A dual-time-phase sentinel-2 methane plume dataset construction method and a methane plume detection method

Abstract

The application discloses a greenhouse gas monitoring technical field, and relates to a kind of two-phase Sentinel-2 Methane Plume Dataset construction method and Methane Plume detection method.The method comprises the following steps: according to the obtained pure methane concentration field image set, construct relative plume density matrix, and the reflectivity of shortwave infrared B12 band of target area image at t time is physically attenuated and injected, according to the difference between shortwave infrared B11 band and B12 band of target area image at t-1 time, and the difference between shortwave infrared B11 band and B12 band of target area image at t time after physical attenuation injection, calculate methane index change matrix and standardize, obtain normalized methane index change matrix;The channel dimension of target area image at t time, target area image at t-1 time and normalized methane index change matrix is spliced, and methane plume sample is obtained;Relative plume density matrix is binarized to obtain plume true value label matrix of methane plume sample.

Description

Technical Field

[0001] This invention relates to the field of greenhouse gas monitoring technology, and in particular to a method for constructing a dual-temporal Sentinel-2 methane plume dataset and a method for detecting methane plumes. Background Technology

[0002] With the increasing severity of global climate change, high-resolution monitoring of industrial methane leaks using remote sensing satellites like Sentinel-2 has become crucial. Methane exhibits specific absorption characteristics in the short-wave infrared band, such as the B11 and B12 bands of Sentinel-2. However, because this absorption signal is extremely weak and easily masked by natural variations in surface albedo and complex background noise such as cloud shadows, single-phase detection often faces a very high false alarm rate.

[0003] Introducing deep learning for dual-temporal change detection is an effective approach to solving this problem, but this solution faces two major technical bottlenecks. On the one hand, high-quality training samples are scarce and the synthesis methods are distorted, resulting in an extreme scarcity of real pixel-level labeled methane leak data. Existing data augmentation and synthesis methods mostly rely on idealized two-dimensional Gaussian plume models, generating plumes with smooth edges and uniform shapes, lacking the hydrodynamic features of turbulent tearing, gust gathering, and centerline oscillation in the real atmospheric boundary layer. Such distorted datasets lead to extremely poor generalization ability of deep learning models under real complex weather conditions. On the other hand, weak signals are difficult to extract effectively by conventional networks. Methane plumes account for a very small proportion of the entire image, resulting in severe class imbalance and huge background noise. Conventional convolutional neural networks, such as traditional CNN networks, struggle to adaptively focus on specific bands, such as the weak physical response of the SWIR band, when processing multispectral dual-temporal data, and their limited local receptive fields lead to a high false negative rate. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing a dual-temporal Sentinel-2 methane plume dataset and a method for detecting methane plumes. By extracting the physical manifold boundary of real large eddy simulation (LES) data, a synthetic dataset containing real hydrodynamic features is generated. By setting up a methane plume detection model that includes a spectral attention module, an encoder module, and a decoder module, automatic detection of methane plumes under complex surface backgrounds is achieved.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for constructing a dual-temporal Sentinel-2 methane plume dataset, comprising:

[0007] Based on the obtained pure methane concentration field image set, a relative plume density matrix is ​​constructed;

[0008] Based on the relative plume density matrix, Shortwave infrared of the target area image at any time The reflectivity of the band is physically attenuated and injected to obtain the result after physical attenuation injection. Image of the target area at any given time;

[0009] according to Shortwave infrared of the target area image at any time bands and Differences in bands, and physical attenuation after injection Shortwave infrared of the target area image at any time bands and The methane index variation matrix was calculated based on the differences in the bands and then standardized to obtain the normalized methane index variation matrix.

[0010] Will Image of target area at any time The target region image at any time and the normalized methane index change matrix are stitched together along the channel dimension to obtain the methane plume sample;

[0011] The relative plume density matrix is ​​binarized to obtain the plume truth label matrix of the methane plume sample;

[0012] The methane plume dataset is obtained based on the methane plume samples and their true value label matrix.

[0013] Optionally, constructing the relative plume density matrix based on the acquired pure methane concentration field image set includes:

[0014] Select one pure methane concentration field image from the acquired pure methane concentration field image set for feature extraction, and obtain the methane plume variation coefficient, effective active region covariance matrix, anisotropic stretching ratio, basic eddy spatial scale, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image.

[0015] Based on the initial methane plume coordinates, effective active region covariance matrix, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image, a multidimensional physically constrained methane concentration field is generated.

[0016] The methane concentration field is edge-modulated based on the anisotropic stretching ratio, the basic eddy space scale, and the methane plume variation coefficient to obtain the edge-modulated methane concentration field.

[0017] The edge-modulated methane concentration field is normalized to obtain the relative plume density matrix.

[0018] Optionally, the step of selecting any one pure methane concentration field image from the acquired pure methane concentration field image set for feature extraction, and obtaining the methane plume variation coefficient, effective active region covariance matrix, anisotropic stretching ratio, basic eddy spatial scale, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image, including:

[0019] Obtain the effective active region of the methane plume from the pure methane concentration field image; where the methane concentration is higher than... The maximum methane concentration was determined as the effective active region of the methane plume. This is the default value;

[0020] The ratio of the standard deviation of methane concentration to the mean methane concentration within the effective active region of the methane plume is calculated to obtain the coefficient of variation of the methane plume in the pure methane concentration field image.

[0021] Calculate the effective active region covariance matrix of the methane plume;

[0022] The anisotropic stretching ratio, the spatial scale of the basic eddy current, and the standard deviation of the centerline oscillation are calculated based on the covariance matrix of the effective active region.

[0023] Random sampling was performed based on a one-dimensional concentration fluctuation empirical model to obtain the longitudinal gust amplitude and longitudinal gust period;

[0024] The basic eddy current spatial scale is obtained by the following formula:

[0025] ,

[0026] in, Indicates the spatial scale of the basic eddy current. The eigenvalues ​​of the covariance matrix representing the effective active region of the methane plume, perpendicular to the prevailing wind direction;

[0027] The standard deviation of the centerline oscillation is obtained by the following formula:

[0028] ,

[0029] in, This represents the standard deviation of the centerline oscillation. Indicates the standard deviation of the grid scale. , The pixel resolution of the pure methane concentration field image;

[0030] The anisotropic stretch ratio is obtained by the following formula:

[0031] ,

[0032] in, Indicates the anisotropic stretch ratio. The eigenvalues ​​of the covariance matrix representing the prevailing wind direction of the effective active region of the methane plume.

[0033] Optionally, the step of generating a multidimensional physically constrained methane concentration field based on the initial methane plume coordinates, effective active region covariance matrix, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image includes:

[0034] Calculate the time series sequence of methane plume coordinate offset based on the covariance matrix of the effective active region, the standard deviation of the centerline oscillation, and the initial methane plume coordinates;

[0035] Based on the time series sequence of methane plume coordinate offset, the time series sequence of horizontal diffusion standard deviation of methane plume is calculated using the Gaussian coefficient diffusion formula;

[0036] Based on the time series sequence of methane plume coordinate offset and the initial center coordinates of the methane plume, the time series sequence of methane plume center coordinates is obtained;

[0037] Based on the time series of methane plume center coordinates, the horizontal diffusion standard deviation of the methane plume, and the boundary coordinates of the pure methane concentration field image, the time series of methane plume mass fraction is calculated using the cumulative distribution function (CDF).

[0038] Based on the longitudinal gust amplitude and longitudinal gust period, the gas mass time series of the methane plume after superimposed gust effect is calculated by using a sinusoidal perturbation function;

[0039] A multidimensional physically constrained methane concentration field is generated based on the time series sequence of methane plume mass fraction and the time series sequence of methane plume gas mass after superimposing gust effects.

[0040] The time series sequence of methane plume coordinate offset is obtained by the following formula:

[0041] ,

[0042] ,

[0043] in, express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis coordinate offset, Represents the natural base. This represents the time step between two adjacent moments. This represents standard normally distributed noise. express The memory time constant of the axis, , This represents the function that takes the maximum value. express The wind speed component of the axis, Indicates the average grid radius. , This represents the average eigenvalue of the covariance matrix of the effective active regions. This represents the pixel resolution of the pure methane concentration field image. This represents the standard deviation of the centerline oscillation. express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis coordinate offset, express The memory time constant of the axis, , express Wind speed component of the axis;

[0044] The time series sequence of the horizontal diffusion standard deviation of the methane plume is obtained by the following formula:

[0045] ,

[0046] in, express The standard deviation of horizontal diffusion of the methane plume at time t. express The physical absolute distance of the methane plume at any given moment. , Indicates the preset wind speed. This represents the wind speed scaling factor. ;

[0047] The time series of the methane plume center coordinates is obtained by the following formula:

[0048] ,

[0049] ,

[0050] in, express methane plume at a moment Axis center coordinates, Indicates the initial state of the methane plume. Axis center coordinates, express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis center coordinates, Indicates the initial state of the methane plume. Axis center coordinates, express methane plume at a moment Axis coordinate offset;

[0051] The time series sequence of the mass fraction of the methane plume is obtained by the following formula:

[0052] ,

[0053] ,

[0054] ,

[0055] in, express At what time is the methane plume at coordinates? Mass fraction at the point, express The methane plume falls within the interval at that moment The probability, The coordinates represent the upper boundary of the pure methane concentration field image. This represents the lower boundary coordinates of the pure methane concentration field image. express The methane plume falls within the interval at that moment The probability, The coordinates of the left boundary of the pure methane concentration field image are represented. The coordinates of the right boundary of the pure methane concentration field image are represented. Represents the cumulative distribution function. Indicates from the start time to methane plume at a moment The set of coordinates of the axis center. Indicates from the start time to methane plume at a moment The set of coordinates of the axis center;

[0056] The time series sequence of methane plume gas mass after superimposed gust effect is obtained by the following formula:

[0057] ,

[0058] in, express The mass of the methane plume after the superposition of gust effects at any given moment. This represents the mass of the methane plume under absolutely constant wind speed. , Indicates the preset emission rate. This indicates the molar mass of methane. Indicates the longitudinal gust amplitude. Indicates the longitudinal gust period, Represents the sine function;

[0059] The multidimensional physical constraint of the methane concentration field is obtained by the following formula:

[0060] ,

[0061] ,

[0062] in, express The methane plume at time [time] at coordinates methane concentration at that location The physical area representing the concentration field image of pure methane. Indicates the methane plume diffusion cessation time. Represents the methane concentration field median coordinate The methane concentration at that location.

[0063] Optionally, the step of edge-modulating the methane concentration field based on the anisotropic stretching ratio, the basic eddy space scale, and the methane plume variation coefficient to obtain the edge-modulated methane concentration field includes:

[0064] Calculate the core protection mask based on the methane concentration field;

[0065] Based on the core protection mask and the variation coefficient of the methane plume, the effective turbulence intensity matrix is ​​extracted;

[0066] The turbulence disturbance factor is calculated based on the anisotropic stretching ratio and the basic eddy space scale.

[0067] The edge-modulated methane concentration field is calculated based on the effective turbulence intensity matrix, methane concentration field, and turbulence perturbation factor.

[0068] The core protective mask is obtained through the following formula:

[0069] ,

[0070] in, Indicates the core protective mask of the main trunk. median coordinate The protection coefficient at the location, Represents the methane concentration field median coordinate methane concentration at that location Represents the methane concentration field The maximum methane concentration in;

[0071] The effective turbulence intensity matrix is ​​obtained by the following formula:

[0072] ,

[0073] in, Representing coordinates Effective turbulence intensity at that location, Indicates the overall basic tear strength. , This represents the function that takes the minimum value. This represents the function that takes the maximum value. This indicates the preset turbulence intensity. Indicates the coefficient of variation of the methane plume;

[0074] The turbulence disturbance factor is obtained by the following formula:

[0075] ,

[0076] ,

[0077] ,

[0078] ,

[0079] ,

[0080] ,

[0081] ,

[0082] ,

[0083] in, This represents the turbulence disturbance factor. This represents the mixing matrix of multi-scale, along-wind, anisotropic random perturbation fields. Indicates rotation operation. Indicates wind direction and The included angle of the axis, This represents the mixing matrix of anisotropic random perturbation fields at multiple scales. This indicates low-frequency scale characteristics. Indicates high-frequency scale characteristics. This represents the anisotropic Gaussian space convolution operation. This represents the preset white noise matrix. Indicates the standard deviation of the minor axis. Indicates the standard deviation of the major axis. Represents the integral scale of turbulence. Indicates the spatial scale of the basic eddy current. Indicates the anisotropic stretch ratio. Indicates the preset wind speed;

[0084] The edge-modulated methane concentration field is obtained by the following formula:

[0085] ,

[0086] ,

[0087] in, Represents the methane concentration field after edge modulation median coordinate methane concentration at that location, Represents the integrated modulation multiplier matrix. Represents the integrated modulation multiplier matrix median coordinate The overall modulation multiplier at the location, This represents the effective turbulence intensity matrix.

[0088] Optionally, the pair Shortwave infrared of the target area image at any time The physical attenuation injection of reflectivity in the band is achieved through the following formula:

[0089] ,

[0090] in, Indicates physical decay injection Shortwave infrared of the target area image at any time Reflectivity of the band express Shortwave infrared of the target area image at any time Reflectivity of the band This represents the natural exponential function. Represents the random optical attenuation coefficient. This represents the relative plume density matrix.

[0091] Optionally, the normalized methane index change matrix is ​​obtained by the following formula:

[0092] ,

[0093] ,

[0094] in, This represents the normalized methane index variation matrix. Represents the methane index variation matrix. Represents the matrix of methane index changes The mean, Represents the matrix of methane index changes standard deviation As a preset value, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix.

[0095] Secondly, the present invention provides a method for detecting a dual-temporal Sentinel-2 methane plume, comprising:

[0096] Input the methane plume sample into the methane plume detection model:

[0097] The methane plume samples were subjected to channel attention weighting using a spectral attention module to obtain a weighted methane plume feature vector.

[0098] The weighted methane plume feature vector is input into the pyramid vision Transformer encoder for multi-scale feature extraction to obtain a multi-scale high-dimensional feature map.

[0099] The multi-scale high-dimensional feature map is input into the U-Net decoder for feature fusion to obtain the fused feature map.

[0100] The fused feature map is input into the output layer for methane concentration detection to obtain the methane concentration detection result. The methane plume detection model uses the bi-temporal Sentinel-2 methane plume dataset construction method provided in the first aspect to construct the training dataset during the training process.

[0101] Optionally, the spectral attention module includes a globally average pooling layer, a multilayer perceptron, and a sigmoid activation layer connected in sequence;

[0102] The weighted methane plume eigenvector is obtained by the following formula:

[0103] ,

[0104] ,

[0105] in, Represents the methane plume weight vector. This represents the Sigmoid activation function. Represents the ReLU activation function. This represents the first learnable weight matrix of the multilayer perceptron. This represents the second learnable weight matrix of the multilayer perceptron. This indicates a global average pooling operation. This indicates a methane plume sample. This represents the weighted methane plume eigenvector.

[0106] Optionally, the joint loss function of the methane plume detection model is as follows:

[0107] ,

[0108] ,

[0109] ,

[0110] ,

[0111] in, Indicates joint loss, Indicates the preset weight. Indicates focal loss. This represents the total number of samples from the methane plume. , Indicates preset parameters. Indicates the detection confidence level. , This represents the natural exponential function. Represents the binary cross-entropy loss. This indicates the methane concentration detection results of the methane plume sample. The plume truth label matrix represents the methane plume samples. express The Middle The truth value label for the feather flow is 1 pixel. express The Middle methane concentration detection results per pixel, This represents the Dice coefficient loss. This is the default value.

[0112] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0113] 1. The proposed dual-temporal Sentinel-2 methane plume dataset construction method fully utilizes the pure methane concentration field image set generated by Les large eddy simulation, avoiding the complex calculations required for a large number of simulations. This not only greatly saves the time cost required to construct a large-scale dataset, but also ensures the physical realism of the simulated plume in the constructed dataset. Based on reducing the false alarm rate of the background surface based on dual-temporal simulation, the method improves the extraction efficiency of plume features by adding a methane index change channel as a physical prior channel. Due to the large scale of the dataset and the high physical realism of the simulated methane plume, its generalization ability in real-world scenarios can also be enhanced.

[0114] 2. The dual-temporal Sentinel-2 methane plume detection method proposed in this invention improves the detection process's ability to capture long-distance features and extract deep features by introducing a spectral attention module and combining it with a pyramid vision Transformer encoder and a U-Net decoder. Attached Figure Description

[0115] Figure 1 This is a schematic diagram of the method for constructing a dual-temporal Sentinel-2 methane plume dataset provided in an embodiment of the present invention;

[0116] Figure 2 This is a schematic diagram of the spectral attention module structure provided in an embodiment of the present invention;

[0117] Figure 3 ROC curves provided for embodiments of the present invention;

[0118] Figure 4 This is a statistical diagram illustrating the success rate of methane plume detection provided in an embodiment of the present invention. Detailed Implementation

[0119] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0120] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0121] Example 1

[0122] This invention discloses a method for constructing a two-phase Sentinel-2 methane plume dataset, with reference to... Figure 1 As shown, the specific steps include the following:

[0123] S1. Construct a relative plume density matrix based on the obtained pure methane concentration field image set;

[0124] S2, based on the relative plume density matrix, for Shortwave infrared of the target area image at any time The reflectivity of the band is physically attenuated and injected to obtain the result after physical attenuation injection. Image of the target area at any given time;

[0125] S3, according to Shortwave infrared of the target area image at any time bands and Differences in wavebands, and physical attenuation after injection Shortwave infrared of the target area image at any time bands and The methane index variation matrix was calculated based on the differences in the bands and then standardized to obtain the normalized methane index variation matrix.

[0126] S4, Image of target area at any time The target region image at any time and the normalized methane index change matrix are stitched together along the channel dimension to obtain the methane plume sample;

[0127] S5, perform binarization on the relative plume density matrix to obtain the plume truth label matrix of the methane plume sample;

[0128] S6. Based on the methane plume samples and their plume truth label matrix, the methane plume dataset is obtained.

[0129] In step S1, this embodiment collects a set of extremely high spatial resolution pure methane concentration field images generated by real large eddy simulation as a template library for constructing the relative plume density matrix. Using the template library maximizes the realism of the plume simulation, and the range of macroscopic meteorological parameters such as wind speed, wind direction, and emission rate can be set according to actual use. The construction process includes:

[0130] S1.1 Select one pure methane concentration field image from the acquired pure methane concentration field image set for feature extraction to obtain the methane plume variation coefficient, effective active region covariance matrix, anisotropic stretching ratio, basic eddy spatial scale, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image.

[0131] S1.2, Based on the initial methane plume coordinates, effective active region covariance matrix, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image, a multidimensional physically constrained methane concentration field is generated.

[0132] S1.3, Based on the anisotropic stretching ratio, the basic eddy space scale and the methane plume variation coefficient, the methane concentration field is edge-modulated to obtain the edge-modulated methane concentration field.

[0133] S1.4, normalize the edge-modulated methane concentration field to obtain the relative plume density matrix.

[0134] In step S1.1, a single pure methane concentration field image is randomly selected from the acquired pure methane concentration field image set for feature extraction, yielding the methane plume variation coefficient, effective active region covariance matrix, anisotropic stretching ratio, basic eddy spatial scale, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image, including:

[0135] S1.1.1, Obtain the effective active region of the methane plume from the pure methane concentration field image; where the methane concentration is higher than... The maximum methane concentration was determined as the effective active region of the methane plume. As a preset value, this embodiment sets... The value is set to 5 to filter out redundant pixels with a concentration value lower than 5% of the maximum concentration.

[0136] S1.1.2, calculate the ratio of the standard deviation of methane concentration to the mean methane concentration within the effective active region of the methane plume to obtain the coefficient of variation of the methane plume in the pure methane concentration field image. In this embodiment, the following settings are provided. It is calculated by dividing the standard deviation by the mean and then multiplying by a coefficient of 0.5.

[0137] S1.1.3, Calculate the effective active region covariance matrix of the methane plume;

[0138] S1.1.4, Calculate the anisotropic stretching ratio, basic eddy spatial scale, and centerline oscillation standard deviation based on the effective active region covariance matrix:

[0139] The basic eddy current spatial scale is obtained by the following formula:

[0140] ,

[0141] in, Indicates the spatial scale of the basic eddy current. The eigenvalues ​​of the covariance matrix representing the effective active region of the methane plume, perpendicular to the prevailing wind direction;

[0142] The standard deviation of the centerline oscillation is obtained by the following formula:

[0143] ,

[0144] in, This represents the standard deviation of the centerline oscillation. Indicates the standard deviation of the grid scale. , The pixel resolution of the pure methane concentration field image;

[0145] The anisotropic stretch ratio is obtained by the following formula:

[0146] ,

[0147] in, Indicates the anisotropic stretch ratio. The eigenvalues ​​of the covariance matrix representing the prevailing wind direction of the effective active region of the methane plume.

[0148] S1.1.5, Random sampling is performed based on a one-dimensional concentration fluctuation empirical model to obtain the longitudinal gust amplitude and longitudinal gust period.

[0149] In step S1.2, generating a multidimensional physically constrained methane concentration field based on the initial methane plume coordinates, effective active region covariance matrix, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image includes:

[0150] S1.2.1, based on the covariance matrix of the effective active region, the standard deviation of the centerline oscillation, and the initial methane plume coordinates, the time series sequence of methane plume coordinate offset is calculated using the Ornstein-Uhlenbeck process perturbation formula:

[0151] ,

[0152] ,

[0153] in, express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis coordinate offset, Represents the natural base. This represents the time step between two adjacent moments. This represents standard normally distributed noise. express The memory time constant of the axis, , This represents the function that takes the maximum value. express The wind speed component of the axis, Indicates the average grid radius. , This represents the average eigenvalue of the covariance matrix of the effective active regions. This represents the pixel resolution of the pure methane concentration field image. This represents the standard deviation of the centerline oscillation. express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis coordinate offset, express The memory time constant of the axis, , express Wind speed component of the axis;

[0154] S1.2.2, Based on the time series of methane plume coordinate offset, the time series of horizontal diffusion standard deviation of the methane plume is calculated using the Gaussian coefficient diffusion formula:

[0155] ,

[0156] in, express The standard deviation of horizontal diffusion of the methane plume at time t. express The physical absolute distance of the methane plume at any given moment. , Indicates the preset wind speed. This represents the wind speed scaling factor. ;

[0157] S1.2.3, Based on the time series sequence of methane plume coordinate offset and the initial center coordinates of the methane plume, the time series sequence of methane plume center coordinates is obtained:

[0158] ,

[0159] ,

[0160] in, express methane plume at a moment Axis center coordinates, Indicates the initial state of the methane plume. Axis center coordinates, express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis center coordinates, Indicates the initial state of the methane plume. Axis center coordinates, express methane plume at a moment Axis coordinate offset;

[0161] S1.2.4, based on the time series of the methane plume center coordinates, the horizontal diffusion standard deviation of the methane plume, and the boundary coordinates of the pure methane concentration field image, the time series of the methane plume mass fraction is calculated using the cumulative distribution function (CDF):

[0162] ,

[0163] ,

[0164]

[0165] in, express At what time is the methane plume at coordinates? Mass fraction at the point, express The methane plume falls within the interval at that moment The probability, The coordinates represent the upper boundary of the pure methane concentration field image. This represents the lower boundary coordinates of the pure methane concentration field image. express The methane plume falls within the interval at that moment The probability, The coordinates of the left boundary of the pure methane concentration field image are represented. The coordinates of the right boundary of the pure methane concentration field image are represented. Represents the cumulative distribution function. Indicates from the start time to methane plume at a moment The set of coordinates of the axis center. Indicates from the start time to methane plume at a moment The set of coordinates of the axis center;

[0166] S1.2.5, Based on the longitudinal gust amplitude and longitudinal gust period, the gas mass time series of the methane plume after superimposed gust effect is calculated using a sinusoidal perturbation function:

[0167] ,

[0168] in, express The mass of the methane plume after the superposition of gust effects at any given moment. This indicates that, under ideal conditions with absolutely constant wind speed, the constant mass of methane gas that should be uniformly emitted from the methane cloud at each time step is determined by the input emission rate. , Indicates the preset emission rate. This indicates the molar mass of methane. Indicates the longitudinal gust amplitude. This indicates the longitudinal gust period.

[0169] S1.2.6, Based on the time series sequence of methane plume mass fraction and the time series sequence of methane plume gas mass after superimposing gust effects, a multidimensional physically constrained methane concentration field is generated:

[0170] ,

[0171] ,

[0172] in, express The methane plume at time [time] at coordinates methane concentration at that location The physical area representing the concentration field image of pure methane. Indicates the methane plume diffusion cessation time. Represents the methane concentration field median coordinate The methane concentration at that location.

[0173] In step S1.3, the step of edge-modulating the methane concentration field based on the anisotropic stretching ratio, the basic eddy space scale, and the methane plume variation coefficient to obtain the edge-modulated methane concentration field includes:

[0174] S1.3.1, Calculate the core protection mask based on the methane concentration field:

[0175] ,

[0176] in, Indicates the core protective mask of the main trunk. median coordinate The protection coefficient at the location, Represents the methane concentration field median coordinate methane concentration at that location Represents the methane concentration field The maximum methane concentration in [the sample / sample].

[0177] S1.3.2, Based on the core protection mask and the variation coefficient of the methane plume, extract the effective turbulence intensity matrix:

[0178] ,

[0179] in, Representing coordinates Effective turbulence intensity at that location, Indicates the overall basic tear strength. , This represents the function that takes the minimum value. This represents the function that takes the maximum value. This indicates the preset turbulence intensity. This represents the coefficient of variation of the methane plume.

[0180] S1.3.3, Calculate the turbulence disturbance factor based on the anisotropic stretching ratio and the basic eddy space scale:

[0181] ,

[0182] ,

[0183] ,

[0184] ,

[0185] ,

[0186] ,

[0187] ,

[0188] ,

[0189] in, This represents the turbulence disturbance factor. This represents the mixing matrix of multi-scale, along-wind, anisotropic random perturbation fields. Indicates rotation operation. Indicates wind direction and The included angle of the axis, This represents the mixing matrix of multi-scale anisotropic random perturbation fields. This indicates low-frequency scale characteristics. Indicates high-frequency scale characteristics. This represents the anisotropic Gaussian space convolution operation. This represents the preset white noise matrix. Indicates the standard deviation of the minor axis. Indicates the standard deviation of the major axis. Represents the integral scale of turbulence. Indicates the spatial scale of the basic eddy current. Indicates the anisotropic stretch ratio. This indicates the preset wind speed.

[0190] S1.3.4, Calculate the edge-modulated methane concentration field based on the effective turbulence intensity matrix, methane concentration field, and turbulence perturbation factor:

[0191] ,

[0192] ,

[0193] in, Represents the methane concentration field after edge modulation median coordinate methane concentration at that location Represents the integrated modulation multiplier matrix. Represents the integrated modulation multiplier matrix median coordinate The overall modulation multiplier at the location, This represents the effective turbulence intensity matrix.

[0194] In step S2, based on the Beer-Lambert law, the generated relative plume density matrix is ​​used... With random optical attenuation coefficient right Shortwave infrared of the target area image at any time The reflectivity of the band is physically attenuated and injected. After physical attenuation injection... Shortwave infrared of the target area image at any time The reflectivity of the band is as follows:

[0195] ,

[0196] in, Indicates physical decay injection Shortwave infrared of the target area image at any time Reflectivity of the band express Shortwave infrared of the target area image at any time Reflectivity of the band This represents the natural exponential function. Represents the random optical attenuation coefficient. This represents the relative plume density matrix.

[0197] In step S3, the normalized methane index change matrix is ​​obtained by the following formula:

[0198] ,

[0199] ,

[0200] in, This represents the normalized methane index variation matrix. Represents the methane index variation matrix. Represents the matrix of methane index changes The mean, Represents the matrix of methane index changes standard deviation This is a preset value used for data protection to prevent the denominator from being 0. express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix.

[0201] In step S4, by setting a hard threshold of 0.02, the relative plume density matrix is ​​binarized to generate the plume truth label matrix Mask.

[0202] Example 2

[0203] Based on the same inventive concept as Example 1, this embodiment of the invention discloses a dual-temporal Sentinel-2 methane plume detection method, comprising:

[0204] Input the methane plume sample into the methane plume detection model:

[0205] The methane plume samples were subjected to channel attention weighting using a spectral attention module to obtain a weighted methane plume feature vector.

[0206] The weighted methane plume feature vector is input into the pyramid vision Transformer encoder for multi-scale feature extraction to obtain a multi-scale high-dimensional feature map.

[0207] The multi-scale high-dimensional feature map is input into the U-Net decoder for feature fusion to obtain the fused feature map.

[0208] The fused feature map is input into the output layer for methane concentration detection to obtain the methane concentration detection result.

[0209] The methane plume detection model uses the dual-temporal Sentinel-2 methane plume dataset construction method described in Example 1 to construct the training dataset during the training process.

[0210] In this embodiment, the methane plume samples can be generated using the dual-temporal Sentinel-2 methane plume dataset construction method in Example 1, or they can be real methane plume images.

[0211] refer to Figure 2 As shown, in this embodiment, the spectral attention module includes a globally average pooling layer, a multilayer perceptron, and a sigmoid activation layer connected in sequence. The spectral attention module first compresses the spatial information of each channel into a one-dimensional global descriptor using global average pooling; then, it further compresses the spatial information by including the dimensionality reduction ratio... The two-layer multilayer perceptron (MLP) learns the nonlinear correlation between channels and outputs the channel weight vector after passing through a sigmoid activation layer. Finally, the channel weight vector With methane plume sample Channel-level multiplication is performed to adaptively weight the weak methane characteristic channels, resulting in a weighted methane plume feature vector. This weighted methane plume feature vector is obtained using the following formula:

[0212] ,

[0213] ,

[0214] in, Represents the methane plume weight vector. This represents the Sigmoid activation function. Represents the ReLU activation function. This represents the first learnable weight matrix of the multilayer perceptron. This represents the second learnable weight matrix of the multilayer perceptron. This indicates a global average pooling operation. This indicates a methane plume sample. This represents the weighted methane plume eigenvector.

[0215] The Pyramid Visual Transformer encoder captures the long-range spatial dependencies of methane plumes in the atmospheric boundary layer caused by serpentine undulations and turbulent tearing through a global self-attention mechanism, and weights the methane plume feature vectors. Feature extraction at different depths is performed, outputting four sets of high-dimensional feature maps at multiple scales. , respectively corresponding to the original resolution It achieves efficient extraction from shallow details to deep global semantics.

[0216] The U-Net decoder is used to recover spatial resolution for the deepest features output by the encoder. First, upsampling is performed using a transposed convolution with a stride of 2 to obtain the magnified features. and correlate them with the coding features of the corresponding level. The features are concatenated along the channel dimension; the concatenated features are then upsampled by transposed convolution and concatenated with shallow features. Finally, the fused feature map upsampled to the original image size is input into the output layer for prediction to obtain the methane concentration detection result of the methane plume sample.

[0217] Because methane plumes occupy a very small percentage of pixels in the entire remote sensing image, resulting in severe class imbalance and easy confusion with bright surface noise, this embodiment employs a joint loss function combining Focal Loss and Dice coefficient loss during the network training phase. The focal loss can suppress false alarm errors caused by the high albedo of vast bare soil backgrounds, while the Dice coefficient loss extracts weak fragmented plume signals with extremely low total area proportions by forcing the model through spatial region alignment; the joint loss function The expression is as follows:

[0218] ,

[0219] ,

[0220] ,

[0221] ,

[0222] in, Indicates joint loss, Indicates the preset weight. Indicates focal loss. This represents the total number of samples from the methane plume. , Indicates preset parameters. Indicates the detection confidence level. , This represents the natural exponential function. Represents the binary cross-entropy loss. This indicates the methane concentration detection results of the methane plume sample. The plume truth label matrix represents the methane plume samples. express The Middle The truth value label for the feather flow is 1 pixel. express The Middle methane concentration detection results per pixel, This represents the Dice coefficient loss. This is the default value.

[0223] This embodiment uses the dataset constructed using the method in Example 1 as the training and validation set for the methane detection model. Figure 3 To perform detection and pixel-level evaluation on 2000 randomly selected samples from the validation set, the ROC curve relationship between the recall and false alarm rate of the methane detection model in the validation set was obtained. Figure 3 The ROC curve in the figure is used to demonstrate the ability of the methane plume detection model to balance recall and false positive rate. The vertical axis represents the true positive rate (TPR) of the methane plume detection model on the validation set, and the horizontal axis represents the false positive rate (FPR) of the methane plume detection model on the validation set. AUC represents the area under the ROC curve, which is an indicator for evaluating the performance of binary classification models. The closer it is to 1, the better the model's performance. The blue dot in the figure represents the target point, indicating that the false positive rate of the methane plume detection model is 1.3337% while maintaining an 85% recall rate.

[0224] This embodiment also uses real methane emission event samples to evaluate the detection performance of the methane detection model, obtaining the detection success rate for different emission rate ranges. This embodiment downloads a statistical table of methane emission events in the United States for 2020-2021 from the official Carbon Mapper website, then uses the statistical information in the table to retrieve events on Google Cloud Platform and exports Sentinel-2 dual-temporal data within a 2.56km × 2.56km range centered on the emission source. This data is used to construct a real methane emission event test set. The experimental success condition is the detection of continuous pixels representing methane emissions within 200m of the central emission source. (Reference) Figure 4 The figure shows the success rate of the methane plume detection model in detecting real emission events. Successful detection is defined as the detection of consecutive pixels within 200 meters of the leak source. The vertical axis represents the success rate of the methane plume detection model in real methane emission events, and the horizontal axis represents the emission rate range of real methane emission events. N in the bar chart represents the number of real emission events within the emission rate range. The overall average detection rate is 46.4%.

[0225] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0226] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0227] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0228] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0229] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for constructing a two-phase Sentinel-2 methane plume dataset, characterized in that, include: Based on the obtained pure methane concentration field image set, a relative plume density matrix is ​​constructed; Based on the relative plume density matrix, Shortwave infrared of the target area image at any time The reflectivity of the band is physically attenuated and injected to obtain the result after physical attenuation injection. Image of the target area at any given time; according to Shortwave infrared of the target area image at any time bands and Differences in bands, and physical attenuation after injection Shortwave infrared of the target area image at any time bands and The methane index variation matrix was calculated based on the differences in the bands and then standardized to obtain the normalized methane index variation matrix. Will Image of target area at any time The target region image at any time and the normalized methane index change matrix are stitched together along the channel dimension to obtain the methane plume sample; The relative plume density matrix is ​​binarized to obtain the plume truth label matrix of the methane plume sample; The methane plume dataset is obtained based on the methane plume samples and their true value label matrix.

2. The method for constructing a dual-temporal Sentinel-2 methane plume dataset according to claim 1, characterized in that, The step of constructing a relative plume density matrix based on the acquired pure methane concentration field image set includes: Select one pure methane concentration field image from the acquired pure methane concentration field image set for feature extraction, and obtain the methane plume variation coefficient, effective active region covariance matrix, anisotropic stretching ratio, basic eddy spatial scale, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image. Based on the initial methane plume coordinates, effective active region covariance matrix, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image, a multidimensional physically constrained methane concentration field is generated. The methane concentration field is edge-modulated based on the anisotropic stretching ratio, the basic eddy space scale, and the methane plume variation coefficient to obtain the edge-modulated methane concentration field. The edge-modulated methane concentration field is normalized to obtain the relative plume density matrix.

3. The method for constructing a dual-temporal Sentinel-2 methane plume dataset according to claim 2, characterized in that, The process involves selecting one pure methane concentration field image from the acquired pure methane concentration field image set for feature extraction. This yields the methane plume variation coefficient, effective active region covariance matrix, anisotropic stretching ratio, basic eddy spatial scale, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period of the pure methane concentration field image, including: Obtain the effective active region of the methane plume from the pure methane concentration field image; where the methane concentration is higher than... The maximum methane concentration was determined as the effective active region of the methane plume. This is the default value; The ratio of the standard deviation of methane concentration to the mean methane concentration within the effective active region of the methane plume is calculated to obtain the coefficient of variation of the methane plume in the pure methane concentration field image. Calculate the effective active region covariance matrix of the methane plume; The anisotropic stretching ratio, the spatial scale of the basic eddy current, and the standard deviation of the centerline oscillation are calculated based on the covariance matrix of the effective active region. Random sampling was performed based on a one-dimensional concentration fluctuation empirical model to obtain the longitudinal gust amplitude and longitudinal gust period; The basic eddy current spatial scale is obtained by the following formula: ， in, Indicates the spatial scale of the basic eddy current. The eigenvalues ​​of the covariance matrix representing the effective active region of the methane plume, perpendicular to the prevailing wind direction; The standard deviation of the centerline oscillation is obtained by the following formula: ， in, This represents the standard deviation of the centerline oscillation. Indicates the standard deviation of the grid scale. , The pixel resolution of the pure methane concentration field image; The anisotropic stretch ratio is obtained by the following formula: ， in, Indicates the anisotropic stretch ratio. The eigenvalues ​​of the covariance matrix representing the prevailing wind direction of the effective active region of the methane plume.

4. The method for constructing a dual-temporal Sentinel-2 methane plume dataset according to claim 2, characterized in that, The process of generating a multidimensional physically constrained methane concentration field based on the initial methane plume coordinates, effective active region covariance matrix, centerline oscillation standard deviation, longitudinal gust amplitude, and longitudinal gust period from a pure methane concentration field image includes: Calculate the time series sequence of methane plume coordinate offset based on the covariance matrix of the effective active region, the standard deviation of the centerline oscillation, and the initial methane plume coordinates; Based on the time series sequence of methane plume coordinate offset, the time series sequence of horizontal diffusion standard deviation of methane plume is calculated using the Gaussian coefficient diffusion formula; Based on the time series sequence of methane plume coordinate offset and the initial center coordinates of the methane plume, the time series sequence of methane plume center coordinates is obtained; Based on the time series of methane plume center coordinates, the horizontal diffusion standard deviation of the methane plume, and the boundary coordinates of the pure methane concentration field image, the time series of methane plume mass fraction is calculated using the cumulative distribution function (CDF). Based on the longitudinal gust amplitude and longitudinal gust period, the gas mass time series of the methane plume after superimposed gust effect is calculated by using a sinusoidal perturbation function; A multidimensional physically constrained methane concentration field is generated based on the time series sequence of methane plume mass fraction and the time series sequence of methane plume gas mass after superimposing gust effects. The time series sequence of methane plume coordinate offset is obtained by the following formula: ， ， in, express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis coordinate offset, Represents the natural base. This represents the time step between two adjacent moments. This represents standard normally distributed noise. express The memory time constant of the axis, , This represents the function that takes the maximum value. express The wind speed component of the axis, Indicates the average grid radius. , This represents the average eigenvalue of the covariance matrix of the effective active regions. This represents the pixel resolution of the pure methane concentration field image. This represents the standard deviation of the centerline oscillation. express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis coordinate offset, express The memory time constant of the axis, , express Wind speed component of the axis; The time series sequence of the horizontal diffusion standard deviation of the methane plume is obtained by the following formula: ， in, express The standard deviation of horizontal diffusion of the methane plume at time t. express The physical absolute distance of the methane plume at any given moment. , Indicates the preset wind speed. This represents the wind speed scaling factor. ; The time series of the methane plume center coordinates is obtained by the following formula: ， ， in, express methane plume at a moment Axis center coordinates, Indicates the initial state of the methane plume. Axis center coordinates, express methane plume at a moment Axis coordinate offset, express methane plume at a moment Axis center coordinates, Indicates the initial state of the methane plume. Axis center coordinates, express methane plume at a moment Axis coordinate offset; The time series sequence of the mass fraction of the methane plume is obtained by the following formula: ， ， ， in, express At what time is the methane plume at coordinates? Mass fraction at the point, express The methane plume falls within the interval at that moment The probability, The coordinates represent the upper boundary of the pure methane concentration field image. This represents the lower boundary coordinates of the pure methane concentration field image. express The methane plume falls within the interval at that moment The probability, The coordinates of the left boundary of the pure methane concentration field image are represented. The coordinates of the right boundary of the pure methane concentration field image are represented. Represents the cumulative distribution function. Indicates from the start time to methane plume at a moment The set of coordinates of the axis center. Indicates from the start time to methane plume at a moment The set of coordinates of the axis center; The time series sequence of methane plume gas mass after superimposed gust effect is obtained by the following formula: ， in, express The mass of the methane plume after the superposition of gust effects at any given moment. This represents the mass of the methane plume under absolutely constant wind speed. , Indicates the preset emission rate. This indicates the molar mass of methane. Indicates the longitudinal gust amplitude. Indicates the longitudinal gust period, Represents the sine function; The multidimensional physical constraint of the methane concentration field is obtained by the following formula: ， ， in, express The methane plume at time [time] at coordinates methane concentration at that location The physical area representing the concentration field image of pure methane. Indicates the methane plume diffusion cessation time. Represents the methane concentration field median coordinate The methane concentration at that location.

5. The method for constructing a dual-temporal Sentinel-2 methane plume dataset according to claim 2, characterized in that, The step of edge-modulating the methane concentration field based on the anisotropic stretching ratio, the basic eddy space scale, and the methane plume variation coefficient to obtain the edge-modulated methane concentration field includes: Calculate the core protection mask based on the methane concentration field; Based on the core protection mask and the variation coefficient of the methane plume, the effective turbulence intensity matrix is ​​extracted; The turbulence disturbance factor is calculated based on the anisotropic stretching ratio and the basic eddy space scale. The edge-modulated methane concentration field is calculated based on the effective turbulence intensity matrix, methane concentration field, and turbulence perturbation factor. The core protective mask is obtained through the following formula: ， in, Indicates the core protective mask of the main trunk. median coordinate The protection coefficient at the location, Represents the methane concentration field median coordinate methane concentration at that location Represents the methane concentration field The maximum methane concentration in; The effective turbulence intensity matrix is ​​obtained by the following formula: ， in, Representing coordinates Effective turbulence intensity at that location, Indicates the overall basic tear strength. , This represents the function that takes the minimum value. This represents the function that takes the maximum value. This indicates the preset turbulence intensity. Indicates the coefficient of variation of the methane plume; The turbulence disturbance factor is obtained by the following formula: ， ， ， ， ， ， ， ， in, This represents the turbulence disturbance factor. This represents the mixing matrix of multi-scale, along-wind, anisotropic random perturbation fields. Indicates rotation operation. Indicates wind direction and The included angle of the axis, This represents the mixing matrix of anisotropic random perturbation fields at multiple scales. This indicates low-frequency scale characteristics. Indicates high-frequency scale characteristics. This represents the anisotropic Gaussian space convolution operation. This represents the preset white noise matrix. Indicates the standard deviation of the minor axis. Indicates the standard deviation of the major axis. Represents the integral scale of turbulence. Indicates the spatial scale of the basic eddy current. Indicates the anisotropic stretch ratio. Indicates the preset wind speed; The edge-modulated methane concentration field is obtained by the following formula: ， ， in, Represents the methane concentration field after edge modulation median coordinate methane concentration at that location Represents the integrated modulation multiplier matrix. Represents the integrated modulation multiplier matrix median coordinate The overall modulation multiplier at the location, This represents the effective turbulence intensity matrix.

6. The method for constructing a dual-temporal Sentinel-2 methane plume dataset according to claim 1, characterized in that, The pair Shortwave infrared of the target area image at any time The physical attenuation injection of reflectivity in the band is achieved through the following formula: ， in, Indicates physical decay injection Shortwave infrared of the target area image at any time Reflectivity of the band express Shortwave infrared of the target area image at any time Reflectivity of the band This represents the natural exponential function. Represents the random optical attenuation coefficient. This represents the relative plume density matrix.

7. The method for constructing a dual-temporal Sentinel-2 methane plume dataset according to claim 1, characterized in that, The normalized methane index change matrix is ​​obtained by the following formula: ， ， in, This represents the normalized methane index variation matrix. Represents the methane index variation matrix. Represents the matrix of methane index changes The mean, Represents the matrix of methane index changes standard deviation As a preset value, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix, express Shortwave infrared of the target area image at any time Band matrix.

8. A method for detecting methane plumes in a two-phase Sentinel-2 configuration, characterized in that, include: Input the methane plume sample into the methane plume detection model: The methane plume samples were subjected to channel attention weighting using a spectral attention module to obtain a weighted methane plume feature vector. The weighted methane plume feature vector is input into the pyramid vision Transformer encoder for multi-scale feature extraction to obtain a multi-scale high-dimensional feature map. The multi-scale high-dimensional feature map is input into the U-Net decoder for feature fusion to obtain the fused feature map. The fused feature map is input into the output layer for methane concentration detection to obtain the methane concentration detection result. The methane plume detection model constructs a training dataset using the dual-temporal Sentinel-2 methane plume dataset construction method as described in any one of claims 1 to 7 during the training process.

9. The method for detecting a two-phase Sentinel-2 methane plume according to claim 8, characterized in that, The spectral attention module includes a globally average pooling layer, a multilayer perceptron, and a sigmoid activation layer connected in sequence. The weighted methane plume eigenvector is obtained by the following formula: ， ， in, Represents the methane plume weight vector. This represents the Sigmoid activation function. Represents the ReLU activation function. This represents the first learnable weight matrix of the multilayer perceptron. This represents the second learnable weight matrix of the multilayer perceptron. This indicates a global average pooling operation. This indicates a methane plume sample. This represents the weighted methane plume eigenvector.

10. The method for detecting a two-phase Sentinel-2 methane plume according to claim 8, characterized in that, The joint loss function of the methane plume detection model is as follows: ， ， ， ， in, Indicates joint loss, Indicates the preset weight. Indicates focal loss. This represents the total number of samples from the methane plume. , Indicates preset parameters. Indicates the detection confidence level. , This represents the natural exponential function. Represents the binary cross-entropy loss. This indicates the methane concentration detection results of the methane plume sample. The plume truth label matrix represents the methane plume samples. express The Middle The truth value label for the feather flow is 1 pixel. express The Middle methane concentration detection results per pixel, This represents the Dice coefficient loss. This is the default value.

Images