A Method for Identifying Methane Leakage Points and Plumes Applicable to Hyperspectral Imagers

Abstract

The present invention relates to a method for identifying methane leakage points and plumes applicable to hyperspectral imagers, comprising the following steps: S1. Using the matched filtering method and the point-source hyperspectral data of the target area, an inverse methane concentration grayscale image is obtained, and water bodies and shadow areas in the methane concentration grayscale image are removed to obtain basic methane concentration data; S2. Using the density-based DBSCAN noise reduction algorithm to reduce the background noise in the basic methane concentration data and improve the signal-to-noise ratio of the image to obtain noise-reduced methane concentration data; S3. Using the YOLO image classification model to remove the pseudo-methane anomalies caused by other types of objects in the noise-reduced methane concentration data; S4. Using the plume identification algorithm based on pixel connectivity to segment the methane plume from the noise-reduced methane concentration data; S5. Superimposing the methane plume with wind speed data and atmospheric pressure data, and using the IME algorithm to estimate the methane plume emission amount, solving the problem that the real data of the image is affected by filter noise reduction.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image recognition, and particularly relates to a method for identifying methane leakage points and plumes suitable for hyperspectral imagers. Background Art

[0002] Methane (CH4) is a potent greenhouse gas with a much higher global warming potential than carbon dioxide, and its concentration in the atmosphere is increasing year by year. The main sources of methane emissions include natural sources such as fossil fuel extraction, agricultural activities, landfills, and wetlands. Given the huge impact of methane on climate change, how to accurately monitor its emissions and formulate effective emission reduction strategies has become the focus of global attention. In recent years, with the rapid development of remote sensing technology, satellite-based methane monitoring means have gradually matured. In particular, imaging spectrometers in the short-wave infrared (SWIR) band provide important data support for global methane emission monitoring;

[0003] According to different observation purposes, current methane satellite payloads can be divided into two categories: regional and point-source types. Regional satellites such as GOSAT and Sentinel-5P have high spectral resolution and concentration measurement accuracy, and can monitor the atmospheric methane column concentration globally. These satellites usually have a large observation swath and a short revisit period, and are suitable for monitoring methane emissions at regional and global scales. However, due to the low spatial resolution (about 5 - 50 km), regional satellites are difficult to identify small-scale point-source emissions. In contrast, point-source satellites such as GHGSat, PRISMA, and China's GF5-02 have high spectral resolution and spatial resolution (about 10 - 100 m), and can detect high-concentration methane plumes near point sources. These satellites are mainly used to identify and quantify large point-source emissions, such as oil wells, storage tanks, coal mines, and landfills;

[0004] In the research on methane concentration inversion based on point-source hyperspectral satellite data, the Matched Filtering (MF) method is widely used. This algorithm realizes the identification and concentration estimation of the target gas by maximizing the matching degree between the spectral characteristics of the target gas and the known reference spectral library. However, factors such as complex terrain conditions, surface reflection characteristics, spectrally similar features, and instrument noise will interfere with the inversion results, resulting in a large amount of background noise in the inversion results, which is not conducive to the identification of methane leakage points and methane plumes;

[0005] At present, many studies have explored methods for identifying methane leakage points and methane plumes. Most of the existing methane leakage point identification methods use simple Gaussian filtering to remove background noise in the methane concentration grayscale image, and then perform threshold segmentation to identify methane leakage points and leakage areas. This filter-based noise reduction will cause damage to the real data of the image, resulting in problems such as loss of original pixel values and decrease in image resolution, affecting the accuracy of methane leakage point identification. At the same time, due to factors such as complex terrain, uneven surface reflection, and spectrally similar ground objects in reality, there are a large number of false alarm signals in the methane leakage points obtained by the threshold segmentation method;

[0006] Current methods for identifying the range of methane plumes include visual discrimination, using a meteorological data model to perform single-point plume emission simulation on methane leakage points and match with the methane concentration map to obtain the optimal matching plume, using 2-D plume image simulation training convolutional neural network models under various wind speed conditions and background noise to identify methane plumes. Visual discrimination has high labor costs and low accuracy and cannot be widely applied. The meteorological data model matching method has a large amount of model calculations, high operating costs, and low simulation accuracy under complex terrain, and its applicability is not strong in practical applications. The training of convolutional neural network models requires a large number of real and manually labeled methane plume images. In reality, the data volume is insufficient. To pursue high accuracy, large-volume models need to be used, which have extremely high requirements for the amount of calculation and the size of the database. At the same time, the neural network model performs poorly when the methane plume concentration is low. Some studies have tried to use large-eddy simulation (LES) to simulate actual observed plume data for training machine learning models. However, its calculation cost is high and it is not conducive to practical deployment.

[0007] Due to the above factors, there are a large number of false alarm signals in the results identified by the existing methane leakage point identification methods. Methane plume identification methods have problems such as inaccurate methane plume range, large amount of calculation, high labor and data costs. So far, there is no methane leakage point and plume identification method in the public domain with high automation, strong applicability, high accuracy, and small algorithm calculation amount. Summary of the Invention

[0008] The purpose of the present invention is to provide a methane leakage point and plume identification method applicable to a hyperspectral imager to solve the problems existing in the above-mentioned prior art.

[0009] The above technical purpose of the present invention is achieved through the following technical solutions:

[0010] A methane leakage point and plume identification method applicable to a hyperspectral imager includes the following steps:

[0011] Step S1: Use the matched filtering method and the point source type hyperspectral data of the target area to invert the methane concentration grayscale image, and remove the water body and shadow areas in the methane concentration grayscale image to obtain the basic methane concentration data;

[0012] Step S2: Use the density-based DBSCAN noise reduction algorithm to reduce the background noise in the basic methane concentration data and improve the signal-to-noise ratio of the image to obtain methane concentration noise-reduced data. The step S2 further includes:

[0013] Step S21: Use the threshold segmentation method to remove the low-concentration part in the basic methane concentration data. The formula is: where Δα and Δα1 are the basic methane concentration and the methane concentration after threshold segmentation respectively, and Δα t1 is the methane concentration threshold set according to the specific situation of the project;

[0014] Step S22: Use the concave function to perform a non-linear transformation on the segmented basic methane concentration data to obtain the methane concentration after non-linear transformation. The formula is: Δα2 = f(Δα1), where Δα2 is the methane concentration after non-linear transformation, making the data in the plume part denser in the concentration dimension and easier to be clustered into clusters;

[0015] Step S23: Use the DBSCAN algorithm to perform clustering in the pixel horizontal, vertical, and concentration dimensions, and further divide the basic methane concentration data after non-linear transformation in step S22 into a methane plume area and a background noise area. The methane plume area is a set of pixel area collections clustered into the same cluster by the algorithm, and the noise area is a set of pixel area collections that cannot meet the algorithm clustering conditions and thus cannot be clustered into any cluster. The value range of the DBSCAN algorithm parameter min_sample is: 5 - 10, and the value range of the eps parameter is: 1 - 5. The eps parameter refers to the neighborhood distance threshold of a certain sample point, and the min_sample parameter refers to the threshold of the total number of sample points included in the neighborhood with a distance of the eps for a certain sample point; Subtract the background noise area from the basic methane concentration data after non-linear transformation to obtain the methane concentration noise-reduced data; This noise reduction algorithm is a lossless noise reduction method that does not lose the original pixel values and does not change the resolution;

[0016] Step S3: Use the YOLO image classification model to remove the pseudo-methane anomalies caused by other types of objects in the methane concentration noise-reduced data in step S2. The step S3 further includes:

[0017] Step S31: Training dataset production. Collect the historical public data of methane leakage points in the target area, download the Sentinel-2 cloud-free optical image slices with a length and width of 100 pixels and a resolution of 10 meters centered on the leakage point, and re-comb and correct them through manual interpretation, and divide them into a "methane leakage area" and a "non-methane leakage area" as the training dataset;

[0018] Step S32: Use the training dataset described in Step S31 to train the YOLO image classification model, and obtain a binary classification model that divides the Sentinel-2 cloud-free optical image slices into "methane leakage area" and "non-methane leakage area".

[0019] Step S33: Perform threshold division on the methane concentration noise-reduced data described in Step S2. Define the patches with methane concentration Δα > Δα t2 as candidate methane leakage areas, where Δα t2 is a custom threshold. Determine the point corresponding to the highest methane concentration value in the methane leakage area as the candidate methane leakage point; obtain the coordinates of each candidate methane leakage point, and download the Sentinel-2 cloud-free optical image slice with a length and width of 100 pixels each and a resolution of 10 meters centered at this point.

[0020] Step S34: Deploy the binary classification model described in Step S32 and use the model to perform predictive inference on all the Sentinel-2 cloud-free optical image slices described in Step S33. Retain the parts predicted as "methane leakage area" and mark the corresponding candidate methane leakage points as pending methane leakage points.

[0021] Step S4: Use a plume recognition algorithm based on pixel connectivity to segment methane plumes from the methane concentration noise-reduced data described in Step S2. The Step S4 further includes:

[0022] Step S41: Binarize the methane concentration noise-reduced data described in Step S2. Set the pixels with methane concentration greater than Δα t1 described in Step S21 to 1, and set other pixels to 0 to obtain the methane concentration binarized data.

[0023] Step S42: Convolve the binarized image with a convolution kernel of size k×k to obtain a neighborhood connectivity number result map of size (w-k+1)×(h-k+1), where w and h are the width and height of the original methane concentration binarized data image respectively. The value of each pixel in the map is the number of non-zero pixel points in the k-neighborhood around the corresponding pixel point in the original methane concentration binarized data image. Use the threshold n to extract the pixel positions in the neighborhood connectivity number result map with values greater than n to obtain multiple plume patches.

[0024] Step S43: Use the 8-neighborhood connectivity domain algorithm to cluster all the plume patches to obtain all the candidate plume patches.

[0025] Step S44: Traverse the candidate plume patches, and if the methane concentration contained in them is less than the threshold Δα t2Filter out the candidate plume patches; mark the candidate plume patches containing the m pending methane leakage points with m steps S34 as methane plume masks, where m >= 1, and cluster the m pending methane leakage points, with each cluster being a methane leakage point, and its coordinates being the geometric center of the pending methane leakage points belonging to the cluster; filter out the candidate plume patches that do not contain the pending methane leakage points; filter out the pending methane leakage points that do not belong to any candidate plume patches;

[0026] Step S45: Traverse the methane plume mask, index the methane concentration values in the methane concentration basic data map in step S1 according to the coverage range thereof, and fill the plume mask range to obtain all methane plumes.

[0027] Step S5: Superimpose the methane plume with the wind speed data and the atmospheric pressure data, use the IME algorithm to estimate the methane plume emission amount, and finally perform error analysis.

[0028] In summary, the present invention has the following beneficial effects:

[0029] 1. The present invention applies the DBSCAN clustering algorithm to denoise the methane concentration grayscale image inversed by the matched filtering algorithm, solves the problem that the real data of the image is affected by filter denoising, and improves the signal-to-noise ratio of the image without losing the original pixel values and without changing the resolution. The algorithm is simple and efficient.

[0030] 2. The present invention uses Sentinel-2 optical images and the YOLO model to remove pseudo-methane anomaly points, can accurately filter a large number of abnormal pseudo-methane leakage areas, and greatly improves the accuracy and efficiency of methane leakage point recognition compared with the prior art, with high innovation.

[0031] 3. In view of the characteristics of methane plume diffusion, the present invention specifically designs a plume recognition algorithm based on pixel connectivity, with small computational complexity and strong applicability to actual cases.

[0032] 4. The present invention adopts the idea of an integrated algorithm to improve the accuracy of the final result: make the candidate plume patches obtained by the methane plume recognition algorithm intersect and fuse with the pending methane leakage points obtained by the YOLO classification algorithm, and finally obtain the methane leakage points and the methane plume, with high accuracy of the obtained methane leakage points and more accurate methane plume range. Description of the Drawings

[0033] Figure 1 is a flowchart of a method for identifying methane leakage points and plumes applicable to a hyperspectral imager according to the present invention;

[0034] Figure 2 is an image before processing and denoising the methane concentration grayscale image using the DBSCAN denoising algorithm;

[0035] Figure 3 It is the image in the noise reduction obtained by processing the methane concentration grayscale image using the DBSCAN noise reduction algorithm;

[0036] Figure 4 It is the image after noise reduction obtained by processing the methane concentration grayscale image using the DBSCAN noise reduction algorithm;

[0037] Figure 5 It is the normalized confusion matrix of the test set of the YOLO image classification model;

[0038] Figure 6 It is the identification result of the methane leakage point and methane plume in the embodiment. Specific implementation mode

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] Among them, the same components are denoted by the same reference numerals. It should be noted that the terms "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to the directions in the accompanying Figure 1 drawings, and the terms "bottom surface" and "top surface", "inner" and "outer" refer to the directions towards or away from the geometric center of a specific component. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of this specification, "a plurality" means two or more unless otherwise specifically defined.

[0041] Example 1: As Figure 1 shown, a method for identifying methane leakage points and plumes applicable to a hyperspectral imager includes:

[0042] Step S1: Using the matched filtering method and the point source type hyperspectral data of the target area, the methane concentration grayscale image is inversely calculated, and the water body and shadow area in the methane concentration grayscale image are removed to obtain the basic methane concentration data;

[0043] Step S2: Using the density-based DBSCAN noise reduction algorithm to reduce the background noise in the basic methane concentration data and improve the signal-to-noise ratio of the image to obtain the methane concentration noise reduction data. The specific steps are as follows;

[0044] Step S21: Using the threshold segmentation method to remove the low-concentration part in the basic methane concentration data. The formula is: where Δα and Δα1 are the basic methane concentration and the methane concentration after threshold segmentation respectively, and is the methane concentration threshold set according to the specific situation of the project. In this embodiment, it is taken as 100 ppb;

[0045] Step S22: Use a concave function Perform a non-linear transformation on the segmented methane concentration basic data to obtain the methane concentration after non-linear transformation. The formula is: Δα2 = f(Δα1), where Δα2 is the methane concentration after non-linear transformation, making the data of the plume part denser in the concentration dimension and easier to be clustered into clusters;

[0046] Step S23: Use the DBSCAN algorithm to perform clustering in the pixel horizontal, vertical and concentration dimensions, and further divide the methane concentration basic data after the non-linear transformation in step S22 into a methane plume area and a background noise area. The methane plume area is a set of pixel areas clustered into the same cluster by the algorithm, and the noise area is a set of pixel areas that cannot meet the algorithm clustering conditions and thus cannot be clustered into any cluster. The value range of the DBSCAN algorithm parameter min_sample is: 8, and the value range of the eps parameter is: 2. The eps parameter refers to the neighborhood distance threshold of a certain sample point, and the min_sample parameter refers to the threshold of the total number of sample points included in the neighborhood where the distance of a certain sample point is the eps; Subtract the background noise area from the methane concentration basic data after non-linear transformation to obtain the methane concentration noise reduction data; This noise reduction algorithm is a lossless noise reduction method, which does not lose the original pixel value and does not change the resolution;

[0047] After Δα' = f(Δα), The transformed methane concentration value is as Figure 2 shown. After separating the background noise by the DBSCAN noise reduction algorithm, as Figure 3 shown. Finally, an image of the methane concentration value minus the separated background noise is obtained, as Figure 4 shown. It can be seen that this noise reduction algorithm can well separate the background noise and extract the methane concentration basic data;

[0048] Step S3: Use the YOLO image classification model to remove the pseudo-methane anomalies caused by other types of objects in the methane concentration noise reduction data in step S2. The specific steps are as follows;

[0049] Step S31: Training dataset production. Collect the historical public data of methane leakage points in the target area, download the Sentinel-2 cloud-free optical image slices centered on the leakage point with a length and width of 100 pixels each and a resolution of 10 meters, and re-comb and correct them through manual interpretation, and divide them into a "methane leakage area" and a "non-methane leakage area" as the training dataset;

[0050] Step S32: Use the training dataset described in Step S31 to train the YOLO image classification model, and obtain a binary classification model that divides Sentinel-2 cloud-free optical image slices into "methane leakage area" and "non-methane leakage area".

[0051] The types of underlying surfaces corresponding to methane leakage points are limited in variety and are concentrated in certain categories. Table 1 shows the classification statistics of methane leakage points in the CarbonMapper aerial methane leakage point dataset in the United States from 2016 to 2024. Among them, the proportion of the top 3 types of methane leakage points can reach 90%, namely oil and gas facilities, solid waste, and power facilities. These ground objects can be uniformly identified as the "methane leakage area", and other methane leakage points with a proportion of less than 10%, such as wastewater, grassland, photovoltaic panels, and plastic greenhouses, are grouped into the "non-methane leakage area".

[0052]

[0053]

[0054] Divide the training dataset into a training set and a test set in a ratio of 2:8, and use the YOLO11 s image classification model for training. The main parameters are shown in Table 2:

[0055]

[0056] Step S33: Perform threshold division on the methane concentration noise reduction data described in Step S2, and define the patches with methane concentration Δα > Δα t2 as candidate methane leakage areas, where Δα t2 is a user-defined threshold, and determine the point corresponding to the highest methane concentration in the methane leakage area as the candidate methane leakage point; obtain the coordinates of each candidate methane leakage point, and download Sentinel-2 cloud-free optical image slices with a length and width of 100 pixels each centered on this point and a resolution of 10 meters.

[0057] Step S34: Deploy the binary classification model described in Step S32 and use the model to perform predictive inference on all Sentinel-2 cloud-free optical image slices described in Step S33. Retain the part predicted as the "methane leakage area" and mark the corresponding candidate methane leakage points as pending methane leakage points.

[0058] As Figure 5 shown, the normalized confusion matrix of the classification model test set can achieve a classification accuracy of 0.90 in the test set, indicating that the model classification results are reliable.

[0059] Step S4. According to the Gaussian plume model, the methane plume dissipates from the leakage point. Along the axial and radial increasing directions, the edge of the plume gradually expands and the methane concentration gradually decreases until it converges with the background concentration. Correspondingly, in the methane concentration grayscale image, the pixels contained in the methane plume should have a certain density in space. Using a plume recognition algorithm based on pixel connectivity, the methane plume is segmented from the methane concentration noise reduction data described in step S2. The specific steps are as follows:

[0060] Step S41. Binarize the methane concentration noise reduction data described in step S2. Set the pixels with methane concentration greater than Δα described in step S21 to 1, and set other pixels to 0 to obtain the binarized methane concentration data; t1 of the pixels to 1, and set other pixels to 0, obtaining the binarized methane concentration data;

[0061] Step S42. Convolve the binarized image with a convolution kernel of size k×k to obtain a neighborhood connectivity number result image of size (w - k + 1)×(h - k + 1), where w and h are the width and height of the original binarized methane concentration data image respectively. The value of each pixel in the image is the number of non-zero pixels in the k-neighborhood around the corresponding pixel point in the original binarized methane concentration data image. Use the threshold n to extract the pixel positions in the neighborhood connectivity number result image with values greater than n to obtain multiple plume patches. In this embodiment, k = 11 and n = 20 are taken;

[0062] Step S43. Use the 8-neighborhood connectivity domain algorithm to cluster all the plume patches to obtain all candidate plume patches;

[0063] Step S44. Traverse the candidate plume patches, filter out the candidate plume patches whose contained methane concentrations are all less than the threshold Δα described in step S33 t2 ; Mark the candidate plume patches containing m undetermined methane leakage points described in step S34 as methane plume masks, where m >= 1, and cluster the m undetermined methane leakage points. Each cluster is a methane leakage point, and its coordinates are the geometric center of the undetermined methane leakage points belonging to the cluster; Filter out the candidate plume patches that do not contain the undetermined methane leakage points; Filter out the undetermined methane leakage points that do not belong to any candidate plume patches;

[0064] Step S45. Traverse the methane plume mask, index the methane concentration values in the methane concentration basic data image described in step S1 according to the coverage range it covers, and fill the plume mask range to obtain all methane plumes;

[0065] Such as Figure 6As shown, the methane leakage points and the identification results of methane plumes. The basic methane concentration data is overlaid with the pending methane leakage points predicted as "methane leakage points" by the YOLO classification algorithm, which are represented by red circles; the basic methane concentration data is also overlaid with the methane plumes identified by the methane plume identification algorithm. The intersection of the pending methane leakage points and the methane plumes is the methane leakage points and methane plumes finally determined in this study.

[0066] Visually judge the methane plumes obtained by the methane plume identification algorithm, screen out the real part of the methane plumes as the true value of the methane plumes, and compare it with the methane leakage points and methane plumes described in step S45. The accuracy rate of the present invention in the embodiment can be obtained. The comparison results are shown in Table 3, and the identification accuracy rate can reach 96.36%.

[0067]

[0068] Step S5: Overlay the methane plumes with wind speed data and atmospheric pressure data, use the IME algorithm to estimate the methane plume emissions, and finally conduct error analysis.

[0069] Specific implementation process: Use the matched filtering method and the point-source hyperspectral data of the target area to feedback the methane concentration grayscale image, remove the water body and shadow areas in the methane concentration grayscale image to obtain the basic methane concentration data, use the threshold segmentation method to remove the low-concentration part in the basic methane concentration data, perform a non-linear transformation on the segmented basic methane concentration data to obtain the non-linearly transformed methane concentration, and then further divide the non-linearly transformed basic methane concentration data into methane plume areas and background noise areas. Then subtract the background noise area from the non-linearly transformed basic methane concentration data to obtain the methane concentration noise reduction data.

[0070] In the embodiments disclosed in the present invention, terms such as "installation", "connection", "connection", "fixation", etc. should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; "connection" can be a direct connection or an indirect connection through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in the embodiments disclosed in the present invention can be understood according to specific circumstances.

[0071] This specific embodiment is only an explanation of the present invention and does not limit the present invention. Those skilled in the art can make modifications without creative contributions to this embodiment after reading this specification, but as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A method for identifying methane leakage points and plumes applicable to hyperspectral imagers, characterized in that: It includes the following steps: Step S1: Using the matched filtering method and the point-source hyperspectral data of the target area, the methane concentration grayscale image is inverted, and the water body and shadow areas in the methane concentration grayscale image are removed to obtain the basic methane concentration data; Step S2: Using the density-based DBSCAN denoising algorithm to reduce the background noise in the basic methane concentration data and improve the signal-to-noise ratio of the image to obtain the methane concentration denoised data. The step S2 further includes: Step S21. Remove the low-concentration part from the basic methane concentration data using the threshold segmentation method. The formula is as follows: where Δα and Δα1 are the basic methane concentration and the methane concentration after threshold segmentation respectively, and Δα t1 is the methane concentration threshold set according to the specific situation of the project; Step S22: Use a concave function Perform a non-linear transformation on the basic methane concentration data after segmentation to obtain the basic methane concentration data after non-linear transformation; Step S23: Using the DBSCAN algorithm to cluster in the pixel horizontal, vertical, and concentration dimensions, and further dividing the basic methane concentration data after the non-linear transformation in step S22 into a methane plume area and a background noise area. The methane plume area is a set of pixel areas clustered into the same cluster by the algorithm, and the background noise area is a set of pixel areas that cannot meet the algorithm clustering conditions and thus cannot be clustered into any cluster. Subtracting the background noise area from the basic methane concentration data after the non-linear transformation to obtain the denoised data of the basic methane concentration data after the non-linear transformation; Step S3: Using the YOLO image classification model to remove the pseudo-methane anomalies caused by other types of objects in the methane concentration denoised data in step S2. The step S3 further includes: Step S31: Training dataset production. Collect the historical public data of the methane leakage points in the target area, download the Sentinel-2 cloud-free optical image slices with a length and width of 100 pixels and a resolution of 10 meters centered on the leakage point, and re-comb and correct them through manual interpretation, and divide them into a methane leakage area and a non-methane leakage area as the training dataset; Step S32: Using the training dataset in step S31 to train the YOLO image classification model to obtain a binary classification model that divides the Sentinel-2 cloud-free optical image slices into a methane leakage area and a non-methane leakage area; Step S33: Perform threshold division on the methane concentration noise reduction data described in step S2, and define the patches where the methane concentration Δα > Δα t2 as candidate methane leakage areas. Δα t2 is a user-defined threshold, and determine the point corresponding to the highest methane concentration value in the methane leakage area as the candidate methane leakage point; obtain the coordinates of each candidate methane leakage point, and download a Sentinel-2 cloud-free optical image slice with a length and width of 100 pixels each centered on this point and a resolution of 10 meters; Step S34: Deploy the binary classification model in step S32 and use the model to perform predictive inference on all the Sentinel-2 cloud-free optical image slices in step S33, retain the part predicted as the methane leakage area, and mark the corresponding candidate methane leakage points as pending methane leakage points; Step S4: Using the plume recognition algorithm based on pixel connectivity to segment the methane plume from the methane concentration denoised data in step S2. The step S4 further includes: Step S41: Binarize the methane concentration noise-reduced data described in step S2. Set the pixels with methane concentration greater than Δα described in step S21 to 1, and set other pixels to 0 to obtain the binarized methane concentration data; t1 ​ Step S42: Using a convolution kernel of size k×k to convolve the methane concentration binary data image to obtain a neighborhood connectivity number result image of size (w-k+1)×(h-k+1), where w and h are the width and height of the original methane concentration binary data image respectively. The value of each pixel in the image is the number of non-zero pixel points in the k-neighborhood around the corresponding pixel point in the original methane concentration binary data image. Using the threshold n to extract the pixel positions in the neighborhood connectivity number result image with values greater than n to obtain multiple plume patches; Step S43: Using the 8-neighborhood connectivity domain algorithm to cluster all the plume patches to obtain all the candidate plume patches; Step S44: Traverse the candidate plume patches, and filter out the candidate plume patches whose methane concentrations are all less than the threshold Δα described in step S33. t2 Filter out the candidate plume patches that contain m pending methane leakage points in step S34; Mark the candidate plume patches containing the pending methane leakage points as methane plume masks; Filter out the candidate plume patches that do not contain the pending methane leakage points; Filter out the pending methane leakage points that do not belong to any candidate plume patches; Step S45: Traverse the methane plume mask, index the methane concentration values in the methane concentration basic data map described in step S1 according to the covered range, and fill the plume mask range to obtain all methane plumes. Step S5: Superimpose the methane plumes with the wind speed data and atmospheric pressure data, estimate the methane plume emissions using the IME algorithm, and finally conduct error analysis.

Images