Methane point source remote sensing detection inversion method based on physical guidance deep learning

Abstract

The application discloses a methane point source remote sensing detection inversion method based on physical guidance deep learning, which comprises the following steps: obtaining a sample data set of a research area; inputting the sample data set into a radiation transmission combination model to obtain a synthetic methane remote sensing image radiance sample set; a multi-task prediction head of a multi-task deep learning model comprises a background concentration prediction head, a concentration enhancement prediction head and a plume mask prediction head; the multi-task deep learning model is trained by using the synthetic methane remote sensing image radiance sample set; and the synthetic methane remote sensing image radiance data of the research area is input into the multi-task deep learning model to obtain concentration enhancement distribution data and a plume pixel mask. The application realizes plume injection to generate a synthetic training sample through radiation transmission simulation, realizes joint optimization of plume mask prediction, concentration enhancement regression and background concentration estimation through a multi-task deep learning model, and improves detection precision and emission inversion reliability under complex background conditions.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of environmental remote sensing monitoring and artificial intelligence, and in particular to a method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning. Background Technology

[0002] Methane (CH4) is a significant greenhouse gas influencing global climate change. It exhibits significant absorption characteristics in the shortwave infrared band, allowing for the identification and retrieval of near-surface methane concentration changes through remote sensing. With the development of imaging spectrometers, hyperspectral remote sensing images possess fine and nearly continuous spectral resolution, providing denser spectral information in methane absorption-sensitive bands. This allows for a more comprehensive characterization of subtle changes in absorption patterns, making them valuable for methane point source identification, plume region extraction, and quantitative estimation of enhanced concentration. Compared to multispectral images, which rely on a limited number of discrete bands, hyperspectral images offer advantages in weak plume signal extraction, complex background suppression, and spectral differentiation under different land surface conditions, making them a crucial data source for refined methane monitoring.

[0003] Existing methods for methane retrieval and detection based on hyperspectral remote sensing imagery mainly include physical inversion algorithms, CO2 surrogate methods, and matched filtering methods. Physical inversion algorithms, based on the radiative transfer equation, model methane concentration inversion by inverting atmospheric absorption, scattering, and surface albedo processes. While physically meaningful, they are sensitive to changes in aerosols, clouds, surface reflectance, and observation geometry. Furthermore, the numerous model parameters and complex solutions can easily lead to increased inversion uncertainty. CO2 surrogate methods utilize the shared absorption characteristics of CO2 and methane in specific bands to correct for changes in the optical path caused by scattering. However, they are highly dependent on prior CO2 conditions, spectral quality, and noise levels, and may still produce errors in complex backgrounds or under varying observation conditions. Matched filtering methods can directly output methane concentration enhancement information with high computational efficiency, making them suitable for point source detection and plume identification tasks. However, they are sensitive to background changes, noise interference, and threshold selection, and struggle to achieve stable quantitative estimation of enhancement while maintaining the integrity of the plume morphology. In practical applications, hyperspectral methane point source detection and emission inversion still face several technical challenges: First, hyperspectral images have high dimensionality and strong information redundancy, making them susceptible to sensor noise, strip noise, and observational geometric changes, which can mask subtle changes in methane absorption characteristics with background disturbances. Second, methane plumes are significantly affected by factors such as wind speed and direction, boundary layer stability, and topography, resulting in strong nonlinearity and spatiotemporal variation in plume morphology, leading to significant differences in the enhancement distribution of the same emission intensity under different meteorological conditions. Third, accurate manual labeling of real pixel-level plume masks and concentration enhancement quantitative labels is difficult, resulting in insufficient supervised learning training samples and high label noise, which limits the model's generalization ability. Fourth, methane point source monitoring not only requires locating suspected emission areas but also outputting concentration enhancement distributions and background concentration benchmarks that can be used for emission inversion. This requires algorithms to simultaneously possess detection, segmentation, and quantitative estimation capabilities within the same framework, while existing methods struggle to balance plume boundary positioning accuracy, enhancement regression stability, and background estimation reliability.

[0004] In recent years, deep learning methods have demonstrated strong feature learning capabilities in remote sensing image target detection, semantic segmentation, and physical quantity inversion tasks. They can automatically extract multi-scale discriminative features through end-to-end training, providing a new technical approach for hyperspectral methane point source detection and concentration enhancement inversion. However, deep learning methods are significantly dependent on high-quality supervised data and are prone to false detections, boundary fragmentation, or regression instability under weak plume signals and complex background conditions. Therefore, there is an urgent need to propose a method that can accurately extract methane point source plumes and stably estimate their concentration in hyperspectral remote sensing images, thereby improving robustness and generalization ability under different observation conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning. This method generates physically consistent synthetic training data through forward simulation of radiative transfer. A multi-task deep learning model forms a model processing architecture encompassing data generation, feature extraction, collaborative learning, and plume prediction. This enables end-to-end synchronous output of methane plume masks, pixel-level concentration enhancement maps, and background concentration estimates from multi-source remote sensing images. Then, by combining meteorological parameters and accurately inverting point source emission fluxes, this method achieves high-precision, automated detection and emission inversion of methane point sources, providing complete and reliable technical support for automated, high-precision satellite inspection and quantitative monitoring of methane point source emissions.

[0006] The objective of this invention is achieved through the following technical solution: A method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning, comprising the following steps: S1. Obtain the sample dataset of the study area. The sample dataset includes remote sensing image samples, methane concentration plume labels, and meteorological parameter data. The methane concentration plume labels include background concentration pixel labels, plume concentration enhancement pixel labels, and plume pixel mask labels. S2. Construct a combined radiative transfer model that includes an atmospheric diffusion physics model and a plume diffusion concentration field simulation model. Input the sample dataset into the combined radiative transfer model to obtain a synthetic methane remote sensing image radiance sample set. S3. Construct a multi-task deep learning model that includes a spatial-frequency domain joint encoder, a cross-scale feature fusion module, a task-adaptive feature generation module, and a multi-task prediction head. The multi-task prediction head includes a background concentration prediction head, a concentration enhancement prediction head, and a plume mask prediction head. The background concentration prediction head is used to output the estimated background concentration, the concentration enhancement prediction head is used to output the concentration enhancement distribution data, and the plume mask prediction head is used to segment and output the plume pixel mask. The multi-task deep learning model is trained using a synthetic methane remote sensing image radiance sample set. S4. Obtain remote sensing image data and meteorological parameter data of the study area and input them into the radiative transfer combination model to obtain synthetic methane remote sensing image radiance data; input the synthetic methane remote sensing image radiance data into the multi-task deep learning model to obtain concentration enhancement distribution data and plume pixel mask.

[0007] To better achieve the present invention, the present invention also includes the following methods: S5. Constructing a point source emission inversion model based on the integral mass enhancement method, plume integral mass enhancement. The expression is as follows: , The feather region defined by the feather pixel mask Inner pixel The column concentration enhancement value, Let Q be the actual land surface area corresponding to the pixel; the expression for the point source emission rate Q is as follows: , The effective wind speed is extracted from the meteorological parameter data, and L is the plume feature length in the plume pixel mask.

[0008] Preferably, in method S2, the radiance of the synthesized methane remote sensing image in the synthesized methane remote sensing image radiance sample set of the radiative transfer combination model after plume injection is... The expression for the composite observation result is as follows: , For the plume injection process radiation term, The wavelength of the satellite band in the remote sensing image sample. Inject a gain term into the feather stream. For radiance observation, For background process radiation term, This is the background gain term.

[0009] Preferably, the spatial-frequency domain co-encoder SF-DE includes a two-dimensional convolution Conv2d, a layer normalization LN, a spatial-frequency fusion module SF-Fusion Block, and an upsampling operation module UpSample. The spatial-frequency domain co-encoder SF-DE sequentially extracts five-scale pyramid features SF1, SF2, SF3, SF4, and SF5 from the synthetic methane remote sensing image. The spatial-frequency fusion module SF-Fusion Block includes a spatial domain branch SC-Block, a frequency domain branch FC-Block, and a channel attention gating module CAG. The spatial domain branch SC-Block processes and enhances the expressive power, including plume boundaries and morphological contours, through depthwise separable convolution, layer normalization, and residual connections. The frequency domain branch FC-Block maps features to the frequency domain through fast Fourier transform and enhances the changing features caused by methane absorption using frequency domain convolution units. The channel attention gating module CAG is used to perform adaptive weighted fusion processing on the feature outputs of the spatial domain branch SC-Block and the frequency domain branch FC-Block.

[0010] Preferably, the cross-scale feature fusion module includes a first feature processing branch and a second feature processing branch. The first feature processing branch performs multi-attention feature extraction and cross-weight fusion on features SF2 and SF3 through a multi-attention selection fusion block (MASF) and a cross-attention gating module (CAG). Then, it enhances the high-frequency details of the plume boundary and weak signal region through a fast Fourier detail-aware block (FFDP) to obtain feature FF1, which focuses on general representation. The second feature processing branch performs multi-attention feature extraction and cross-weight fusion on features SF4 and SF5 through a multi-attention selection fusion block (MASF) and a cross-attention gating module (CAG). Then, it enhances the high-frequency details of the plume boundary and weak signal region through a fast Fourier detail-aware block (FFDP) to obtain feature FF2, which focuses on deep semantics.

[0011] Preferably, the task-adaptive feature generation module takes features SF1 and FF1 as inputs, and performs cross-gating weighted average fusion of features SF1 and FF1 through a cross-gating mechanism to generate differentiated features PF and CF for different tasks.

[0012] Preferably, the background concentration prediction head performs global average pooling and linear regression on feature FF2 and outputs an estimated background concentration; the concentration enhancement prediction head performs convolution and softplus activation on feature CF and outputs a concentration enhancement distribution map. From the concentration enhancement distribution map Concentration-enhanced distribution data are extracted; the plume mask prediction head performs convolution processing and sigmoid activation processing on the feature PF to output a plume probability map. The plume pixel mask is obtained by segmenting from the plume probability map.

[0013] Preferably, the multi-task deep learning model uses the following multi-task joint loss function for loss constraint: ,in The total loss of the model, Indicates background concentration loss. This indicates that the regression loss increases with concentration. This indicates the plume mask segmentation loss. This represents the mask-enhanced consistency constraint loss; , , To preset the trade-off parameters, This is an adaptive tradeoff parameter for background concentration loss.

[0014] Preferably, the mask enhances consistency constraint loss. Used to constrain plume mask prediction results With concentration enhancement prediction results To ensure consistency in spatial distribution, the following methods are employed: suppress concentration enhancement output in the region outside the plume mask to avoid spatial drift where the mask serves as the background but concentration enhancement is significant; enhance concentration enhancement output in the region inside the plume mask and maintain consistency with the plume mask prediction space so that concentration enhancement is mainly concentrated in the plume region.

[0015] Preferably, the background concentration pixel labels in the sample dataset are obtained by deploying high-precision continuous methane monitoring equipment in a clean background area selected upwind of the prevailing wind direction or far from the emission source in the study area to continuously detect the methane background concentration; the meteorological parameter data includes wind speed, wind direction, temperature, air pressure, specific humidity, solar zenith angle, downwind / upwind atmospheric transmittance, boundary layer height, and atmospheric stability, and the meteorological parameter data are derived from ERA5 reanalysis data; the background concentration pixel labels, plume concentration enhancement pixel labels, and plume pixel mask labels in the synthetic methane remote sensing image radiance sample set are collected by pixel matching in the remote sensing image sample or synthetic methane remote sensing image.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention generates physically consistent synthetic training data through forward simulation of radiative transmission. The multi-task deep learning model forms a model processing architecture from data generation, feature extraction, collaborative learning to plume prediction. It realizes end-to-end synchronous output of methane plume mask, pixel-level concentration enhancement map and background concentration estimation from multi-source remote sensing images. Then, it combines meteorological parameters and accurately inverts point source emission flux, realizing high-precision and automated detection and emission inversion of methane point sources. It provides complete and reliable technical support for automated and high-precision satellite inspection and quantitative supervision of methane point source emissions.

[0017] (2) The present invention includes a spatial-frequency domain joint encoder, a cross-scale feature fusion module, a task-adaptive feature generation module, and a multi-task prediction head multi-task deep learning model. The spatial-frequency domain joint encoder introduces a channel frequency domain feature processing mechanism during the step-by-step downsampling encoding process, so that the downsampling process integrates spatial domain structural information and channel dimension frequency response information. The cross-scale feature fusion module is used to fuse feature information under different spatial scales. The multi-task prediction head module is used to output methane plume mask prediction results, methane concentration enhancement prediction results, and background methane concentration prediction results respectively.

[0018] (3) This invention uses physical mechanisms as a data engine and constraint, embedding them into the training and application of the data-driven model. Based on point source parameters and meteorological fields, it simulates plume diffusion, generates concentration enhancement distributions and mask labels, and fuses physical disturbances with images through radiative transfer to construct a physically consistent synthetic training sample library. The multi-task deep learning model takes the synthetic samples as input and outputs background concentration estimates, concentration enhancement distributions, and plume masks end-to-end synchronously. It also performs joint optimization by introducing a multi-task loss function with mask-enhancement consistency constraints. The multi-task deep learning model is applied to the image under test to directly obtain the spatial range and enhancement intensity of the plume, and couples meteorological parameters such as wind speed to achieve accurate inversion of point source emissions. This invention achieves high-precision detection, pixel-level quantification, and reliable emission estimation of methane point sources through the deep integration of physical mechanisms and data-driven approaches, improving the detection accuracy and emission inversion reliability of weak plumes. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method for remote sensing detection and inversion of methane point sources according to the present invention; Figure 2 This is a structural diagram of the multi-task deep learning model in the embodiment; Figure 3 This is a schematic diagram of the spatial frequency domain joint encoder in the embodiment; Figure 4 This is the original structural diagram of the spatial frequency fusion module in the embodiment; Figure 5 This is the original structural diagram of the spatial domain branch in the spatial frequency fusion module of the embodiment; Figure 6 This is the original structural diagram of the frequency domain branch in the spatial frequency fusion module of the embodiment; Figure 7 This is the original structural diagram of the cross-scale feature fusion module in the embodiment; Figure 8 This is the original structural diagram of the multi-attention selection fusion block in the embodiment; Figure 9 This is the original structural diagram of the Fast Fourier Transform detail-aware block in the embodiment; Figure 10 This is the original structural diagram of the task adaptive feature generation module in the embodiment; Figure 11 This is the original structural diagram of the multi-task prediction head in the embodiment; Figure 12 This is a schematic diagram of the methane concentration enhancement mask segmentation in the embodiment; Figure 13 This is a schematic diagram illustrating the synthesis of methane remote sensing images and plume methane concentration enhancement mask in the embodiment. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to embodiments: Example like Figure 1 As shown, a method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning is described, the method comprising: S1. Obtain a sample dataset for the study area. The sample dataset includes remote sensing image samples, methane concentration plume labels, and meteorological parameter data. The study area contains emission point sources (i.e., sources emitting methane). The emission point source data includes the location and emission amount of the point sources. Obtain meteorological parameters (including wind speed and direction) corresponding to the remote sensing image data to construct a set of point source emission scenarios. Based on the emission scenario set, use plume diffusion simulation to obtain the concentration enhancement distribution of the methane plume. The methane concentration plume labels include background concentration pixel labels, plume concentration enhancement pixel labels, and plume pixel mask labels. See [link to relevant documentation]. Figure 12 , Figure 12 This example demonstrates the pixel regions of the study area corresponding to the plume pixel mask labels in the methane concentration plume label. In this embodiment, the preferred remote sensing image sample is L1A-level hyperspectral image data of the study area. The L1A-level hyperspectral image data is obtained from a data platform. Radiometric calibration and geometric correction are performed on the L1A-level hyperspectral image data. The L1A-level hyperspectral image data undergoes quality screening, which can be done pixel-by-pixel, removing low-quality data affected by clouds, shadows, or anomalies. The L1A-level hyperspectral image data is methane hyperspectral remote sensing image data, which includes multiple continuous spectral channels covering the methane absorption-sensitive band. Low-quality pixels are removed or masked based on cloud masks and quality labels to form an effective pixel set. Preprocessing operations such as georegistration, projection unification, and regional cropping are then performed to obtain the target area remote sensing image for subsequent processing. The background concentration pixel labels in the sample dataset are obtained by deploying high-precision continuous methane monitoring equipment in clean background areas selected upwind of the prevailing wind direction or far from emission sources within the study area to continuously detect methane background concentration. The high-precision continuous methane monitoring equipment performs continuous observations at a high frequency of ≥1Hz and strictly records the exact time of satellite (i.e., the satellite corresponding to the remote sensing image sample) transit over the target area (i.e., the study area) to ensure that the equipment time is synchronized with the satellite time (UTC time alignment). The high-precision continuous methane monitoring equipment obtains a continuous methane background concentration sequence. Generally, the median of the methane background concentration sequence within a certain time range is taken as the background methane concentration value (referred to as background concentration). In some embodiments, to quantify the reliability of the background methane concentration value, its robust statistical uncertainty is calculated to characterize the inherent fluctuation of the background atmosphere during satellite transit. First, the relationship between each observed value C(t) in the methane background concentration sequence and the background median is calculated. The absolute deviations are calculated, then the median of these absolute deviations is calculated, and finally converted to standard deviations. Robust estimates are obtained by multiplying approximately normally distributed data by a constant factor (k≈1.4826). Quality control and correction are performed on the observed data, and background methane concentration and its uncertainty are calculated within a preset time window centered on the satellite transit time. These are used for subsequent synthetic sample construction and model training constraints. Meteorological parameters include key parameters such as wind speed, wind direction, temperature, air pressure, specific humidity, solar zenith angle, down / up atmospheric transmittance, boundary layer height, and atmospheric stability. These meteorological parameters are derived from ERA5 reanalysis data.

[0021] S2. Construct a combined radiative transfer model, including an atmospheric diffusion physics model and a plume diffusion concentration field simulation model. The combined radiative transfer model includes an atmospheric radiative transfer model (i.e., the MODTRAN model). The atmospheric diffusion physics model simulates atmospheric diffusion based on meteorological parameter data. The plume diffusion concentration field simulation model establishes a three-dimensional diffusion coordinate system with the emission point source of the study area as the origin, the downwind direction as the x-axis, the crosswind direction as the y-axis, and the vertical direction as the φ-axis. This system is then combined with the atmospheric diffusion physics model to simulate the methane plume diffusion concentration field. Input the sample dataset into the combined radiative transfer model to obtain a synthetic methane remote sensing image radiance sample set. For each synthetic methane remote sensing image sample, construct a corresponding set of supervision labels. The supervision label set includes at least background concentration labels, pixel-level concentration enhancement labels, and plume mask labels. The pixel-level concentration enhancement labels and plume mask labels maintain the same spatial resolution as the synthetic samples. Figure 12 , Figure 13 As shown, background concentration pixel labels, plume concentration enhancement pixel labels, and plume pixel mask labels in the radiance sample set of the synthetic methane remote sensing image are grouped by pixel matching in the synthetic methane remote sensing image sample. Figure 12 This shows the pixel regions of the study area corresponding to the plume pixel mask labels in the methane concentration plume label. Figure 12 Excluding the black areas, the colored areas represent the plume pixel areas of the study area, while the black areas represent the non-plume areas of the study area. The color indicates the corresponding pixel. plume concentration enhancement value (Increasing from blue to yellow); Figure 13 This diagram illustrates the fusion of plume pixel mask labels with synthetic methane remote sensing image samples. Figure 12 The pixel regions corresponding to the mid-feather pixel mask labels in the study area are fused together in the synthetic methane remote sensing image samples. Figure 13 Thresholding was applied to the methane concentration enhancement map in the plume diffusion concentration field simulation using the combined radiative transfer model. Segmentation, when pixels Increased plume concentration Not less than the threshold At that time, the pixel A pixel is marked as a feather pixel, and the binary feather mask label M(p) is set to 1; otherwise, it is marked as a background pixel, and the binary feather mask label M(p) is set to 0. The expression for the binary feather mask label M(p) is as follows: Based on the plume concentration enhancement distribution, threshold discrimination is used to generate plume pixel mask labels. The plume mask labels can be further processed morphologically and / or filtered to remove isolated noise and enhance plume coherence, thereby obtaining plume mask labels for supervised learning.

[0022] The radiance of the synthetic methane remote sensing image after plume injection in the radiative transfer combined model. The expression for the synthetic observation result (based on radiative transfer, the plume concentration enhancement distribution is injected into the remote sensing image data to generate a synthetic methane remote sensing image sample) is as follows: , For the plume injection process radiation term, The wavelength of the satellite band in the remote sensing image sample. Inject a gain term into the feather stream. For radiance observation, For background process radiation term, This is the background gain term.

[0023] The expression for the radiance of the top of the atmosphere under the Lambertian surface assumption is as follows: ,in For sensors (Also an L1A product) Received radiance, Indicates the wavelength of the satellite band. For atmospheric path radiation, For surface reflectance, Sun zenith angle, Atmospheric gain term (including solar irradiance) Downstream With uplink transmittance Its expression is as follows: .

[0024] Given only background methane concentration, atmospheric profile parameters for the background scenario are constructed. These parameters include at least temperature, pressure, water vapor, and background methane profiles. The atmospheric profile parameters, observational geometric parameters (solar zenith angle, observational zenith angle, azimuth angle, etc.), and surface parameters are input into the MODTRAN model within the radiative transfer composite model. A forward radiative transfer simulation is performed to obtain the remote sensing response parameters under the background methane scenario, i.e., the background path radiation term. and background gain term , The wavelength represents the satellite band; therefore, the observation in the context of background methane is expressed as follows: ; Pixel-level plume concentration enhancement distribution was obtained using a plume diffusion model. The plume concentration enhancement distribution is converted into a methane concentration perturbation, which is then injected into the background methane profile within a preset height range to construct atmospheric profile parameters for the plume injection scenario. These atmospheric profile parameters, observation geometry parameters, and surface parameters are then input into the MODTRAN model within the radiative transfer composite model to perform a forward radiative transfer simulation, yielding the remote sensing response parameters under the plume injection scenario, i.e., the plume injection path radiative term. and feather injection gain term Therefore, the observation expression for the plume injection case is as follows: Since the background conditions and the plume injection conditions correspond to the same surface area, that is: Then we have the following expression: Therefore, the equivalent expression for surface reflectance can be obtained: Substituting the reflectance into the expression for the plume injection case, we obtain the composite observation results after plume injection: , For the plume injection process radiation term, The wavelength of the satellite band in the remote sensing image sample. Inject a gain term into the feather stream. For radiance observation, For background process radiation term, This is the background gain term.

[0025] S3. Construct a multi-task deep learning model including a spatial-frequency domain joint encoder SF-DE, a cross-scale feature fusion module CGFF Block, a task-adaptive feature generation module CGAF Block, and a multi-task prediction head. The structural principle diagram of the multi-task deep learning model in this embodiment is as follows: Figure 2 As shown, see Figure 3 The input synthetic methane remote sensing image first extracts features through the spatial-frequency domain co-encoder SF-DE. Then, it sequentially undergoes multi-level feature interaction and enhancement via the cross-guided attention fusion module CGAF Block and the cross-scale feature fusion module CGFF Block. Based on the fused features, the multi-task prediction head branches out in parallel into three dedicated task heads. The multi-task prediction head includes a background concentration prediction head, a concentration enhancement prediction head, and a plume mask prediction head. The background concentration prediction head outputs the estimated background concentration value, showing the global background concentration distribution; the concentration enhancement prediction head (or methane concentration value regression head, i.e....) Figure 2The CH4 value head module outputs concentration enhancement distribution data to calculate the concentration enhancement value within the plume region, providing a high-precision data foundation for quantitative inversion; the plume mask prediction head segments the output plume pixel mask, achieving pixel-level plume region segmentation (see the image for the segmented plume pixel mask). Figure 12 The multi-task deep learning model is trained using a radiance sample set of synthetic methane remote sensing images. Through multi-task collaboration and cross-guided fusion, the multi-task deep learning model significantly improves the robustness and accuracy of weak hyperspectral methane signal detection and quantification.

[0026] In some embodiments, the principle of the spatial frequency domain co-encoder SF-DE structure is as follows: Figure 3 As shown, the spatial-frequency domain co-encoder SF-DE includes multiple 2D convolutional modules (Conv2d), multiple layer normalization modules (LN), multiple spatial-frequency fusion modules (SF-Fusion Block), and multiple upsampling operation modules (UpSample). The structure of the spatial-frequency domain co-encoder SF-DE is as follows: Figure 3 As shown. The spatial-frequency domain co-encoder SF-DE extracts highly discriminative multi-scale features from synthetic methane remote sensing images, such as... Figure 3 As shown, the spatial-frequency domain co-encoder SF-DE sequentially extracts five-scale pyramid features SF1, SF2, SF3, SF4, and SF5 from the synthetic methane remote sensing image. The original structural diagram of the preferred spatial-frequency fusion module SF-Fusion Block in this embodiment is shown below. Figure 4 As shown, the spatial-frequency fusion module SF-Fusion Block includes a spatial domain branch SC-Block, a frequency domain branch FC-Block, and a channel attention gating module CAG. The preferred structural diagram of the spatial domain branch SC-Block in this embodiment is shown below. Figure 5 As shown, the spatial domain branch SC-Block enhances the representation of details such as plume boundaries and morphological contours through depthwise separable convolution, layer normalization, and residual connections. The original structure of the preferred frequency domain branch FC-Block in this embodiment is shown below. Figure 6As shown, the frequency domain branch FC-Block maps features to the frequency domain through Fast Fourier Transform and enhances the variation features caused by methane absorption (mainly high-frequency spectral variation features) using frequency domain convolutional units. Then, it performs an inverse transform back to the spatial domain, thereby improving the separability of weak signals in complex backgrounds. The channel attention gating module CAG is used to adaptively weight and fuse the feature outputs of the spatial domain branch SC-Block and the frequency domain branch FC-Block. This allows the multi-task deep learning model to retain both the spatial structure information and spectral frequency domain information that are crucial for detection during the process of progressive downsampling and expanding the receptive field. Finally, the encoder outputs a feature pyramid {SF1, SF2, SF3, SF4, SF5} containing five scales, providing multi-resolution feature primitives for subsequent processing.

[0027] In some embodiments, the preferred structural principle of the cross-scale feature fusion module CGFF Block in this embodiment is as follows: Figure 7 As shown, the purpose of the cross-scale feature fusion module CGFFBlock is to efficiently integrate the multi-scale features output by the cross-scale feature fusion module CGFFBlock to form a coherent and complete representation of the plume. The cross-scale feature fusion module CGFFBlock includes a first feature processing branch and a second feature processing branch. The first feature processing branch processes features SF2 and SF3 through a multi-attention selection fusion block MASF (the structural principle of the multi-attention selection fusion block MASF in this embodiment is shown below). Figure 8 As shown), the cross-attention gating module (CAG) performs multi-attention feature extraction and cross-weight fusion. Then, the Fast Fourier Detail-Aware Block (FFDP) is used to enhance high-frequency details in plume boundaries and weak signal regions, resulting in feature FF1, which focuses on general representation (and will serve subsequent plume segmentation and enhanced regression). The second feature processing branch extracts multi-attention features SF4 and SF5 through the multi-attention selection fusion block (MASF) and the cross-attention gating module (CAG). Then, the Fast Fourier Detail-Aware Block (FFDP) is used (the structure and principle of the FFDP in this embodiment are shown below). Figure 9 (As shown) to enhance the high-frequency details of the plume boundary and weak signal region and obtain the feature FF2 which focuses on deep semantics (which will be directly used for background concentration estimation).

[0028] In some embodiments, the preferred structural principle of the task adaptive feature generation module CGAF Block in this embodiment is as follows: Figure 10As shown, the task-adaptive feature generation module CGAF Block is a key design element in resolving the contradiction between "segmentation requiring fine boundaries" and "regression requiring global smoothness." CGAF Block takes features SF1 (detail-rich shallow high-resolution features) and FF1 (semantic-rich deep fusion features) as input. It uses a cross-gating mechanism to perform cross-gating weighted averaging fusion of features SF1 and FF1, generating differentiated features PF and CF tailored to different tasks. Figure 10 As shown, one output feature PF focuses more on the details of SF1 and is specifically designed for plume mask prediction tasks that are sensitive to boundaries; the other output feature CF focuses more on the context of FF1 and is specifically designed for concentration-enhanced regression tasks with stable values. This strategy of "homogeneous features, heterogeneous expression" is the core of achieving synergistic optimization of high-precision detection and quantitative inversion.

[0029] In some embodiments, the preferred structural principle of the multi-task prediction head in this embodiment is as follows: Figure 11 As shown, the multi-task prediction head includes a background concentration prediction head and a concentration enhancement prediction head (also known as a methane concentration value regression head, corresponding to...). Figure 2 The CH4 value head and the plume mask prediction head are used. The background concentration prediction head performs global average pooling and linear regression on the feature FF2 and outputs the estimated background concentration value. The concentration enhancement prediction head performs convolution processing on the feature CF and Softplus activation processing to output a concentration enhancement distribution map. From the concentration enhancement distribution map Concentration-enhanced distribution data are extracted. The plume mask prediction head performs convolution and sigmoid activation on the feature power vector (PF) to output a plume probability map. The plume pixel mask is obtained by segmenting from the plume probability map.

[0030] In some embodiments, the multi-task deep learning model performs end-to-end optimization using a multi-task joint loss function (i.e., the multi-task joint loss function). This function simultaneously constrains three tasks: background concentration prediction, concentration enhancement regression, and plume mask segmentation. A consistency constraint term enhances the spatial consistency between mask prediction and concentration enhancement prediction, thereby reducing false detections and spatial drift. The multi-task deep learning model uses the following multi-task joint loss function for loss constraint: ,in The total loss of the model, Indicates background concentration loss. This indicates that the regression loss increases with concentration. This indicates the plume mask segmentation loss. This represents the mask enhancement consistency constraint loss. , , To preset the trade-off parameters, An adaptive tradeoff parameter for background concentration loss. Mask enhancement consistency constraint loss. Used to constrain plume mask prediction results With concentration enhancement prediction results To ensure spatial consistency, the following methods are employed: Suppress concentration enhancement output outside the plume mask to avoid spatial drift where the mask serves as background but concentration enhancement is significant. Enhance concentration enhancement output within the plume mask and maintain spatial consistency with the plume mask prediction, ensuring concentration enhancement is primarily concentrated within the plume region. The consistency constraint loss can be defined as a spatially weighted suppression term based on mask prediction: , among which when Approaching 0 (background area), this item will affect This results in a stronger penalty, thus suppressing spurious concentration enhancement responses in the background region. To further enhance response consistency within the plume region, a plume region enhancement term can be introduced, giving the concentration enhancement output a higher response intensity in the plume region; therefore, the consistency loss can be extended to: ,in, The coefficient for the enhancement term within the plume is the weighting factor. The minimum concentration enhancement response threshold is set to constrain the concentration enhancement output within the plume region to be no lower than a preset response level, thereby improving the consistency and stability of the mask and regression output. A multi-task joint loss function is established based on the supervised label set to train and optimize the model. The trained multi-task deep learning model is obtained when the loss function converges or meets a preset stopping condition. This invention's multi-task deep learning model, through the aforementioned multi-task joint loss function, can simultaneously optimize three tasks during training: background estimation, plume segmentation, and concentration enhancement regression. It also utilizes consistency constraints to reduce the mismatch between the mask and concentration enhancement output, thereby reducing false detections and spatial drift, and improving plume extraction accuracy and emission inversion accuracy. Simultaneously, an adaptive weighting mechanism for background uncertainty reduces the impact of background label noise on training, improving the robustness of background estimation under different observation conditions.

[0031] S4. Acquire remote sensing image data and meteorological parameter data of the study area and input them into the radiative transfer composite model to obtain synthetic methane remote sensing image radiance data. Meteorological parameter data includes key meteorological parameters such as wind speed, wind direction, temperature, air pressure, specific humidity, solar zenith angle, down / up atmospheric transmittance, boundary layer height, and atmospheric stability. The meteorological parameter data is derived from ERA5 reanalysis data. Input the synthetic methane remote sensing image radiance data into a multi-task deep learning model to obtain concentration enhancement distribution data and plume pixel masks. The multi-task deep learning model obtains background concentration estimates (i.e., background methane concentration prediction results), concentration enhancement distribution data (i.e., methane concentration enhancement prediction results), and plume pixel masks (i.e., methane plume mask prediction results; see the plume pixel segmentation mask for details). Figure 12 This provides stable and reliable data input for the inversion of final methane emissions in the study area.

[0032] S5. Construct a point source emission inversion model based on the integral mass enhancement method to obtain the methane concentration enhancement map output by the multi-task deep learning model. With binarized plume mask Next, it needs to be converted into a quantitative point source emission flux. The final inversion is completed using the integral mass enhancement method, which follows the law of mass conservation. The core idea is: under the steady-state assumption, the emission rate of a point source equals the total methane enhancement accumulated in its downwind plume, divided by the time required for the gas mass to be transported through the characteristic length of the plume. (Plume integral mass enhancement) The expression is as follows: , The feather region defined by the feather pixel mask Inner pixel The column concentration enhancement value, The actual surface area corresponding to the pixel; plume integral quality enhancement. The net methane mass of the plume contained in the atmospheric column and captured by the satellite during its transit was quantified. The integral mass enhancement (IME) characterizes the total methane mass exceeding the background value within the plume space and is a core physical quantity connecting the concentration field and emission flux. The expression for the point source emission rate Q is as follows: , Effective wind speed is the wind speed extracted from meteorological parameter data. The expression is as follows: , The wind speed at a height of 10 meters is derived from ERA5 reanalysis data; the coefficients 0.47 and 0.31 are obtained from the calibration of the simulated plume based on the WRF-Chem model. L is the characteristic length of the plume in the plume pixel mask (i.e., the spatial extension scale of the plume in the downwind direction).

[0033] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning, characterized in that: The methods include: S1. Obtain the sample dataset of the study area. The sample dataset includes remote sensing image samples, methane concentration plume labels, and meteorological parameter data. The methane concentration plume labels include background concentration pixel labels, plume concentration enhancement pixel labels, and plume pixel mask labels. S2. Construct a combined radiative transfer model that includes an atmospheric diffusion physics model and a plume diffusion concentration field simulation model. Input the sample dataset into the combined radiative transfer model to obtain a synthetic methane remote sensing image radiance sample set. S3. Construct a multi-task deep learning model that includes a spatial-frequency domain joint encoder, a cross-scale feature fusion module, a task-adaptive feature generation module, and a multi-task prediction head. The multi-task prediction head includes a background concentration prediction head, a concentration enhancement prediction head, and a plume mask prediction head. The background concentration prediction head is used to output the estimated background concentration, the concentration enhancement prediction head is used to output the concentration enhancement distribution data, and the plume mask prediction head is used to segment and output the plume pixel mask. The multi-task deep learning model is trained using a synthetic methane remote sensing image radiance sample set. S4. Obtain remote sensing image data and meteorological parameter data of the study area and input them into the radiative transfer combination model to obtain synthetic methane remote sensing image radiance data; input the synthetic methane remote sensing image radiance data into the multi-task deep learning model to obtain concentration enhancement distribution data and plume pixel mask.

2. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 1, characterized in that: It also includes the following methods: S5. Constructing a point source emission inversion model based on the integral mass enhancement method, plume integral mass enhancement. The expression is as follows: , The feather region defined by the feather pixel mask Inner pixel The column concentration enhancement value, Let Q be the actual land surface area corresponding to the pixel; the expression for the point source emission rate Q is as follows: , The effective wind speed is extracted from the meteorological parameter data, and L is the plume feature length in the plume pixel mask.

3. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 1 or 2, characterized in that: In method S2, the radiance of the synthesized methane remote sensing imagery radiance sample set in the radiative transfer combined model after plume injection is... The expression for the composite observation result is as follows: , For the plume injection process radiation term, The wavelength of the satellite band in the remote sensing image sample. Inject a gain term into the feather stream. For radiance observation, For background process radiation term, This is the background gain term.

4. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 1, characterized in that: The spatial-frequency domain co-encoder SF-DE includes a two-dimensional convolution Conv2d, a layer normalization LN, a spatial-frequency fusion module SF-Fusion Block, and an upsampling operation module UpSample. The spatial-frequency domain co-encoder SF-DE sequentially extracts five-scale pyramid features SF1, SF2, SF3, SF4, and SF5 from the synthetic methane remote sensing image. The spatial-frequency fusion module SF-Fusion Block includes a spatial domain branch SC-Block, a frequency domain branch FC-Block, and a channel attention gating module CAG. The spatial domain branch SC-Block processes and enhances the expressive power, including plume boundaries and morphological contours, through depthwise separable convolution, layer normalization, and residual connections. The frequency domain branch FC-Block maps features to the frequency domain through fast Fourier transform and enhances the changing features caused by methane absorption using frequency domain convolution units. The channel attention gating module CAG is used to perform adaptive weighted fusion processing on the feature outputs of the spatial domain branch SC-Block and the frequency domain branch FC-Block.

5. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 4, characterized in that: The cross-scale feature fusion module includes a first feature processing branch and a second feature processing branch. The first feature processing branch extracts and performs cross-weighted fusion of features SF2 and SF3 through the multi-attention selection fusion block MASF and the cross-attention gating module CAG. Then, the high-frequency details of the plume boundary and weak signal region are enhanced by the fast Fourier detail perception block FFDP to obtain feature FF1, which focuses on general representation. The second feature processing branch extracts and performs cross-weighted fusion of features SF4 and SF5 through the multi-attention selection fusion block MASF and the cross-attention gating module CAG. Then, it enhances the high-frequency details of the plume boundary and weak signal region through the fast Fourier detail-aware block FFDP to obtain feature FF2, which focuses on deep semantics.

6. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 5, characterized in that: The task-adaptive feature generation module takes features SF1 and FF1 as inputs. The task-adaptive feature generation module performs cross-gating weighted average fusion of features SF1 and FF1 through a cross-gating mechanism to generate differentiated features PF and CF for different tasks.

7. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 6, characterized in that: The background concentration prediction head performs global average pooling and linear regression on feature FF2 and outputs an estimated background concentration value; the concentration enhancement prediction head performs convolution and softplus activation on feature CF and outputs a concentration enhancement distribution map. From the concentration enhancement distribution map Concentration-enhanced distribution data are extracted; the plume mask prediction head performs convolution processing and sigmoid activation processing on the feature PF to output a plume probability map. The plume pixel mask is obtained by segmenting from the plume probability map.

8. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 1, characterized in that: The multi-task deep learning model uses the following multi-task joint loss function for loss constraint: ,in The total loss of the model, Indicates background concentration loss. This indicates that the regression loss increases with concentration. This indicates the plume mask segmentation loss. This represents the mask-enhanced consistency constraint loss; , , To preset the trade-off parameters, This is an adaptive tradeoff parameter for background concentration loss.

9. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 1, characterized in that: The mask-enhanced consistency constraint loss Used to constrain plume mask prediction results With concentration enhancement prediction results To ensure consistency in spatial distribution, the following methods are employed: suppress concentration enhancement output in the region outside the plume mask to avoid spatial drift where the mask serves as the background but concentration enhancement is significant; enhance concentration enhancement output in the region inside the plume mask and maintain consistency with the plume mask prediction space so that concentration enhancement is mainly concentrated in the plume region.

10. The method for remote sensing detection and inversion of methane point sources based on physics-guided deep learning according to claim 1, characterized in that: The background concentration pixel labels in the sample dataset were obtained by deploying high-precision continuous methane monitoring equipment in clean background areas selected upwind of the prevailing wind direction or far from emission sources in the study area to continuously detect methane background concentration. The meteorological parameter data included wind speed, wind direction, temperature, air pressure, specific humidity, solar zenith angle, downwind / upwind atmospheric transmittance, boundary layer height, and atmospheric stability. The meteorological parameter data were obtained from ERA5 reanalysis data. The background concentration pixel labels, plume concentration enhancement pixel labels, and plume pixel mask labels in the synthetic methane remote sensing image radiance sample set were collected by pixel matching in the remote sensing image samples or synthetic methane remote sensing images.

Images