High-resolution remote sensing suspected polluted water body classification and identification method

Abstract

The invention relates to a high-resolution remote sensing suspected polluted water body classification and recognition method, which comprises the following steps of: cutting surface reflectance remote sensing images in sequence by utilizing water body vector data to obtain a high-resolution remote sensing image of each water body vector data, so that a suspected polluted water body is not found from the high-resolution remote sensing images by utilizing deep learning any more; instead, a target identification problem is converted into a problem of performing water body type identification from a high-resolution remote sensing image of each water body so as to solve the problems of large data volume, high cost and a large amount of useless work done when a model continuously searches and calculates in a water-body-free region in a target identification method. According to the method, the polluted water bodies are divided into the duckweed type, the black and odorous type and the garbage type according to the existing experience, most polluted water body types are covered, and the comprehensiveness of polluted water body screening through high-resolution remote sensing is ensured.

Description

Technical Field

[0001] The invention relates to a water quality classification and identification method, in particular to a suspected polluted water body classification and identification method using high-resolution remote sensing. Background Art

[0002] With the acceleration of urbanization, the problem of polluted water bodies has become increasingly serious, causing serious impacts on the ecological environment and the quality of life of residents in urban and rural areas. It is of great significance to investigate and control them. As an efficient, rapid and large-scale water monitoring method, remote sensing technology has unique advantages in identifying polluted water bodies. At present, with the further improvement of the environment of large rivers and urban water bodies, large-scale polluted water bodies have been basically eliminated, while a large number of smaller and more hidden water bodies are still polluted. Optical remote sensing images with sub-meter high spatial resolution can be used to identify and find the location of polluted water bodies.

[0003] According to the current research status at home and abroad, there are three methods for identifying polluted water bodies based on high-resolution remote sensing images: The first is to conduct manual identification and interpretation based on expert basic knowledge. This method requires solid and rich experience, is difficult to popularize, and has limited manpower, making it difficult to identify polluted water bodies in large areas. The second is to use the characteristic that the band ratio of polluted water bodies on remote sensing images is different from that of normal water bodies for identification. This method is convenient and fast, but there is a fuzzy area in the calculation results and the accuracy is low. The third is to use deep learning to identify targets based on the characteristics of polluted water bodies on remote sensing images to find polluted water bodies from remote sensing images. This method is an algorithm based on image recognition, which has a high recognition accuracy, but has high requirements for machine configuration and takes a long time. The time cost and equipment cost required for large-scale investigation of polluted water bodies are high.

[0004] When using deep learning to identify suspected polluted water bodies, the high resolution of high-resolution remote sensing images results in a large amount of data, and the processing process requires high time and equipment costs. In addition, there are many types of polluted water bodies, and target recognition can only solve the problem of identifying a single type of polluted water body. Summary of the invention

[0005] The purpose of the present invention is to provide a method for classifying and identifying suspected polluted water bodies using high-resolution remote sensing, so as to solve the problems of large data processing volume and high cost in identifying suspected polluted water bodies using deep learning.

[0006] The object of the present invention is achieved in that: A method for classifying and identifying suspected polluted water bodies using high-resolution remote sensing includes the following steps: S1. Acquire high-resolution remote sensing image data of the target area, pre-process the acquired high-resolution remote sensing image data to obtain image fusion data, correct the image fusion data, and form a surface reflectance remote sensing image of the target area.

[0007] S2. Use a water body recognition algorithm to identify the water body vector range of the suspected polluted water body from the surface reflectance remote sensing image, and then merge it with the water body vector data of the suspected polluted water body collected from the target area to form water body vector merged data; buffer the water body vector merged data according to the specified distance to generate a buffer zone, and obtain buffer zone vector data containing the surrounding environment of the suspected polluted water body.

[0008] S3. Use the buffer vector data to crop the surface reflectance remote sensing image to obtain the water body image raster data, and make a test set based on it.

[0009] S4. With reference to the water body vector data of suspected polluted water bodies collected from the target area and the public opinion data on water pollution, manual remote sensing interpretation is performed on various typical types of suspected polluted water bodies on the surface reflectance remote sensing images, and field verification is carried out simultaneously to determine whether the manual remote sensing interpretation results belong to polluted water bodies; the polluted water body locations with correct interpretation are sorted to form a polluted water body sample set.

[0010] S5. Based on the surface reflectance remote sensing image, the image slices and mask slices required for deep learning are produced according to the polluted water sample set, and the produced image slices and mask slices are randomly divided into a prediction set and a verification set in proportion.

[0011] S6. Based on the Pytorch framework, use the prediction set to train two or more deep learning models separately; use the intersection over union (IoU) and F1-score indicators to test the accuracy of each deep learning model; use multi-category cross entropy as the loss function, set the learning rate and optimizer parameters; use the WANDB online model training visualization tool to track and record the parameters and indicators of each deep learning model in operation, and iterate the parameters of each deep learning model according to the recorded indicators to improve the accuracy of each deep learning model. Finally, select the deep learning model with the highest accuracy as the deep learning model for suspected polluted water classification.

[0012] S7. Use the deep learning model for suspected polluted water body classification to automatically classify the test set by computer to obtain the classification results of suspected polluted water bodies; manually screen the classification results of suspected polluted water bodies, eliminate non-water body images and misclassified images, and compile a list of suspected polluted water bodies in the target area. After field verification, obtain the remote sensing monitoring results of polluted water bodies in the target area.

[0013] Furthermore, the preprocessing of the high-resolution remote sensing image data in step S1 includes radiation correction, atmospheric correction, orthorectification and image fusion.

[0014] Furthermore, step S2 includes the following sub-steps: S2-1 obtains the pbf vector data of the target area, retrieves various water body vectors including water, reservoir and wetland, and saves them as various water body vector files; merges various water body vector files to form the water body vector data of suspected polluted water bodies collected from the target area.

[0015] S2-2 sets the threshold for distinguishing water bodies from non-water bodies, uses the water body recognition algorithm to extract water body rasters from the surface reflectance remote sensing image, and converts the extracted water body rasters into water body vectors; the obtained water body vectors are merged with the water body vectors of suspected polluted water bodies collected from the target area to form water body vector merged data.

[0016] S2-3 buffers the water body vector merge data according to the specified distance, generates a buffer zone, and obtains buffer zone vector data including the surrounding environment of the water body.

[0017] Furthermore, the preparation of the polluted water sample set in step S4 includes the following sub-steps: S4-1 obtains recent water pollution information in the target area, finds the location of the polluted water body from the surface reflectance remote sensing image based on the location of the water pollution, and analyzes and determines whether the water body in the surface reflectance remote sensing image is still polluted. If the judgment result is "yes", the point of the water body is recorded.

[0018] S4-2 sets the water pollution threshold, uses the band ratio method to calculate the surface reflectance remote sensing image, obtains the pollution area of ​​the suspected polluted water body, and records the location of the suspected polluted water body.

[0019] S4-3 analyzes and summarizes the characteristics of the suspected polluted water bodies found through the above two sub-steps, and based on this, continues to search for similar suspected polluted water bodies from the surface reflectance remote sensing images.

[0020] S4-4 conducts field verification based on the recorded location information of suspected polluted water bodies, and determines whether the water body at the verification point is polluted through manual observation and instrument detection. If the judgment result is "yes", the location of the water body is recorded, and the type of pollution is also recorded.

[0021] S4-5 sorts out the correctly interpreted polluted water points based on the field verification results, and searches for other suspected polluted water bodies that are highly similar to the correctly interpreted points on the surface reflectance remote sensing images, expands the number of samples, and forms a polluted water sample set.

[0022] The present invention extracts the image grid data of each water body in the target area, classifies suspected polluted water bodies, and transforms the target identification problem of suspected polluted water bodies into a classification problem of water bodies. According to the pollution source and pollution cause, the suspected polluted water bodies are subdivided into three categories, plus normal water bodies, a total of four environmental types of water body classification and identification, thereby greatly improving the identification efficiency of suspected polluted water bodies. The present invention classifies and identifies suspected polluted water bodies in combination with the causes of pollution, which can more comprehensively understand and grasp the polluted water bodies in the target area.

[0023] The present invention uses water body vector data to crop the surface reflectivity remote sensing image in sequence to obtain high-resolution remote sensing images of each water body vector data. Therefore, deep learning will no longer be used, but suspected polluted water bodies will be directly found from high-resolution remote sensing images, and the target recognition problem will be converted into the problem of water body type recognition from the high-resolution remote sensing images of each water body. The amount of high-resolution remote sensing image data of each water body cropped is very small, and the amount of non-water body information in the image is small, thereby solving the problems of large data volume, high processing cost, and useless work done by the model in continuous search and calculation in areas without water bodies in traditional target recognition methods. According to existing experience, the present invention divides polluted water bodies into three types: duckweed, black and smelly, and garbage, covering most types of polluted water bodies, ensuring the comprehensiveness of polluted water body screening using high-resolution remote sensing image technology.

[0024] This invention breaks through the technical ideas of traditional remote sensing target identification, transforms the target identification problem into a classification problem, adopts deep learning technology, and constructs remote sensing interpretation signs of different types of polluted water bodies according to the characteristics of different types of polluted water bodies in surface reflectivity remote sensing images. It also establishes machine automatic recognition models for different types of polluted water bodies, achieves the purpose of rapid investigation and identification of polluted water bodies in a large area, and lays a technical foundation for the classified management of polluted water bodies. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flowchart of the identification method of the present invention. DETAILED DESCRIPTION

[0026] Explanation of terms: 1. The “high resolution” mentioned in the present invention refers to high spatial resolution.

[0027] The four types of water bodies are: 1. Normal water body: refers to the water body that appears blue-green in color in high-resolution remote sensing images.

[0028] 2. Duckweed water bodies: refers to water bodies that appear emerald green in color in high-resolution remote sensing images. Due to eutrophication, a large amount of emerald green duckweed floats on the surface of this type of water body.

[0029] 3. Black and smelly water bodies: refers to water bodies that appear gray or black in high-resolution remote sensing images. This type of water body is mostly caused by industrial wastewater discharge or livestock and poultry manure discharge.

[0030] 4. Garbage-type water bodies: refers to water bodies that appear as small grayish-white spots on high-resolution remote sensing images. Garbage is piled up on the banks or on the surface of such water bodies, and the environmental conditions are poor.

[0031] The present invention will be further described below in conjunction with the accompanying drawings.

[0032] like Figure 1 As shown, the suspected polluted water body classification and identification method of the present invention using high-resolution remote sensing includes the following steps: Step 1: Retrieve sub-meter high-resolution remote sensing image data according to the target area and time on the "Land Observation Satellite Data Service" website. Select GF7-BWD (0.8m panchromatic and 3.2m multispectral) and GF2-PMS (1m panchromatic and 4m multispectral) as satellite sensors. Select clear high-resolution remote sensing image data with less than 10% cloud cover. After the high-resolution remote sensing image data is collected, use Python's gdal library to write batch preprocessing code for high-resolution remote sensing image data, and perform data preprocessing such as radiation correction, atmospheric correction, orthorectification, and image fusion to obtain image fusion data. Use the "Register Raster" tool in ArcGIS to perform geometric precision correction on the image fusion data to form a surface reflectance remote sensing image of the target area. Use the Watchdog library to monitor the data processing process. When the surface reflectance remote sensing image in tif format is monitored, delete the remote sensing data compression package to ensure that the computer has sufficient space to receive and process the data in a rolling manner.

[0033] The preprocessing method of surface reflectance remote sensing image is: (1) Use Python's tarfile library to batch decompress high-resolution remote sensing image data compression packages in the tar.gz format. Since the data names of different satellites and sensors are different, it is necessary to parse the image data file names and extract parameters such as the satellite name and sensor name. For different data sources, set the corresponding tiff image file path, xml metadata file path, rpb geometric correction file path, and image resolution for its multispectral data and panchromatic data.

[0034] (2) Obtain the absolute radiation calibration coefficient file of the land observation satellite from the China Resources Satellite Application Center. According to the satellite type, sensor type and shooting time of the image data, obtain the corresponding gain value and offset value from the calibration coefficient file, and perform radiation correction calculation on the multispectral data of four bands and the panchromatic data of one band. The formula for radiation correction calculation is: L = Gain DN + Offset Among them, DN is the observed digital value of the remote sensing image, L is the equivalent radiation brightness at the entrance pupil of the satellite payload channel, and the unit is W m-2 sr-1 μm-1, Gain is the calibration coefficient gain value, and Bias is the calibration coefficient offset value.

[0035] (3) From the XML metadata file of the image data, obtain parameters such as the latitude and longitude of the image data center, aerosol type, underlying surface type, 550nm aerosol optical depth, work area altitude, and satellite sensor orbit altitude, and use the 6S model to carry out atmospheric correction.

[0036] (4) Use the gdal library in Python to orthorectify the multispectral data and panchromatic data after atmospheric correction, input the workspace DEM file, and output the orthorectified panchromatic band raster file and multispectral raster file.

[0037] (5) Use the gdal library in Python to perform image fusion on the results of the previous step and increase the resolution of the multispectral raster file to the full-color band resolution.

[0038] (6) The “Register Raster” tool in ArcGIS is written as a callable function, which uses the star map earth layer file as the base map to automatically and precisely correct the image fusion data to the base map to form a surface reflectance remote sensing image.

[0039] Step 2: Use the water body recognition algorithm to identify the latest water body vector range of the suspected polluted water body from the surface reflectance remote sensing image of the target area, and then merge it with the water body vector data of the suspected polluted water body collected from the target area to form new water body vector merged data. Water body vector data can be downloaded from the OpenStreetMap official website, and the water body recognition algorithm can be used to extract the latest water body in the target area. Merging with the collected suspected polluted water body vector data can ensure the integrity of the water body vector. Water body vector data is the key data for realizing the classification and identification of suspected polluted water bodies. Only by using water body vector data to extract water bodies from surface reflectance remote sensing images can the pollution type be classified. Buffer the water body vector merged data according to the specified distance to generate a buffer zone, and obtain the buffer zone vector data containing the surrounding environment of the suspected polluted water body.

[0040] The specific operation method includes the following sub-steps: (1) Download the pbf vector data of the target area from the OpenStreetMap official website, use the Osmosis command-line Java application to convert the pbf vector data into osm format, and retrieve various water body vectors such as water (water area), reservoir (reservoir, reservoir) and wetland (wetland, swamp) from the fclass field of the vector, and save them as various water body vector files in osm. Use the Load OSM tool in the ArcGIS Editor for OSM plug-in to save various water body vector files in osm format into shp format. After merging, the water body vector data of suspected polluted water bodies collected from the target area is formed.

[0041] (2) Set the threshold for distinguishing water bodies from non-water bodies, and use the gdal tool in Python to extract four bands from the surface reflectance remote sensing image; use water body identification algorithms such as the normalized difference water index (NDWI) to calculate the surface reflectance remote sensing image, and extract the water body raster from the surface reflectance remote sensing image according to the set threshold; use the raster to vector tool to convert the extracted water body raster into a water body vector; use the merge tool to merge the obtained water body vector with the water body vector of the suspected polluted water body collected from the target area to form the water body vector merge data.

[0042] (3) Use the Buffer tool to buffer the water body vector merge data according to the specified distance to generate a buffer zone, and obtain the buffer zone vector data that includes the surrounding environment of the water body, which is used to crop the surface reflectance remote sensing image of the target area.

[0043] Step 3: Use the buffer vector data to crop the surface reflectance remote sensing image of the target area to obtain a large amount of water body image raster data with a small single file data volume, and use this to make a test set for deep learning model training. Since the target area is covered by multiple single-scene surface reflectance remote sensing images, and the water body buffer vector data is also composed of multiple data, it is necessary to loop to determine whether each remote sensing image overlaps with each water body vector data record. If they overlap, they are cropped, and if they do not overlap, they are skipped, and this cycle repeats.

[0044] The test set is made as follows: (1) To determine whether each remote sensing image overlaps with each water body vector data record, it is necessary to extract the effective value area boundary of the remote sensing image and then determine whether the image boundary overlaps with the water body vector. Therefore, the raster extent tool of ArcGIS is used to batch generate raster extent face vectors for all images along the raster surface boundary. The file name is the same as the raster file name, followed by the extent suffix.

[0045] (2) If the band order of the multi-view surface reflectance remote sensing images is inconsistent, use Python software combined with the "Generate Raster from Raster Function" tool in ArcGIS to extract the bands of the remote sensing images in layers, and then use the "Band Synthesis" tool to synthesize the layered bands to make the band order of all multi-view surface reflectance remote sensing images consistent.

[0046] (3) According to the tif suffix condition, the remote sensing image files are searched in a loop and saved in the remote sensing image file list. According to the extent suffix condition, the raster range files are searched in a loop and saved in the raster range file list. According to the remote sensing image file list, each remote sensing image file is looped, and each row of the water body buffer vector data is searched in the loop. The intersection tool is used to calculate the intersection area of ​​each raster range and each water body vector data record, and the number of intersection area entries is calculated: if it is greater than 0, the ArcGIS clipping tool is used for clipping; if it is 0, it indicates that the two do not overlap and are skipped; this cycle is repeated.

[0047] After the test set is completed, it is necessary to create samples required for the classification of polluted water bodies, tell the deep learning recognition model the characteristics of various polluted water bodies, and then carry out automatic learning and classification on the test set.

[0048] Step 4, referring to the water body vector data of suspected polluted water bodies collected from the target area and the water pollution public opinion data, artificial remote sensing interpretation is performed on various typical types of suspected polluted water bodies on the surface reflectance remote sensing image of the target area. The present invention divides polluted water bodies into three categories: duckweed water bodies, black and smelly water bodies, and garbage water bodies. When performing artificial remote sensing interpretation, field verification is carried out simultaneously to determine whether the artificial remote sensing interpretation results belong to polluted water bodies. The polluted water body points with correct interpretation are sorted to form a c sample set.

[0049] The polluted water sample set is prepared as follows: (1) Obtain recent water pollution information in the target area from news platforms, short video platforms, search websites and other channels. According to the location of the water pollution, find the specific location of the polluted water body from the surface reflectance remote sensing image of the target area. Analyze and determine whether the water body in the surface reflectance remote sensing image is still polluted. If the judgment result is "yes", record the location of the water body.

[0050] (2) Set a water pollution threshold, use the band ratio method to calculate the surface reflectance remote sensing image, obtain the polluted area of ​​suspected water pollution, and record the location of the suspected polluted water body.

[0051] (3) Analyze and summarize the characteristics of the suspected polluted water bodies found in the above two steps, and based on this, continue to search for similar suspected polluted water bodies from surface reflectance remote sensing images.

[0052] (4) Based on the recorded location information of suspected polluted water bodies, field inspections are carried out simultaneously in the target areas. In combination with the "Guidelines for the Treatment of Black and Smelly Water Bodies in Rural Areas" issued by the Ministry of Ecology and Environment and the "Guidelines for the Treatment of Black and Smelly Water Bodies in Cities" issued by the Ministry of Housing and Urban-Rural Development, the water bodies are manually observed for black or abnormal colors, and for domestic garbage, domestic wastewater, industrial solid waste, and livestock and poultry feces in or around the water bodies, to observe whether the water bodies have odors; instruments are used to detect whether water quality indicators such as transparency, dissolved oxygen, redox potential, and ammonia nitrogen are normal. Through a variety of means and methods, it is determined whether the water bodies at the inspection points are polluted. If the judgment result is "yes", the location of the water body is recorded, and the type of pollution is also recorded.

[0053] (5) Based on the field verification results, the correctly interpreted polluted water points are sorted out, and other suspected polluted water bodies with high similarity to the correctly interpreted points are found on the surface reflectance remote sensing images of the target area to expand the number of samples and form a relatively rich polluted water sample set. Finally, surface features are created for various types of polluted water bodies on the ArcGIS platform.

[0054] Step 5: After the sample sets of various polluted water bodies are prepared, they need to be made into a format that can be recognized by deep learning, namely image slices and mask slices. Image slices refer to image images containing polluted water bodies. Mask slices are the same size as image slices, and the range of polluted water bodies is represented by the two values ​​of "0" and "1". Image slices and mask slices are randomly divided into prediction sets and validation sets in proportion. The prediction set is used for learning and parameter optimization of deep learning models, and the validation set is used to adjust the parameters and model selection of deep learning models.

[0055] The preparation of prediction set and validation set includes the following sub-steps: (1) Use Python software and the "Export training data for deep learning" tool in ArcGIS to create a batch callable function, use the os module in Python software to traverse all remote sensing images used in the workspace, and call the above function. Loop through each vector boundary of a polluted water body and each scene surface reflectance remote sensing image, and set the output format to RCNN_Masks. When the vector boundary and the remote sensing image overlap, they become image slices, and a segmentation mask is generated, and the image slices and mask slices with the same name in tif format are output. The image slices are stored in the images folder, and the mask slices are stored in the labels folder. When the vector boundary and the remote sensing image do not overlap, skip this combination and continue the loop.

[0056] (2) Using the PIL image processing module in Python software, we batch-extracted the three RGB bands from the image slices in tif format, and applied linear transformation to adjust the color histogram of each band in the image slices. After band synthesis, we formed image slices in JPG format, thereby unifying the colors of image slices from different sources and at different times.

[0057] Once the prediction set and validation set of the deep learning model are completed, a variety of deep learning models can be used to classify polluted water bodies.

[0058] Step 6. Based on the Pytorch framework, use the prediction set to train each deep learning model including U-Net, SegNet and DeepLabv3+. Use the intersection over union (IoU) and F1-score indicators to test the accuracy of each deep learning model. Use multi-category cross entropy (Categorical_crossentropy) as the loss function, and set the learning rate and optimizer parameters according to the actual situation. Use the WANDB online model training visualization tool to track and record the parameters and indicators of each deep learning model in operation. According to the recorded indicators, iterate the parameters of each deep learning model to improve the accuracy of each deep learning model. Finally, select the deep learning model with the highest accuracy as the deep learning model for suspected polluted water classification.

[0059] Among them, the calculation formula of the intersection over union (IoU) indicator is: IoU = | A ∩ B | / | A ∪ B | = | I | / | U | In the formula, A represents the predicted bounding box range, B represents the real bounding box range, I represents the intersection area of ​​A and B, and U represents the joint area of ​​A and B.

[0060] The calculation formula of the F1-score indicator is: F1-score = 2 ( Precision Recall ) / ( Precision + Recall ) In the formula, Precision represents the accuracy and Recall represents the recall.

[0061] Step 7: Use the selected suspected polluted water classification deep learning model to perform computer automatic classification on the test set to obtain the suspected polluted water classification results. Manually screen the classification results of each type of suspected polluted water, remove obvious non-water body images and misclassified images, and then compile a list of suspected polluted water bodies in the target area, and conduct field verification in a timely manner to finally obtain the remote sensing monitoring results of polluted water bodies in the target area.

[0062] The deep learning classification method for the test set is: (1) The test set data is made into a format that can be recognized by deep learning. The test set is made into image slices and mask slices by using the image slice and mask slice production process in the prediction set and validation set and the "Export Training Data for Deep Learning" tool in ArcGIS.

[0063] (2) Use the Get Raster Properties tool to obtain the top, bottom, left, and right coordinate values ​​of the image slice range, calculate the center point coordinate value, and store the image slice file name and the center point coordinate value of the image slice in a text file to link the classification results with the geographic location.

[0064] (3) Using the deep learning model for classifying suspected polluted water bodies, the image slices of the test set were automatically classified by computer to obtain the classification results of the suspected polluted water bodies.

[0065] (4) Move the images of each type of suspected polluted water body classification results to the corresponding folder, and remove the obvious non-water body images and misclassified images. Based on the file names of the remaining suspected polluted water body classification result images, retrieve the coordinate values ​​of their center points from the text file to obtain the geographical location of the suspected polluted water body.

[0066] (5) Convert the location of the suspected polluted water body into a point vector. Use the Buffer tool to buffer it according to the specified distance to generate a buffer zone. Merge the water points that are close to each other into one water body to form a surface vector of the suspected polluted water body.

[0067] (6) Convert the surface vector of the suspected polluted water body into a KMZ file, import it into the Aowei Interactive Earth mobile software, plan the route and conduct field verification in a timely manner, and obtain the remote sensing monitoring results of the polluted water body in the target area according to the above-mentioned field verification method.

Claims

1. A method for classifying and identifying suspected polluted water bodies using high-resolution remote sensing, characterized in that: The following steps are involved: S1. Acquire high-resolution remote sensing image data of the target area, pre-process the acquired high-resolution remote sensing image data to obtain image fusion data, correct the image fusion data, and form a surface reflectance remote sensing image of the target area; S2. Using a water body recognition algorithm, identify the water body vector range of the suspected polluted water body from the surface reflectance remote sensing image, and then merge it with the water body vector data of the suspected polluted water body collected from the target area to form water body vector merged data; buffer the water body vector merged data according to a specified distance to generate a buffer zone, and obtain buffer zone vector data including the surrounding environment of the suspected polluted water body; S3, using the buffer vector data, cropping the surface reflectance remote sensing image to obtain the water body image raster data, and making a test set based on it; S4. Refer to the water body vector data of suspected polluted water bodies collected from the target area and the public opinion data on water pollution, conduct manual remote sensing interpretation of various typical types of suspected polluted water bodies on the surface reflectance remote sensing images, and simultaneously conduct field verification to determine whether the manual remote sensing interpretation results belong to polluted water bodies; organize the polluted water body points with correct interpretation to form a polluted water body sample set; S5. Based on the polluted water sample set, image slices and mask slices required for deep learning are produced on the surface reflectance remote sensing image, and the produced image slices and mask slices are randomly divided into a prediction set and a verification set according to proportion; S6. Based on the Pytorch framework, use the prediction set to train two or more deep learning models separately; use the intersection over union (IoU) and F1-score indicators to test the accuracy of each deep learning model; use multi-category cross entropy as the loss function to set the learning rate and optimizer parameters; use the WANDB online model training visualization tool to track and record the parameters and indicators of each deep learning model in operation, and iterate the parameters of each deep learning model according to the recorded indicators to improve the accuracy of each deep learning model. Finally, select the deep learning model with the highest accuracy as the deep learning model for suspected polluted water classification; S7. Using the suspected polluted water body classification deep learning model, the test set is automatically classified by computer to obtain the suspected polluted water body classification result; The classification results of suspected polluted water bodies are manually screened, non-water body images and misclassified images are eliminated, and a list of suspected polluted water bodies in the target area is compiled. After on-site verification, the remote sensing monitoring results of polluted water bodies in the target area are obtained.

2. The method for classifying and identifying suspected polluted water bodies using high-resolution remote sensing according to claim 1 is characterized in that: The preprocessing of high-resolution remote sensing image data described in step S1 includes radiation correction, atmospheric correction, orthorectification and image fusion.

3. The method for classifying and identifying suspected polluted water bodies using high-resolution remote sensing according to claim 1 is characterized in that: Step S2 includes the following sub-steps: S2-1 obtains the pbf vector data of the target area, retrieves various water body vectors including water, reservoir and wetland, and saves them as various water body vector files; merges various water body vector files to form water body vector data of suspected polluted water bodies collected from the target area; S2-2 sets a threshold for distinguishing water bodies from non-water bodies, uses a water body recognition algorithm, extracts water body rasters from surface reflectance remote sensing images, and converts the extracted water body rasters into water body vectors; merges the obtained water body vectors with the water body vectors of suspected polluted water bodies collected from the target area to form water body vector merged data; S2-3 buffers the water body vector merge data according to the specified distance, generates a buffer zone, and obtains buffer zone vector data including the surrounding environment of the water body.

4. The method for classifying and identifying suspected polluted water bodies using high-resolution remote sensing according to claim 1 is characterized in that: The preparation of the polluted water sample set in step S4 includes the following sub-steps: S4-1 obtains the recent water pollution information of the target area, finds the location of the polluted water body from the surface reflectance remote sensing image according to the location of the water pollution, analyzes and determines whether the water body on the surface reflectance remote sensing image is still polluted, and if the judgment result is "yes", records the point of the water body; S4-2 sets the water pollution threshold, calculates the surface reflectance remote sensing image using the band ratio method, obtains the pollution area of ​​the suspected polluted water body, and records the location of the suspected polluted water body; S4-3 analyzes and summarizes the characteristics of the suspected polluted water bodies found through the above two sub-steps, and based on this, continues to search for similar suspected polluted water bodies from the surface reflectance remote sensing image; S4-4 Based on the recorded location information of suspected polluted water bodies, field inspections are carried out simultaneously to determine whether the water bodies at the inspection points are polluted through manual observation and instrument detection. If the judgment result is "yes", the location of the water body is recorded, and the type of pollution is also recorded; S4-5 sorts out the correctly interpreted polluted water points based on the field verification results, and searches for other suspected polluted water bodies that are highly similar to the correctly interpreted points on the surface reflectance remote sensing images, expands the number of samples, and forms a polluted water sample set.

Images