate the height adjustment amount. ,in The baseline cruising altitude value, For safety limit factors; the reference cruising altitude value With limited height adjustment The values ​​are added together to obtain the temporary altitude adjustment value; the temporary altitude adjustment value is then compared with the minimum safe flight altitude. The adjusted cruising altitude value is obtained by comparing and selecting the larger value. Arranged in order of geographic coordinates to form a sequence of waypoints containing the adjusted cruise altitude values; Calculate the adjusted cruise altitude values ​​of adjacent waypoints The absolute value of the difference between them will determine the drone's maximum climb speed. Distance between grid cells The product divided by the drone's flight speed Obtain highly smooth constraint values When the absolute value of the difference between the cruise altitude values ​​of adjacent grid cells after adjustment exceeds the altitude smoothing constraint value, at the adjacent path point that violates the constraint... Calculate the number of interpolation path points Insertion between adjacent path points through equally spaced linear interpolation An intermediate path point is selected, ensuring that the absolute value of the height difference between all adjacent path points does not exceed the height smoothness constraint value. The constraint check and path point insertion process is repeated until the entire path sequence meets the smoothness requirement. The final path point sequence is then arranged in chronological order to generate a drone cruise path containing geographic coordinates and cruise altitude values.

[0014] (3) Beneficial effects Compared with existing technologies, the beneficial effect of this invention is that it uses a hyperspectral camera to monitor water pollution at different types and concentrations, thereby improving spectral resolution and reducing data processing volume, and adjusting the drone's cruising altitude. Attached Figure Description

[0015] Figure 1 This is a flowchart of the pollution data analysis method based on deep learning according to Embodiment 1 of the present invention; Figure 2 This is a block diagram of the pollution data analysis system based on deep learning according to Embodiment 2 of the present invention; Figure 3 This is a schematic diagram illustrating the influence of the drone's cruising altitude on spectral data acquisition in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Before providing examples, it's necessary to describe the application scenario of this invention. In a pollution data analysis process, the target of the survey was a section of a tributary of the Yangtze River within the city's jurisdiction, with a main stream length of approximately 11 kilometers and a drainage area of ​​53.5 square kilometers. Equipped with a 400-1300nm hyperspectral camera, using a line-scan hyperspectral method, the drone's cruise altitude was fixed at 500 meters. At this altitude, the maximum range for a single drone sortie was 15 kilometers. Compared to traditional camera-captured RGB images for water body identification, the 400-1700nm wavelength range was divided into 162 bands, with a spectral resolution of 8nm. After looping, the relevant data was imported into ENVI software, and the HDR file acquired by the hyperspectral camera was opened to analyze the pixels of the drainage basin. The image formed using 162 band channels across the entire 11-kilometer drainage basin significantly increased the data processing volume exponentially compared to images obtained from the traditional RGB images using only 3 channels. We selected the reflectance of the observed water bodies in our corresponding area at different bands, plotted the corresponding spectral curves for analysis, and determined whether pollution was present. As distance from the pollution source increases, the spectral camera receives the reflection characteristics of the entire water body. The lower the proportion of wastewater, the less obvious the spectral characteristics of the polluted object become (because spectral imaging is a mixture of the spectral characteristics of each pixel). For example, when shooting from 500 meters down, the spectral characteristics of the water body reflected by a single pixel are not as obvious as those captured from 300 meters down. At a cruising altitude of 500 meters, the ground resolution (GSD), a core indicator of the hyperspectral camera in this test scenario, refers to the actual ground object size corresponding to a single pixel being 12 meters (the spectrum returned by a single pixel is the reflection characteristics of the mixed superposition of this 12m*12m area). At a cruising altitude of 300 meters, the accuracy of the actual ground object size should be around 7 meters. Figure 3 As shown, the observation area of ​​the water surface covered by a single pixel of a hyperspectral camera is used as an example. Different flight altitudes correspond to different observation areas. Lower cruising altitudes bring higher observation accuracy, but they also increase the amount of processing computation.

[0018] Example 1: As Figure 1 As shown, this embodiment provides a deep learning-based method for analyzing contamination data, the method comprising: Obtain historical pollution datasets for the target water area, including pollution type, pollution concentration values ​​recorded by geographic coordinates and timestamps, and corresponding drone cruising altitudes. Obtain pollution source coordinates from the pollution concentration values. Use a convolutional neural network to obtain a pollution diffusion model based on the pollution type, pollution source coordinates, and pollution concentration values ​​recorded by geographic coordinates and timestamps. Divide the target water area into grid cells. A drone cruise plan equipped with a hyperspectral camera is generated. The cruise plan consists of multiple individuals. Each individual includes the cruise status, cruise altitude value, and corresponding pollution concentration prediction value of each grid cell. The cruise status value is either 0 or 1. A cruise status value of 1 indicates that the drone passes through the corresponding grid cell. The cruise altitude value is randomly generated within the safe cruise altitude threshold of the drone. The pollution concentration prediction value is obtained by using a pollution diffusion model to obtain the pollution type, the geographical coordinates of the grid cell, the coordinates of the pollution source, and the pollution concentration of the pollution source. The cruise strategy is iteratively optimized, including calculating the information entropy of each hyperspectral camera unit pixel, selecting a preset number of winning individuals based on the information entropy using a roulette wheel selection method, cross-crossing and mutating the winning individuals and removing non-adjacent individuals from the grid cells to obtain offspring individuals, replacing the initial population with offspring individuals, and repeating the iterative evolution process until the fitness improvement values ​​of the offspring and parent individuals are lower than the set fitness threshold, and smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individuals after iterative evolution and using them as the UAV cruise path.

[0019] For example, a pollution monitoring experiment was conducted on an 11-kilometer section of the Yangtze River in a concentrated industrial area in Qingshan District, Wuhan. The target water area has a drainage area of ​​53.5 square kilometers. Historical pollution data for the past 24 months was obtained from the environmental protection department's database. The dataset contains three main types of pollution: heavy metal pollution, organic compound pollution, and ammonia nitrogen pollution. Historical data records show that at 10:30 AM on March 12, 2023, a chromium ion concentration of 0.15 mg / L was detected at a location of 114.3452°E, 30.6237°N, with the drone conducting the monitoring mission hovering at an altitude of 450 meters. At 2:20 PM on June 18 of the same year, a benzene series organic compound concentration of 0.28 mg / L was detected at a location of 114.3567°E, 30.6183°N, with the drone hovering at an altitude of 380 meters. At 9:45 AM on September 25th of the same year, an ammonia nitrogen concentration of 2.8 mg / L was detected at a location of 114.3398°E and 30.6291°N, with the drone's cruising altitude at 520 meters. Spatial analysis of the 682 effective pollution concentration sampling points identified the geographical coordinates of three main pollution sources: the steel plant's wastewater discharge outlet at 114.3420°E and 30.6210°N; the chemical industrial park's comprehensive sewage discharge outlet at 114.3580°E and 30.6150°N; and the city's wastewater treatment plant's effluent discharge outlet at 114.3380°E and 30.6310°N.

[0020] The extracted pollution type identifiers, pollution source geographic coordinates, and pollution concentration data recorded by time geographic coordinates were input into a convolutional neural network for training. A deep network structure containing 5 convolutional layers and 3 fully connected layers was constructed, with a learning rate of 0.001 and a batch size of 32. After 12,000 iterations of training, a pollution diffusion prediction model was obtained, achieving a prediction accuracy of 92.3% on the validation set. The 11-kilometer-long target water area was divided into 275 grid cells with a grid size of 200 meters by 200 meters, and each grid cell was assigned a unique number using latitude and longitude coordinates. An initial population of 60 individuals was generated for drone patrol schemes, with each individual representing a possible patrol strategy configuration. In the individual numbered 1, the patrol status of 45 grid cells from 45th to 89th was set to 1, indicating that the drone needed to fly over these areas for monitoring; the patrol status of the remaining 230 grid cells was set to 0, indicating that the drone would skip the patrol. For these 45 selected grid cells, cruise altitudes were set to 420 meters, 380 meters, 450 meters, 390 meters, and 480 meters, respectively. The pollution diffusion model predicted pollution concentrations of 0.12 mg / L, 0.08 mg / L, 0.25 mg / L, 0.06 mg / L, and 0.31 mg / L for these grid cells, respectively.

[0021] A genetic algorithm was executed on the 60 generated cruise path individuals. Specifically, the information entropy value per unit pixel of the hyperspectral camera at the selected grid cell location was calculated for each individual. Since the pollution concentration is spatially unevenly distributed, information entropy can effectively quantify the information acquisition value of each monitoring point. Based on the calculated information entropy value, a roulette wheel selection strategy was used to select 24 high-fitness winners from the 60 individuals. Crossover and mutation operations were performed on these 24 winners to generate 48 new offspring individuals, with the crossover probability set to 0.8 and the mutation probability set to 0.1. During the offspring generation process, invalid individuals whose selected grid cells were not spatially discontinuous were removed to ensure the continuity and executability of the cruise path. After 280 iterations, the genetic algorithm iteration process was terminated when the fitness improvement value of offspring individuals relative to their parents was less than 0.005 for 10 consecutive generations. From the final winning individual, 38 adjacent grid cells with a cruise status of 1 were extracted, including cells numbered 52, 53, 67, 68, 69, 82, 83, and 84. The geographic center coordinates and corresponding cruise altitude values ​​of these 38 grid cells were extracted and processed using a path smoothing algorithm to generate a drone cruise trajectory containing 51 waypoints. The total flight distance was 8.6 kilometers, and the cruise altitude was dynamically adjusted within the range of 350 to 500 meters based on the predicted pollution concentration distribution.

[0022] Furthermore, the method for obtaining the coordinates of the pollution source from the pollution concentration value includes: Pollution concentration values ​​from historical pollution datasets By geographical coordinates Spatial interpolation yields the concentration distribution function. ; Calculate the gradient field of the concentration distribution function The maximum point of the concentration distribution function is found using the gradient ascent method, which serves as the coordinate of the pollution source.

[0023] For example, based on the 682 valid pollution concentration sampling points obtained above, the pollution concentration values ​​recorded in the historical pollution dataset are spatially interpolated according to their corresponding geographic coordinates. The Kriging interpolation method is selected to spatially continuousize the irregularly distributed sampling point data, with the interpolation grid resolution set to 50 meters by 50 meters. Taking heavy metal pollution as an example, interpolation calculations are performed on 156 heavy metal pollution concentration data points from the 682 sampling points to generate a concentration distribution function covering the entire 11-kilometer river section. The interpolation results show that the heavy metal concentration in the area near 114.3420°E and 30.6210°N is 0.18 mg / L, the concentration at 114.3425°E and 30.6215°N is 0.22 mg / L, and the concentration at 114.3415°E and 30.6205°N is 0.14 mg / L. By spatially connecting the concentration values ​​of all interpolation grid points, a complete concentration distribution function is constructed.

[0024] The gradient field value of the concentration distribution function at each grid point is calculated. The central difference method is used to calculate the concentration gradient at each grid point. For the grid point located at 114.3420°E and 30.6210°N, the east-west gradient is calculated as follows: the concentration value 50 meters east of the point (0.22 mg / L) minus the concentration value 50 meters west of the point (0.14 mg / L), divided by a distance interval of 100 meters, yielding an east-west gradient of 0.0008 mg / L / m. The north-south gradient is calculated as follows: the concentration value 50 meters north of the point (0.25 mg / L) minus the concentration value 50 meters south of the point (0.16 mg / L), divided by a distance interval of 100 meters, yielding a north-south gradient of 0.0009 mg / L / m. The east-west gradient of 0.0008 and the north-south gradient of 0.0009 are combined to form the gradient vector for that grid point. The gradient vectors were calculated for each of the 4400 interpolation grid points covering the entire target water area to construct a complete concentration gradient field distribution.

[0025] The gradient ascent method was used to search for the maximum point of the concentration distribution function to determine the pollution source coordinates. 114.3400°E, 30.6200°N was chosen as the starting point, with a concentration of 0.08 mg / L. Moving 10 meters along the gradient direction, the search reached 114.3401°E, 30.6201°N, where the concentration increased to 0.09 mg / L. Continuing along the gradient direction, after 83 iterations, the search reached 114.3419°E, 30.6209°N, where the concentration reached 0.31 mg / L. Continuing the search along the gradient direction at this location, it was found that the concentration values ​​in the surrounding eight directions were all lower than at the current location, confirming this point as a local maximum. The same gradient ascent search process was performed on both organic compound pollution and ammonia nitrogen pollution, ultimately determining the coordinates of the three pollution sources to be 114.3419°E, 30.6209°N, 114.3579°E, 30.6151°N, and 114.3381°E, 30.6309°N. The obtained pollution source coordinates were compared with the actual locations of the discharge outlets, and the coordinate deviations were all within 15 meters, meeting the requirements for pollution source location accuracy.

[0026] Furthermore, the method for calculating the information entropy of each individual hyperspectral camera unit pixel includes: Through concentration distribution function Spatial sampling is performed within the pixel coverage area to obtain the concentration values ​​of the sampling points. ,in For the first Coordinates of each sampling point; cruising altitude values ​​through each grid cell. Calculate ground pixel resolution ,in For sensor pixel size, The focal length of the lens; calculate the standard deviation of pollution concentration in the area covered by a unit pixel. ,in This refers to the number of sampling points within the coverage area of ​​a single pixel. Let K be the pollution concentration value at the kth sampling point. This represents the average pollution concentration value within the coverage area of ​​a single pixel. Establish the information entropy of a unit pixel ,in A preset stability constant is set when the pollution concentration values ​​of all sampling points within the unit pixel coverage area are equal. = 0.

[0027] For example, based on the 60 individual cruise scenarios generated above and the determined pollution source coordinates, the hyperspectral camera unit pixel information entropy calculation process is performed for each individual. Taking grid cell number 52 in individual number 1 as an example, the center coordinates of this grid cell are 114.3456 degrees east longitude and 30.6198 degrees north latitude, corresponding to a cruise altitude of 420 meters. Spatial sampling is performed within the pixel coverage area according to the established concentration distribution function. Within a 200-meter by 200-meter area covered by the pixel, grid sampling is performed at 25-meter intervals, resulting in a total of 64 sampling points. The pollution concentration value of each sampling point is calculated using the concentration distribution function. The concentration value of the first sampling point is 0.12 mg / L, the concentration value of the second sampling point is 0.15 mg / L, the concentration value of the third sampling point is 0.08 mg / L, and so on, to obtain the complete concentration dataset of the 64 sampling points. The concentration values ​​of the sampling points range from 0.06 mg / L to 0.28 mg / L, and the average concentration of the sampling points is 0.145 mg / L.

[0028] Ground pixel resolution was calculated based on the cruising altitude of 420 meters for the grid cells. The hyperspectral camera sensor pixel size is 5.2 micrometers, and the lens focal length is 35 millimeters. Multiplying the cruising altitude of 420 meters by the sensor pixel size of 5.2 micrometers and then dividing by the lens focal length of 35 millimeters yields a ground pixel resolution of 0.624 meters. The ground area covered by a single pixel is 0.624 meters multiplied by 0.624 meters, which equals 0.389 square meters. Dense sampling was conducted within this pixel coverage area at 0.1-meter intervals, acquiring pollution concentration data from 39 valid sampling points. The concentration value at the first sampling point was 0.13 mg / L, the second at 0.11 mg / L, and the third at 0.16 mg / L. Data collection continued until all 39 sampling points were collected. The average concentration of the 39 sampling points was calculated to be 0.142 mg / L. Calculate the squared deviation of the concentration value at each sampling point from the average value. The squared deviation of the first sampling point is 0.13 minus 0.142, which equals 0.000144. The squared deviation of the second sampling point is 0.11 minus 0.142, which equals 0.001024. Calculate the squared deviations for all 39 sampling points. Sum the 39 squared deviations, divide by the number of sampling points (39), and take the square root to obtain the standard deviation of the pollution concentration, which is 0.0386 mg / L.

[0029] A formula for calculating the information entropy per unit pixel is established. Specifically, a stability constant of 0.01 mg / L is set. The standard deviation of the pollution concentration (0.0386) is divided by the sum of the standard deviation and the stability constant (0.0486), yielding the first coefficient of 0.794. The base-2 logarithm of the first coefficient 0.794 is calculated to be -0.331. The stability constant (0.01) is divided by the sum of the standard deviation and the stability constant (0.0486), yielding the second coefficient of 0.206. The base-2 logarithm of the second coefficient 0.206 is calculated to be -2.281. Multiplying the first coefficient 0.794 by its logarithm (-0.331) yields -0.263, and multiplying the second coefficient 0.206 by its logarithm (-2.281) yields -0.470. The negatives of the two products are added together to obtain the information entropy value of 0.733. For each of the 38 grid cells in the individual designated as "Cruise State 1" with the same information entropy calculation process, information entropy values ​​of 0.733, 0.542, 0.681, 0.595, 0.724, 0.458, and 0.612 were obtained. When all sampling points within the cell coverage area had completely equal pollution concentration values, the standard deviation of the pollution concentration was 0, and the information entropy value of that cell was directly set to 0. After completing the information entropy calculation for 60 individual cruise schemes, a fitness function was provided to select the winning scheme in the subsequent roulette wheel selection process.

[0030] Furthermore, the method for removing non-adjacent individuals from the grid cells includes: For each individual, an adjacency graph is created for the grid cells with a cruise state of 1, and the grid cells are... As grid nodes, when two grid nodes and satisfy and Edges are established between corresponding nodes; the adjacency graph is traversed using a depth-first search algorithm to count the number of connected components. ; When the number of connected components At that time, calculate the sum of the information entropy of each connected component. ,in For the first The set of grid cells contained in each connected component; retain the connected component with the largest sum of information entropy, and set the cruise state of the grid cells in other connected components to 0.

[0031] For example, based on the calculated entropy values ​​of the 60 individual cruise plans, a grid cell adjacency check and non-adjacent individual removal process are performed. Taking individual number 15 as an example, the grid cells with a cruise state of 1 in this individual include 12 grid cells: number 34, 35, 49, 50, 51, 78, 79, 94, 95, 127, 128, and 129. These 12 grid cells are used as nodes in graph theory to construct an adjacency graph. According to the grid partitioning rules, the row and column coordinates of grid cell number 34 are in the 3rd row and 4th column, and the row and column coordinates of grid cell number 35 are in the 3rd row and 5th column. The absolute value of the difference between their row coordinates is 0, and the absolute value of the difference between their column coordinates is 1, satisfying the adjacency condition that the difference between their row and column coordinates is less than or equal to 1. Therefore, a connecting edge is established between nodes number 34 and 35. Grid cell 49 is located in row 4, column 9, and grid cell 50 is located in row 4, column 10. The absolute difference between their row coordinates is 0, and the absolute difference between their column coordinates is 1, satisfying the adjacency condition and establishing a connecting edge. Grid cell 78 is located in row 6, column 8, and grid cell 94 is located in row 7, column 4. The absolute difference between their row coordinates is 1, and the absolute difference between their column coordinates is 4, not satisfying the adjacency condition, therefore no connecting edge is established.

[0032] Continuing to traverse the adjacency graph constructed using the depth-first search algorithm, the number of connected components is counted. Starting from node 34, the depth-first search is performed, visiting its neighbor node 35. Continuing the search, it is found that node 35 is adjacent to node 50, and node 50 is adjacent to nodes 49 and 51, completing the traversal of the first connected component, which contains 5 grid cells: nodes 34, 35, 49, 50, and 51. Continuing from the unvisited node 78, a new depth-first search is started, finding that node 78 is adjacent to node 79, and node 79 is adjacent to nodes 94 and 95, completing the traversal of the second connected component, which contains 4 grid cells: nodes 78, 79, 94, and 95. Starting the search from the remaining unvisited node 127, it was found that nodes 127, 128, and 129 are adjacent to each other, forming a third connected component containing 3 grid cells. The number of connected components is 3, which is greater than 1, so a connected component filtering operation needs to be performed.

[0033] The information entropy of each connected component is calculated and then filtered for retention. For example, the first connected component contains grid cells 34, 35, 49, 50, and 51, with corresponding information entropy values ​​of 0.642, 0.578, 0.695, 0.723, and 0.601, respectively, for a total information entropy of 3.239. The second connected component contains grid cells 78, 79, 94, and 95, with corresponding information entropy values ​​of 0.456, 0.498, 0.567, and 0.523, respectively, for a total information entropy of 2.044. The third connected component contains grid cells 127, 128, and 129, with corresponding information entropy values ​​of 0.385, 0.421, and 0.394, for a total information entropy of 1.200. Comparing the sum of the information entropy of the three connected components, the first connected component has the largest sum of information entropy (3.239), therefore, the cruise state of all grid cells in this connected component is retained as 1. The cruise states of grid cells 78, 79, 94, and 95 in the second connected component are changed to 0, and the cruise states of grid cells 127, 128, and 129 in the third connected component are changed to 0. After modification, individual number 15 retains only five adjacent grid cells (numbers 34, 35, 49, 50, and 51) with a cruise state of 1. The same adjacency check and elimination process is performed on each of the 60 cruise scheme individuals to ensure that all grid cells with a cruise state of 1 form a single connected region, satisfying the executability requirements of continuous UAV cruise paths.

[0034] Furthermore, the method for smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individual after iterative evolution and using them as the drone's cruise path includes: Extract all grid cells with a cruise state of 1 from the winning individuals after the complete iterative evolution, and obtain the geographic coordinate center point and corresponding cruise altitude value of each grid cell. Calculate the pollution concentration gradient magnitude of neighboring grid cells surrounding the target grid cell. The partial derivatives pass through a radius of The concentration distribution within the sampling area was calculated; the distance between grid cells was calculated based on the coordinates of the center points of adjacent grid cells. ,in For adjacent grid cells, the minimum detectable pollution concentration difference is... Divide by the distance between grid cells Obtain the gradient threshold ; When the pollution concentration gradient amplitude Exceeding the gradient threshold At that time, calculate the height adjustment amount. ,in The baseline cruising altitude value, For safety limit factors; the reference cruising altitude value With limited height adjustment The values ​​are added together to obtain the temporary altitude adjustment value; the temporary altitude adjustment value is then compared with the minimum safe flight altitude. The adjusted cruising altitude value is obtained by comparing and selecting the larger value. Arranged in order of geographic coordinates to form a sequence of waypoints containing the adjusted cruise altitude values; Calculate the adjusted cruise altitude values ​​of adjacent waypoints The absolute value of the difference between them will determine the drone's maximum climb speed. Distance between grid cells The product divided by the drone's flight speed Obtain highly smooth constraint values When the absolute value of the difference between the cruise altitude values ​​of adjacent grid cells after adjustment exceeds the altitude smoothing constraint value, at the adjacent path point that violates the constraint... Calculate the number of interpolation path points Insertion between adjacent path points through equally spaced linear interpolation An intermediate path point is selected, ensuring that the absolute value of the height difference between all adjacent path points does not exceed the height smoothness constraint value. The constraint check and path point insertion process is repeated until the entire path sequence meets the smoothness requirement. The final path point sequence is then arranged in chronological order to generate a drone cruise path containing geographic coordinates and cruise altitude values.

[0035] For example, based on the results of 280 iterations of optimization after the adjacency screening, the grid cells with a cruise state of 1 in the final winning individual are extracted and path smoothing is performed. For instance, the optimal individual contains 38 adjacent grid cells: cells 52, 53, 67, 68, 69, 82, 83, and 84. The geographic coordinate center point of grid cell 52 is extracted as 114.3445°E, 30.6215°N, corresponding to a cruise altitude of 380 meters. The center point of grid cell 53 is 114.3447°E, 30.6215°N, with a cruise altitude of 420 meters. The center point of grid cell 67 is 114.3445°E, 30.6200°N, with a cruise altitude of 450 meters. Calculate the pollution concentration gradient amplitude of adjacent grid cells surrounding grid cell 52. Extract concentration distribution data within a sampling area with a radius of 400 meters around the center point of this grid cell. The concentration value at 200 meters east is 0.18 mg / L, and at 200 meters west is 0.12 mg / L. The partial derivative in the east-west direction is 0.18 minus 0.12, divided by the 400-meter distance, yielding 0.00015 mg / L / m. The concentration value at 200 meters north is 0.22 mg / L, and at 200 meters south is 0.14 mg / L. The partial derivative in the north-south direction is 0.22 minus 0.14, divided by the 400-meter distance, yielding 0.0002 mg / L / m. Calculate the gradient amplitude as the square root of the square of the partial derivative in the east-west direction (0.00015) plus the square of the partial derivative in the north-south direction (0.0002), resulting in 0.00025 mg / L / m.

[0036] The gradient threshold is calculated based on the distance between adjacent grid cells to determine whether the cruising altitude needs adjustment. For example, the distance between the center coordinates of grid cell 52 and grid cell 53 is calculated as follows: (114.3447°E - 114.3445°E) + (30.6215°N - 30.6215°N) + square root of the result, multiplied by the Earth's radius, yielding a distance of 200 meters between grid cells. The minimum detectable pollution concentration difference of 0.01 mg / L is divided by the grid cell distance of 200 meters, resulting in a gradient threshold of 0.00005 mg / L / m. The pollution concentration gradient amplitude of grid cell 52, at 0.00025 mg / L / m, exceeds the gradient threshold of 0.00005 mg / L / m, requiring altitude adjustment. The baseline cruising altitude is set to 400 meters, with a safety limit factor of 0.3. The altitude adjustment is calculated as follows: 400 meters multiplied by 1, minus 0.00005 divided by 0.00005 plus 0.00025, then multiplied by 0.3, yielding an altitude adjustment of 80 meters. The baseline cruise altitude of 400 meters is added to this adjustment, resulting in a temporary altitude adjustment of 480 meters. A minimum safe flight altitude of 300 meters is set. The temporary altitude adjustment of 480 meters is compared to the minimum safe flight altitude of 300 meters, and the larger value is used to determine the adjusted cruise altitude as 480 meters.

[0037] Arrange 38 grid cells in geographic coordinate order to form a pathpoint sequence. Perform an altitude smoothing constraint check. For example, after adjustment, the cruising altitude of grid cell 52 is 480 meters, and after adjustment, the cruising altitude of grid cell 53 is 420 meters. The absolute value of the altitude difference between adjacent pathpoints is 60 meters. Set the UAV's maximum climb speed to 3 meters per second, the distance between grid cells to 200 meters, and the UAV's flight speed to 15 meters per second. Calculate the altitude smoothing constraint value as 3 meters per second multiplied by 200 meters and then divided by 15 meters per second, resulting in 40 meters. The absolute value of the altitude difference between grid cells 52 and 53, 60 meters, exceeds the altitude smoothing constraint value of 40 meters, requiring the insertion of an intermediate pathpoint. Calculate the number of interpolated pathpoints as 60 meters divided by 40 meters, rounded up, and then subtracted by 1, resulting in the need to insert one intermediate pathpoint. Between grid cell 52 (coordinates 114.3445°E, 30.6215°N, altitude 480m) and grid cell 53 (coordinates 114.3447°E, 30.6215°N, altitude 420m), an intermediate pathpoint (coordinates 114.3446°E, 30.6215°N, altitude 450m) is inserted using equally spaced linear interpolation. The same constraint checks and pathpoint insertion process are performed on all 38 grid cells, ultimately generating a complete cruise path sequence containing 67 waypoints. The cruise altitude smoothly varies between 350m and 500m, ensuring the UAV can safely and continuously execute the entire pollution monitoring mission.

[0038] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a pollution data analysis system based on deep learning, the system comprising: The pollution diffusion analysis module is used to acquire historical pollution datasets of the target water area, including pollution type, pollution concentration values ​​recorded by geographic coordinates and timestamps, and the corresponding drone cruising altitude. The pollution source coordinates are obtained through the pollution concentration values. The pollution type, pollution source coordinates, and pollution concentration values ​​recorded by geographic coordinates and timestamps are used to obtain a pollution diffusion model through a convolutional neural network. The target water area is divided into grid cells. The data initialization module is used to generate a drone cruise plan equipped with a hyperspectral camera. The cruise plan consists of multiple individuals, each of which includes the cruise status, cruise altitude value, and corresponding pollution concentration prediction value of each grid cell. The cruise status value is either 0 or 1, with a cruise status value of 1 indicating that the drone passes through the corresponding grid cell. The cruise altitude value is randomly generated within the drone's safe cruise altitude threshold. The pollution concentration prediction value is obtained by using a pollution diffusion model based on the pollution type, the geographic coordinates of the grid cell, the coordinates of the pollution source, and the pollution concentration of the pollution source. The data analysis module is used to iteratively optimize the cruise plan. This includes calculating the information entropy of each hyperspectral camera unit pixel, selecting a preset number of winning individuals based on the information entropy using a roulette wheel selection method, cross-crossing and mutating the winning individuals, and removing non-adjacent individuals from the grid cells to obtain offspring individuals; replacing the initial population with offspring individuals, and repeating the iterative evolution process until the fitness improvement values ​​of the offspring and parent individuals are lower than a set fitness threshold; and smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individuals after iterative evolution and using them as the drone's cruise path.

[0039] Furthermore, the system also includes: The pollution source location module is used to locate pollution concentration values ​​from historical pollution datasets. By geographical coordinates Spatial interpolation yields the concentration distribution function. ; Calculate the gradient field of the concentration distribution function The maximum point of the concentration distribution function is found using the gradient ascent method, which serves as the coordinate of the pollution source.

[0040] Furthermore, the system also includes: The fitness calculation module is used to calculate the fitness using the concentration distribution function. Spatial sampling is performed within the pixel coverage area to obtain the concentration values ​​of the sampling points. ,in For the first Coordinates of each sampling point; cruising altitude values ​​through each grid cell. Calculate ground pixel resolution ,in For sensor pixel size, The focal length of the lens; calculate the standard deviation of pollution concentration in the area covered by a unit pixel. ,in This refers to the number of sampling points within the coverage area of ​​a single pixel. Let K be the pollution concentration value at the kth sampling point. This represents the average pollution concentration value within the coverage area of ​​a single pixel. Establish the information entropy of a unit pixel ,in A preset stability constant is set when the pollution concentration values ​​of all sampling points within the unit pixel coverage area are equal. = 0.

[0041] Furthermore, the system also includes: The mesh connectivity check module is used to build an adjacency graph for mesh cells with a cruise state of 1 in each individual, and to connect the mesh cells... As grid nodes, when two grid nodes and satisfy and Edges are established between corresponding nodes; the adjacency graph is traversed using a depth-first search algorithm to count the number of connected components. ; When the number of connected components At that time, calculate the sum of the information entropy of each connected component. ,in For the first The set of grid cells contained in each connected component; retain the connected component with the largest sum of information entropy, and set the cruise state of the grid cells in other connected components to 0.

[0042] Furthermore, the system also includes: The path optimization generation module is used to extract all grid cells with a cruising state of 1 from the winning individuals after iterative evolution, and obtain the geographic coordinate center point and corresponding cruising altitude value of each grid cell. Calculate the pollution concentration gradient magnitude of neighboring grid cells surrounding the target grid cell. The partial derivatives pass through a radius of The concentration distribution within the sampling area was calculated; the distance between grid cells was calculated based on the coordinates of the center points of adjacent grid cells. ,in For adjacent grid cells, the minimum detectable pollution concentration difference is calculated. Divide by the distance between grid cells Obtain the gradient threshold ; When the pollution concentration gradient amplitude Exceeding the gradient threshold At that time, calculate the height adjustment amount. ,in The baseline cruising altitude value, For safety limit factors; the reference cruising altitude value With limited height adjustment The values ​​are added together to obtain the temporary altitude adjustment value; the temporary altitude adjustment value is then compared with the minimum safe flight altitude. The adjusted cruising altitude value is obtained by comparing and selecting the larger value. Arranged in order of geographic coordinates to form a sequence of waypoints containing the adjusted cruise altitude values; Calculate the adjusted cruise altitude values ​​of adjacent waypoints The absolute value of the difference between them will determine the drone's maximum climb speed. Distance between grid cells The product divided by the drone's flight speed Obtain highly smooth constraint values When the absolute value of the difference between the cruise altitude values ​​of adjacent grid cells after adjustment exceeds the altitude smoothing constraint value, at the adjacent path point that violates the constraint... Calculate the number of interpolation path points Insertion between adjacent path points through equally spaced linear interpolation An intermediate path point is selected, ensuring that the absolute value of the height difference between all adjacent path points does not exceed the height smoothness constraint value. The constraint check and path point insertion process is repeated until the entire path sequence meets the smoothness requirement. The final path point sequence is then arranged in chronological order to generate a drone cruise path containing geographic coordinates and cruise altitude values.

[0043] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0044] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for analyzing contaminated data, characterized in that, The method includes: Obtain historical pollution datasets for the target water area, including pollution type, pollution concentration values ​​recorded by geographic coordinates and timestamps, and corresponding drone cruising altitudes. Obtain pollution source coordinates from the pollution concentration values. Use a convolutional neural network to obtain a pollution diffusion model based on the pollution type, pollution source coordinates, and pollution concentration values ​​recorded by geographic coordinates and timestamps. Divide the target water area into grid cells. A drone cruise plan equipped with a hyperspectral camera is generated. The cruise plan consists of multiple individuals. Each individual includes the cruise status, cruise altitude value, and corresponding pollution concentration prediction value of each grid cell. The cruise status value is either 0 or 1. A cruise status value of 1 indicates that the drone passes through the corresponding grid cell. The cruise altitude value is randomly generated within the safe cruise altitude threshold of the drone. The pollution concentration prediction value is obtained by using a pollution diffusion model to obtain the pollution type, the geographical coordinates of the grid cell, the coordinates of the pollution source, and the pollution concentration of the pollution source. The cruise strategy is iteratively optimized, including calculating the information entropy of each hyperspectral camera unit pixel, selecting a preset number of winning individuals based on the information entropy using a roulette wheel selection method, cross-crossing and mutating the winning individuals and removing non-adjacent individuals from the grid cells to obtain offspring individuals, replacing the initial population with offspring individuals, and repeating the iterative evolution process until the fitness improvement values ​​of the offspring and parent individuals are lower than the set fitness threshold, and smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individuals after iterative evolution and using them as the UAV cruise path.

2. The deep learning-based contamination data analysis method according to claim 1, characterized in that, The method for obtaining the coordinates of the pollution source from the pollution concentration value includes: Pollution concentration values ​​from historical pollution datasets By geographic coordinates Spatial interpolation yields the concentration distribution function. ; Calculate the gradient field of the concentration distribution function The maximum point of the concentration distribution function is found using the gradient ascent method, which serves as the coordinate of the pollution source.

3. The deep learning-based contamination data analysis method according to claim 2, characterized in that, The method for calculating the information entropy of a unit pixel in the hyperspectral camera for each individual includes: Through concentration distribution function Spatial sampling is performed within the pixel coverage area to obtain the concentration values ​​of the sampling points. ,in For the first Coordinates of each sampling point; cruising altitude values ​​through each grid cell. Calculate ground pixel resolution ,in For sensor pixel size, The focal length of the lens; calculate the standard deviation of pollution concentration in the area covered by a unit pixel. ,in This refers to the number of sampling points within the coverage area of ​​a single pixel. Let K be the pollution concentration value at the kth sampling point. This represents the average pollution concentration value within the coverage area of ​​a single pixel. Establish the information entropy of a unit pixel ,in A preset stability constant is set when the pollution concentration values ​​of all sampling points within the unit pixel coverage area are equal. = 0.

4. The deep learning-based contamination data analysis method according to claim 3, characterized in that, The method for removing non-adjacent individuals from the grid cells includes: For each individual, an adjacency graph is created for the grid cells with a cruise state of 1, and the grid cells are... As grid nodes, when two grid nodes and satisfy and Edges are established between corresponding nodes; the adjacency graph is traversed using a depth-first search algorithm to count the number of connected components. ; When the number of connected components At that time, calculate the sum of the information entropy of each connected component. ,in For the first The set of grid cells contained in each connected component; retain the connected component with the largest sum of information entropy, and set the cruise state of the grid cells in other connected components to 0.

5. The deep learning-based contamination data analysis method according to claim 4, characterized in that, The method for smoothing the center point coordinates and cruise altitude values ​​of the geographic coordinates of the grid cells with a cruise state of 1 after iterative evolution and using them as the drone's cruise path includes: Extract all grid cells with a cruise state of 1 from the winning individuals after the complete iterative evolution, and obtain the geographic coordinate center point and corresponding cruise altitude value of each grid cell. Calculate the pollution concentration gradient magnitude of neighboring grid cells surrounding the target grid cell. The partial derivatives pass through a radius of The concentration distribution within the sampling area was calculated; the distance between grid cells was calculated based on the coordinates of the center points of adjacent grid cells. ,in For adjacent grid cells, the minimum detectable pollution concentration difference is calculated. Divide by the distance between grid cells Obtain the gradient threshold ; When the pollution concentration gradient amplitude Exceeding the gradient threshold At that time, calculate the height adjustment amount. ,in The baseline cruising altitude value, For safety limit factors; the reference cruising altitude value With limited height adjustment The values ​​are added together to obtain the temporary altitude adjustment value; the temporary altitude adjustment value is then compared with the minimum safe flight altitude. The adjusted cruising altitude value is obtained by comparing and selecting the larger value. Arranged in order of geographic coordinates to form a sequence of waypoints containing the adjusted cruise altitude values; Calculate the adjusted cruise altitude values ​​of adjacent waypoints The absolute value of the difference between them will determine the drone's maximum climb speed. Distance between grid cells The product divided by the drone's flight speed Obtain highly smooth constraint values When the absolute value of the difference between the cruise altitude values ​​of adjacent grid cells after adjustment exceeds the altitude smoothing constraint value, at the adjacent path point that violates the constraint... Calculate the number of interpolation path points Insertion between adjacent path points through equally spaced linear interpolation An intermediate path point is selected, ensuring that the absolute value of the height difference between all adjacent path points does not exceed the height smoothness constraint value. The constraint check and path point insertion process is repeated until the entire path sequence meets the smoothness requirement. The final path point sequence is then arranged in chronological order to generate a drone cruise path containing geographic coordinates and cruise altitude values.

6. A pollution data analysis system based on deep learning, characterized in that, The system includes: The pollution diffusion analysis module is used to acquire historical pollution datasets of the target water area, including pollution type, pollution concentration values ​​recorded by geographic coordinates and timestamps, and the corresponding drone cruising altitude. The pollution source coordinates are obtained through the pollution concentration values. The pollution type, pollution source coordinates, and pollution concentration values ​​recorded by geographic coordinates and timestamps are used to obtain a pollution diffusion model through a convolutional neural network. The target water area is divided into grid cells. The data initialization module is used to generate a drone cruise plan equipped with a hyperspectral camera. The cruise plan consists of multiple individuals, each of which includes the cruise status, cruise altitude value, and corresponding pollution concentration prediction value of each grid cell. The cruise status value is either 0 or 1, with a cruise status value of 1 indicating that the drone passes through the corresponding grid cell. The cruise altitude value is randomly generated within the drone's safe cruise altitude threshold. The pollution concentration prediction value is obtained by using a pollution diffusion model based on the pollution type, the geographic coordinates of the grid cell, the coordinates of the pollution source, and the pollution concentration of the pollution source. The data analysis module is used to iteratively optimize the cruise plan. This includes calculating the information entropy of each hyperspectral camera unit pixel, selecting a preset number of winning individuals based on the information entropy using a roulette wheel selection method, cross-crossing and mutating the winning individuals, and removing non-adjacent individuals from the grid cells to obtain offspring individuals; replacing the initial population with offspring individuals, and repeating the iterative evolution process until the fitness improvement values ​​of the offspring and parent individuals are lower than a set fitness threshold; and smoothing the center point coordinates and cruise altitude values ​​of the grid cells with a cruise state of 1 for the winning individuals after iterative evolution and using them as the drone's cruise path.

7. The deep learning-based contamination data analysis system according to claim 6, characterized in that, The system also includes: The pollution source location module is used to locate pollution concentration values ​​from historical pollution datasets. By geographic coordinates Spatial interpolation yields the concentration distribution function. ; Calculate the gradient field of the concentration distribution function The maximum point of the concentration distribution function is found using the gradient ascent method, which serves as the coordinate of the pollution source.

8. The deep learning-based contamination data analysis system according to claim 7, characterized in that, The system also includes: The fitness calculation module is used to calculate the fitness using the concentration distribution function. Spatial sampling is performed within the pixel coverage area to obtain the concentration values ​​of the sampling points. ,in For the first Coordinates of each sampling point; cruising altitude values ​​through each grid cell. Calculate ground pixel resolution ,in For sensor pixel size, The focal length of the lens; calculate the standard deviation of pollution concentration in the area covered by a unit pixel. ,in This refers to the number of sampling points within the coverage area of ​​a single pixel. Let K be the pollution concentration value at the kth sampling point. This represents the average pollution concentration value within the coverage area of ​​a single pixel. Establish the information entropy of a unit pixel ,in A preset stability constant is set when the pollution concentration values ​​of all sampling points within the unit pixel coverage area are equal. = 0.

9. The pollution data analysis system based on deep learning according to claim 8, characterized in that, The system also includes: The mesh connectivity check module is used to build an adjacency graph for mesh cells with a cruise state of 1 in each individual, and to connect the mesh cells... As grid nodes, when two grid nodes and satisfy and Edges are established between corresponding nodes; the adjacency graph is traversed using a depth-first search algorithm to count the number of connected components. ; When the number of connected components At that time, calculate the sum of the information entropy of each connected component. ,in For the first The set of grid cells contained in each connected component; retain the connected component with the largest sum of information entropy, and set the cruise state of the grid cells in other connected components to 0.

10. The deep learning-based contamination data analysis system according to claim 9, characterized in that, The system also includes: The path optimization generation module is used to extract all grid cells with a cruising state of 1 from the winning individuals after iterative evolution, and obtain the geographic coordinate center point and corresponding cruising altitude value of each grid cell. Calculate the pollution concentration gradient magnitude of neighboring grid cells surrounding the target grid cell. The partial derivatives pass through a radius of The concentration distribution within the sampling area was calculated; the distance between grid cells was calculated based on the coordinates of the center points of adjacent grid cells. ,in For adjacent grid cells, the minimum detectable pollution concentration difference is calculated. Divide by the distance between grid cells Obtain the gradient threshold ; When the pollution concentration gradient amplitude Exceeding the gradient threshold At that time, calculate the height adjustment amount. ,in The baseline cruising altitude value, For safety limit factors; the reference cruising altitude value With limited height adjustment The values ​​are added together to obtain the temporary altitude adjustment value; the temporary altitude adjustment value is then compared with the minimum safe flight altitude. The adjusted cruising altitude value is obtained by comparing and selecting the larger value. Arranged in order of geographic coordinates to form a sequence of waypoints containing the adjusted cruise altitude values; Calculate the adjusted cruise altitude values ​​of adjacent waypoints The absolute value of the difference between them will determine the drone's maximum climb speed. Distance between grid cells The product divided by the drone's flight speed Obtain highly smooth constraint values When the absolute value of the difference between the cruise altitude values ​​of adjacent grid cells after adjustment exceeds the altitude smoothing constraint value, at the adjacent path point that violates the constraint... Calculate the number of interpolation path points Insertion between adjacent path points through equally spaced linear interpolation An intermediate path point is selected, ensuring that the absolute value of the height difference between all adjacent path points does not exceed the height smoothness constraint value. The constraint check and path point insertion process is repeated until the entire path sequence meets the smoothness requirement. The final path point sequence is then arranged in chronological order to generate a drone cruise path containing geographic coordinates and cruise altitude values.

Images