Method and system for automatic detection of methane plumes and quantification of source rate based on hyperspectral satellite data

Abstract

This invention discloses an automatic methane plume detection and source rate quantification method and system based on hyperspectral satellite data. The methane plume detection and source rate quantification are decoupled into two independent stages: In the detection stage, a three-dimensional methane plume is generated and superimposed onto a real, emission-free satellite background scene according to background noise levels. A large-scale synthetic training dataset is constructed, and various deep learning semantic segmentation models are trained. The optimal model is selected through comprehensive evaluation to generate a high-quality binary plume mask. In the quantization stage, the source rate is independently estimated using a CNN regression module and a quantization module, respectively. A five-dimensional fused feature vector is constructed, incorporating mask quality, wind speed uncertainty, background noise level, source rate range, and method consistency. The two estimates are dynamically fused through a piecewise adaptive weighting function and an inverse variance weighted optimization mechanism. This solves the core problems of scarce labeled data, complex noise interference, and the limited accuracy of single methods.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing data processing and atmospheric environment monitoring technology, specifically relating to an automatic detection method and system for methane plumes and point source emission rate quantification based on hyperspectral satellite methane column concentration enhancement products. It is particularly suitable for methane remote sensing monitoring applications that use large eddy simulation synthetic datasets to train deep learning segmentation models and adopt an adaptive weighted hybrid quantization framework that integrates physical priors and deep learning. Background Technology

[0002] Methane (CH4) is the second largest greenhouse gas after carbon dioxide, and its global warming potential (GWP) on a 100-year timescale is 28–34 times that of carbon dioxide. The advancement of global methane emission reduction policies makes the accurate identification and quantification of methane point source emissions a key task in addressing climate change. However, traditional bottom-up inventory methods often systematically underestimate methane emissions, while the emergence of space-based remote sensing technologies based on hyperspectral imaging satellites (such as EMIT and CarbonMapper) provides a new data foundation for large-scale, automated monitoring of methane point source emissions.

[0003] EMIT (Earth Surface Mineral Dust Source Investigation), aboard the International Space Station, provides hyperspectral data across 224 spectral bands with a spatial resolution of approximately 60 m. Its L2B methane column concentration enhancement product can directly reveal the presence and spatial distribution characteristics of methane plumes. However, using EMIT hyperspectral data for the automated detection and quantification of methane plumes faces three major challenges:

[0004] (1) Scarcity of labeled data: High-quality methane plume labeled data in real-world scenarios is extremely limited, which severely restricts the training and generalization of supervised learning methods;

[0005] (2) Background noise interference: Complex surface reflectivity variations and atmospheric scattering form a heterogeneous noise field, which interferes with the accurate extraction of plume signals;

[0006] (3) Insufficient accuracy of source rate quantization: Traditional integrated quality enhancement (IME) methods are highly dependent on wind speed data, and wind speed uncertainty is directly propagated to emission rate estimation; while pure deep learning methods have a systematic bias of regressing to the training mean.

[0007] Existing methods can be broadly categorized into two types: IME methods based on physical models offer advantages such as clear physical meaning and lack of systematic bias, but are highly dependent on accurate wind speed data; deep learning methods like MethaNet and U-Plume exhibit better robustness in complex noisy scenarios, but suffer from limitations such as insufficient interpretability and out-of-range bias. Current methods either use physical models alone or deep learning alone, lacking a systematic framework that can adaptively fuse the advantages of both methods based on scene characteristics. In particular, they lack dynamic weight allocation strategies based on multi-dimensional scene features and inverse variance weighted refinement mechanisms based on method uncertainty.

[0008] Furthermore, most of the synthetic datasets used in existing methods are not constructed in a hierarchical manner to distinguish between background noise levels, resulting in a deviation between the training data distribution and real complex scenes; most methods only optimize a single quantization method and lack a complete framework for decoupling and independently optimizing the detection and quantization stages.

[0009] Regarding uncertainty modeling, existing methods have not yet established a system that balances mask quality (IoU) and wind speed reliability (σ) for satellite methane plume detection scenarios. wind / U 10 ) and background noise (σ B A three-dimensional composite uncertainty assessment model. Summary of the Invention

[0010] The purpose of this invention is to provide an automatic methane plume detection and source rate quantization method and system based on hyperspectral satellite data. By decoupling plume detection and source rate quantization into two independent stages for separate optimization, and constructing an adaptive weight fusion framework based on a five-dimensional fusion feature vector in the quantization stage, this invention overcomes the limitations of existing single methods in complex scenarios, improving physical interpretability while maintaining high accuracy and robustness. The σ proposed in this invention... CNN,rel and σ IME,rel The relative uncertainty model, by directly coupling the mask quality, wind speed relative error and inverse variance weighting mechanism, provides a quantifiable and dynamically responsive confidence assessment framework, which is currently the only publicly available solution of its kind in the field of satellite methane quantification.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] An automatic detection and source rate quantification method for methane plumes based on hyperspectral satellite data includes the following steps:

[0013] S1. Data Acquisition and Preprocessing: Acquire methane column concentration enhancement product data from hyperspectral imaging satellites, perform quality control on the data, remove scenes with cloud cover exceeding 5%, pixels with missing data or abnormal quality indicators, and perform geometric and radiometric corrections on valid data; simultaneously acquire reanalysis wind speed data and interpolate it to the satellite data spatial resolution.

[0014] S2. Construction of Synthetic Training Dataset: A three-dimensional methane plume concentration field is generated based on the large eddy simulation method. The simulated plume is superimposed on real satellite observation data of a background scene with no emissions to construct a large-scale synthetic training dataset containing binary mask labels for the plume. The background scene is divided into three categories: low noise, medium noise, and high noise according to the pixel-level standard deviation of methane column concentration enhancement.

[0015] The method for constructing a synthetic training dataset includes the following steps:

[0016] (1) Enhance the standard deviation σ using methane column concentration. B The ratio to the global average background concentration is used as a classification index to divide real emission-free satellite background scenes into low noise (σ) categories. B <2%), moderate noise (2%≤σ) B <5%) and high noise (σ B ≥5%) are three categories;

[0017] (2) Using the WRF-LES large eddy simulation method, covering a range of 1–10 m·s -1 Wind speed, three atmospheric stability levels (stable / neutral / unstable), 1–1000 kg·h -1 A three-dimensional methane plume concentration field is generated within the parameter space of emission intensity;

[0018] (3) The simulated plume is superimposed onto the corresponding category of emission-free background scene through four data augmentation operations: random placement, random rotation, random scaling and noise matching, according to the background noise level, to generate a large-scale labeled synthetic dataset with real noise distribution characteristics; the synthetic training dataset construction method solves the core problem of the scarcity of real labeled data in methane plume detection, and ensures the matching of the synthetic data distribution with the real complex scene through background noise layering.

[0019] S3. Feather Detection and Mask Generation: Using the synthetic training dataset constructed in S2, a deep learning semantic segmentation model is trained. The optimal model is selected by comparing segmentation performance indicators, and a binary feather mask is generated for the target scene.

[0020] S4. Adaptive Fusion Source Rate Quantization: Based on the plume binary mask and the original methane-enhanced image generated in S3, the source rate is independently estimated using the CNN regression module and the IME physics model module, respectively, and a fusion source rate quantization is constructed, including ① mask quality IoU. mask② Wind speed uncertainty σ wind ③ Source rate range Q range ④ Background noise level σ B Consistency with method ⑤ C consistency =|Q CNN -Q IME | / max(Q CNN Q IME The five-dimensional fusion feature vector; through IoU mask σ wind and C consistency The piecewise adaptive weight function, used as the criterion, dynamically switches between a high-confidence IME mode (α=0.2), a high-confidence CNN mode (α=0.8), and an adaptive mode with continuous sigmoid transitions, where f(Q) range The adjustment factor is physically constrained to correct the weights based on the emission intensity range; the fused output Q hybrid = α·Q CNN + (1-α)·Q IME ;

[0021] S5, Q output from S4 hybrid As a result, based on scene observable variables (IoU) mask σ B σ wind U 10 ) Calculate the relative uncertainty σ of the CNN method CNN,rel =1 / (1+IoU mask )×σ B / 10% and the relative uncertainty of the IME method σ IME,rel =σ wind / U 10 ×1 / (1+IoU mask ), using the inverse variance weighting formula α final = α·(1 / σ² CNN,rel ) / [α·(1 / σ² CNN,rel ) + (1-α)·(1 / σ² IME,rel The weights are refined and integrated to output the final source rate Q. final = α final ·Q CNN + (1-α final )·Q IME And its uncertainty assessment.

[0022] Preferably, the large eddy simulation described in S2 adopts the WRF-LES method, based on the Navier-Stokes equations and scalar transport equations. The simulation parameters are configured as follows: horizontal spatial resolution 60 m, simulation domain size 10.5 km × 10.5 km × 2.4 km, vertical grid spacing 15 m, time step 0.25 s, sampling interval 30 s; the point source is located at one-third of the upwind direction of the simulation domain to ensure that the plume extends fully downwind and is completely contained within the simulation scene; the simulation covers a wind speed range of 1–10 m / s. -1 At least 10 different wind speed scenarios, and three atmospheric stability conditions (stable, neutral, and unstable), with emission intensity ranging from 1 to 1000 kg·h. -1 .

[0023] Preferably, the construction of the synthetic training dataset in S2 includes the following operations: performing four data augmentation operations on the simulated plume—random placement, random rotation, random scaling, and noise matching—before superimposing it onto the background scene; wherein the noise level σ of the background scene... B Calculate using the following formula:

[0024] (1)

[0025] In the formula, x i Let methane column concentration be the value of pixel i. σ represents the average column density within the scene, where n is the total number of pixels; B Expressed as global average background concentration x b = 0.011 kg·m -2 The percentage; low-noise scenarios satisfy σ B < 2%, moderate noise scenarios satisfy 2% ≤ σ B < 5%, high-noise scenarios satisfy σ B ≥ 5%; the synthetic dataset should have at least 10,000 training samples, 2,000 validation samples, and 2,000 test samples.

[0026] Preferably, the deep learning semantic segmentation model in S3 is one or more of U-Net, DeepLabV3+, SegFormer, or PSPNet; the optimal model is selected from the multiple segmentation models by comprehensively comparing and evaluating IoU, Dice coefficient, precision, recall, and F1 score; during training, only successful detection samples with a Jaccard score greater than 0.1 are used to ensure the quality of the input mask.

[0027] Preferably, the CNN regression module in S4 adopts a three-channel explicit functional decoupled input architecture: Channel 1 is the original methane column concentration enhanced image to provide a direct signal of emission intensity; Channel 2 is the binary plume mask generated in S3 to achieve spatial attention guidance; and Channel 3 is the spatially distributed 10-meter height wind speed field to provide meteorological prior constraints. The binary mask in the three-channel input explicitly guides the convolutional feature extraction to focus on the plume region in the form of independent channels, avoiding the influence of background interference on the regression accuracy. The network structure is based on U-Net encoder path truncation and connection of fully connected regression layers. The encoder has 16 initial filters and a maximum channel depth of 256, containing 4 downsampling stages. Each stage contains two 3×3 convolutional layers and a batch normalization layer. The encoder output is compressed into a 256-dimensional feature vector by global average pooling, and then passes through two fully connected layers before a single-node linear output layer outputs the source rate estimate in kg·h. -1 The training loss function is the mean squared error.

[0028] (2)

[0029] In the formula, n is the batch size, Q predicted,i and Q true,i Let $\frac{i}{i}$ be the predicted and true source rates for the $i$-th sample, respectively; the Adam optimizer is used with a learning rate of 10^25. -4 The training rounds consist of 10 epochs, and an early stopping strategy (patience = 3 epochs) and cosine annealing learning rate scheduling are used. Training is terminated when the validation set loss no longer decreases for 3 consecutive epochs. For large datasets (>50,000 samples), the number of training rounds can be appropriately increased to 50-100 epochs to ensure sufficient convergence, and the batch size is 32.

[0030] Preferably, the IME physical model module in S4 calculates the source rate according to the following formula:

[0031] (3)

[0032] In the formula, IME represents integral quality enhancement, calculated as the methane column concentration enhancement ΔX for all pixels within the mask. CH4,i With pixel area A pixel The sum of their products:

[0033] (4)

[0034] For satellite data with a spatial resolution of 60 m, A pixel = 3600 m²; L is the plume length scale, defined as the square root of the mask area:

[0035] (5)

[0036] Where N pixel U represents the total number of pixels within the mask. eff To determine the effective wind speed, the training dataset was used to measure the wind speed U at a height of 10 meters. 10 Linear regression calibration yielded the following:

[0037] (6)

[0038] The calibration coefficients a and b are instrument and platform dependent; for EMIT satellite data, the typical value for intercept a is approximately 0.7 m·s. -1 The typical value of the slope b is about 0.23.

[0039] Preferably, the adaptive weight function α(·) in S4 is in piecewise form:

[0040] When IoU mask > 0.5 and σ wind < 1.5 m·s -1 And C consistency When α < 0.3, α = 0.2, which is a high-confidence IME pattern;

[0041] When IoU mask < 0.3 or σ wind > 3.0 m·s -1 Or C consistency When the confidence level is greater than 0.5, α = 0.8, which is a high-confidence CNN mode;

[0042] In other cases:

[0043] (8)

[0044] Where w1 = 2.0, w2 = 0.5, w3 = 1.0, sigmoid(x) = 1 / (1+e -x ), α adaptive The value range is [0.2, 0.8]; where f(Q) range ) is the source rate range adjustment factor:

[0045] ;

[0046] Weight w4 = 1.0

[0047] The fused output is:

[0048] (9)

[0049] In the five-dimensional fused feature vector

[0050] Method consistency (7),

[0051] The value range is [0, 1].

[0052] Preferably, the inverse variance weighted optimization described in S5 is implemented according to the following steps:

[0053] The relative uncertainty (dimensionless) of the CNN method is defined by the following physical basis: the main source of error in CNN estimation is the blurring of the mask boundaries (due to IoU). mask Quantization) and background noise confusion (by σ) B Both quantization and quantization can lead to CNNs underestimating or overestimating the enhancement of integral quality.

[0054] (10);

[0055] Where (1+IoU) mask Ensure IoU in the denominator position mask When σ approaches 1 CNN,rel Approaching 0 (CNN is highly reliable), IoU mask When σ = 0 CNN,rel =σ B / 10% (degraded to pure noise level); σ B / 10% normalizes the background noise to the range [0,∞); where IoU mask A higher σ indicates better mask quality and more reliable CNN estimation; B The background noise level is expressed as a percentage, with 10% in the denominator being a normalized reference value (corresponding to a moderate noise level); σ CNN,rel This represents a dimensionless relative uncertainty, with a value range of (0,1]. The smaller the value, the more reliable the CNN estimation.

[0056] The relative uncertainty (dimensionless) of the IME method is defined based on the following physical basis: the IME method uses Q... IME =IME×U eff / L calculation, U eff Directly linearly dependent on U 10 Therefore, the relative error of wind speed σ wind / U 10 The first-order approximate direct propagation is used as the relative error of the emission rate; mask quality also affects IME (erroneous mask regions change the IME integral), with 1 / (1+IoU) as the relative error. mask Factor correction:

[0057] (11);

[0058] The key innovation of equations (10) and (11) lies in: unifying the uncertainty of the two methods into an observable scenario variable (IoU).mask σ B σ wind / U 10 The analytical function enables the inverse variance weighting mechanism to respond to scene changes in real time without additional training;

[0059] In the formula σ wind / U 10 The relative uncertainty of wind speed (dimensionless) is the main source of error in the IME method; both equations have the same dimension, being dimensionless relative uncertainties, and can be directly used for inverse variance weighted calculation;

[0060] The final fusion weights are determined using the inverse variance weighting method:

[0061] (12);

[0062] The final source rate is:

[0063] (13) Methods with low uncertainty automatically receive higher weights.

[0064] Preferably, the method further includes a real data verification step S6: using observation data of methane plumes in the same region from an independent satellite platform, cross-validating the detection results of step S3 and the quantization results of step S5 respectively; detection performance is evaluated using IoU, Dice coefficient, precision, recall, and F1 score, and quantization accuracy is evaluated using RMSE, MAPE, Pearson correlation coefficient R, MAE, systematic bias, and coefficient of determination R², wherein:

[0065] (14)

[0066] (15)

[0067] (16)

[0068] (18)

[0069] In the formula, n is the total number of samples, Q predicted,i The predicted source rate (kg·h) for the i-th sample -1 ), Q true,i The true source rate (kg·h) for the i-th sample -1 ), The mean of the true values; MAPE values ​​range from [0, +∞), and the smaller the value, the higher the quantization accuracy, and the unit is percentage (%).

[0070] An automatic methane plume detection and source rate quantification system for implementing the method, comprising:

[0071] The data acquisition module is used to acquire hyperspectral satellite methane column concentration enhancement product data and reanalysis wind speed data, and to perform quality control and preprocessing.

[0072] The synthetic data generation module is used to run large eddy simulation to generate a three-dimensional methane plume distribution, and superimpose the simulated plume onto a zero-emission background scene to construct a large-scale synthetic training dataset according to the background noise level.

[0073] The plume detection module is used to train multiple deep learning semantic segmentation models based on a synthetic training dataset, select the optimal model through comparison and evaluation, and generate a binary plume mask for the input scene.

[0074] The CNN quantization submodule is used to perform regression estimation of methane point source emission rates based on a convolutional neural network with three-channel input.

[0075] The IME quantization submodule is used to perform physical inversion of methane point source emission rates based on an integral mass-enhanced physical model and effective wind speed parameters calibrated with training data.

[0076] The adaptive fusion module is used to construct a five-dimensional fused feature vector. Through piecewise adaptive weight function and inverse variance weighted optimization, it dynamically fuses the estimation results of the CNN quantization submodule and the IME quantization submodule, and outputs the final source rate quantization result and its uncertainty.

[0077] This invention also discloses an adaptive uncertainty-driven fusion method for quantifying methane point source emission rates, applicable to any method that provides mask quality IoU and background noise σ. B And satellite remote sensing platforms for reanalyzing wind speed data:

[0078] For the estimated value Q1 from the first quantization method and the estimated value Q2 from the second quantization method, an uncertainty-driven adaptive fusion is performed through the following steps:

[0079] (a) Constructing the relative uncertainty σ of the first method based on scene-observable variables 1,rel The relative uncertainty σ of the second method 2,rel , where σ 1,rel and σ 2,rel All are dimensionless quantities and are all analytically derived from the same set of observable scenario variables;

[0080] (b) Construct a preliminary fusion weight α based on a multi-dimensional scene criterion, wherein the multi-dimensional scene criterion includes at least three independent indicators reflecting different physical mechanisms;

[0081] (c) Through the inverse variance weighting mechanism α final = α·(1 / σ² 1,rel) / [α·(1 / σ² 1,rel ) + (1-α)·(1 / σ² 2,rel Refine α;

[0082] In the method described, when σ 1,rel < σ 2,rel α final If α < α, automatically increase the weight of the first method; if σ < α, automatically increase the weight of the first method. 1,rel >σ 2,rel α final > α, automatically increase the weight of the second method; the weight adjustment does not require additional training, but only relies on the real-time calculation of observable physical quantities in the current scene.

[0083] One of the key technological innovations of this invention is: (1) for the first time, mask quality IoU and wind speed relative uncertainty σ are combined. wind / U 10 and background noise σ B The three dimensions are uniformly incorporated into a composite relative uncertainty assessment framework, and the framework directly drives the dynamic adjustment of the fusion weight through the inverse variance weighting mechanism, realizing "uncertainty adaptive method fusion"; (2) Through the two-level fusion mechanism of piecewise adaptive weight function and inverse variance weighting, the advantage boundary of prior knowledge (IME physical model) and data-driven method (CNN) is explicitly modeled, which has a stronger scene adaptability than simple weighted average; (3) A synthetic dataset construction strategy based on background noise level is proposed to make up for the lack of existing synthetic datasets that ignore the diversity of noise distribution.

[0084] The beneficial effects of this invention are:

[0085] (1) Large-scale, high-quality synthetic training data was generated by WRF-LES large eddy simulation, which fundamentally solved the problem of scarce methane plume labeling data. The simulation parameters were precisely matched with the EMIT satellite observation characteristics.

[0086] (2) A piecewise adaptive weight function based on five-dimensional fusion feature vector is proposed, which can dynamically allocate the fusion weights of CNN and IME according to scene features such as mask quality, wind speed uncertainty, and background noise, thus overcoming the limitations of a single method;

[0087] (3) Introducing an inverse variance weighted optimization mechanism allows methods with lower uncertainty to automatically obtain higher weights, further improving quantification accuracy;

[0088] (4) The fusion method reduces the RMSE of the synthetic test dataset by about 30%-50% compared with the single CNN method and by about 20%-30% compared with the single IME method, while maintaining the physical interpretability of the IME method. In actual large-scale satellite scene tests, the quantization error of the fusion method is significantly lower than that of any single method, especially in scenarios with high background noise or poor wind speed data quality. Attached Figure Description

[0089] Figure 1 This is a schematic diagram of the overall process of the automatic detection and source rate quantification method for methane plumes based on hyperspectral satellite data described in this invention. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0091] This invention discloses an automatic detection and source rate quantification method for methane plumes based on hyperspectral satellite data, comprising the following steps:

[0092] Step S1: Data Acquisition and Preprocessing – Acquire methane column concentration enhancement product data (L2B level) from hyperspectral imaging satellites, perform quality control, geometric correction, and radiometric correction; acquire reanalysis wind speed data and spatially interpolate to the target resolution;

[0093] Step S2: Construction of synthetic training dataset - Using the WRF-LES large eddy simulation method, a three-dimensional methane plume distribution covering various atmospheric conditions and emission intensities is generated. Based on the background noise classification, the simulated plume is superimposed on a real satellite background scene with no emissions to construct a large-scale labeled synthetic dataset.

[0094] Step S3: Feather Detection and Mask Generation - Train multiple deep learning semantic segmentation models, select the optimal model through comprehensive evaluation of multiple indicators, and generate a binary feather mask;

[0095] Step S4: Adaptive Fusion Source Rate Quantization – Run the CNN regression module and the IME physics model module separately to obtain independent estimates, construct a five-dimensional fusion feature vector, and dynamically fuse the two estimates through a piecewise adaptive weight function;

[0096] Step S5: Inverse variance weighted optimization – Refine the fusion results based on the method uncertainty model, and output the final emission rate and uncertainty assessment.

[0097] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0098] I. Data Acquisition and Preprocessing (Step S1)

[0099] This embodiment uses the Earth Surface Mineral Dust Source Investigation (EMIT) hyperspectral imaging satellite, aboard the International Space Station, as the primary data source. EMIT carries a 224-channel hyperspectral imager, covering the 400–2500 nm spectral range with a spatial resolution of approximately 60 m. The data type used is the EMIT L2B methane column concentration enhancement product, which retrieves the methane column concentration enhancement value from the methane absorption band (primarily in the shortwave infrared (SWIR) region around 2300 nm), with units of mol·m³. -2 This can directly reveal the existence and intensity distribution of methane plumes.

[0100] Data quality control is performed according to the following process:

[0101] (1) Eliminate scenarios with cloud cover greater than 5%;

[0102] (2) Remove areas within the scenario that contain known methane emission sources (confirmed through cross-validation of the global known emission source database).

[0103] (3) Remove pixels with abnormal quality indicators or missing data;

[0104] (4) Perform geometric and radiometric corrections on the effective pixels.

[0105] Simultaneously, ERA5 reanalysis wind speed data (1-hour temporal resolution, 0.25°×0.25° spatial resolution) was acquired and interpolated to the target spatial resolution (0.1°×0.1°) using bilinear interpolation for calculating the effective wind speed in the IME physical model. The technical parameters of each data source are shown in Table 1.

[0106] Table 1 Technical parameters of each observation data

[0107]

[0108] II. Construction of Synthetic Training Dataset (Step S2)

[0109] 2.1 Screening of zero-emission background scenarios

[0110] We selected emission-free background scenes from the EMIT historical observation archive, using the following criteria: cloud cover less than 5%; no known methane emission sources within the scene; a time span of at least one year to capture seasonal variations; and geographical diversity covering various land surface types such as arid / desert, grassland / agricultural, and urban / industrial.

[0111] For each screening scenario, the standard deviation of the pixel-level methane column concentration enhancement is calculated using the following formula to quantify the background noise level σ. B :

[0112] (1)

[0113] In equation (1), x i The methane column concentration (mol·m) of pixel i -2 ), σ represents the average column density within the scene, where n is the total number of pixels. B Using global average background concentration x b = 0.011 kg·m -2 The percentage is expressed as a percentage. Background scenes are divided into three categories according to the following criteria:

[0114] Low-noise scenarios: σ B <2%, typically corresponds to arid desert regions;

[0115] Medium noise scenarios: 2% ≤ σ B < 5%, typically corresponds to grassland or agricultural areas;

[0116] High-noise scenarios: σ B ≥ 5%, usually corresponds to cities or industrial areas.

[0117] 2.2 WRF-LES Large Eddy Simulation

[0118] The Large Eddy Simulation (LES) mode of the WRF (Weather Research and Forecasting) model, based on the Navier-Stokes equations and scalar transport equations, accurately simulates atmospheric boundary layer turbulence and methane plume transport and diffusion processes by directly solving large-scale eddies and parameterizing small-scale eddies. The simulation parameters are configured as follows: horizontal spatial resolution 60 m, simulation domain size 10.5 km × 10.5 km × 2.4 km, vertical grid spacing 15 m, time step 0.25 s, and sampling interval of 30 s to obtain instantaneous plume snapshots. The methane point source is located one-third of the way upwind in the simulation domain to ensure the plume extends sufficiently downwind and is fully contained within the simulation scene. Multiple simulations cover the following parameter range: wind speed U from 1 to 10 m / s. -1 At least 10 different wind speed scenarios were set up; atmospheric stability included three boundary layer conditions: stable, neutral, and unstable; emission intensity Q ranged from 1 to 1000 kg·h. -1 Covering from small leaks (1 kg·h) -1 From super emission sources (1000 kg·h) -1 (The full range of)

[0119] 2.3 Synthetic Data Generation and Labeling

[0120] The 3D plume distribution simulated by WRF-LES was projected onto the EMIT spatial resolution (60 m horizontal) and then overlaid onto a selected emission-free background scene using the following four data augmentation operations:

[0121] Random placement: Randomly select the placement position of the plume within the background scene, requiring the plume to be completely contained within the scene to avoid boundary effects;

[0122] Random rotation: The plume is randomly rotated from 0° to 360° to simulate different wind direction conditions;

[0123] Random scaling: Slightly scales the plume by ±20% to simulate the effects of different emission intensities;

[0124] Noise matching: Adjust the superposition intensity of the simulated plume to match the signal-to-noise ratio with the noise level of the target background scene.

[0125] A corresponding binary mask label is generated for each synthetic sample: plume regions are labeled with 1, and non-plume regions are labeled with 0. The final dataset size is: training set 10,000+ samples, validation set 2,000+ samples, and test set 2,000+ samples. This ensures different noise levels (low / medium / high) and different emission intensities (1–1000 kg·h⁻¹) are met. -1 Different meteorological conditions are evenly distributed in the dataset.

[0126] III. Plume Detection and Mask Generation (Step S3)

[0127] Using the synthetic dataset constructed in step S2, train the following four mainstream deep learning semantic segmentation models:

[0128] U-Net: A classic network that uses an encoder-decoder architecture and fuses multi-scale features through skip connections, making it suitable for small-sample medical image segmentation tasks.

[0129] DeepLabV3+: Employs the Spatial Pyramid Pooling (ASPP) module to capture multi-scale contextual information through dilated convolutions with different dilation rates;

[0130] SegFormer: It adopts the Transformer encoder architecture and uses the self-attention mechanism to capture long-distance spatial dependencies;

[0131] PSPNet employs a pyramid pooling module (PPM) to fuse global context information through pooling operations at different scales.

[0132] All models use a unified training framework, including the same data augmentation strategies (random horizontal flipping, random rotation, random brightness adjustment) and the Adam optimizer (learning rate 10). -4 Early stopping strategy. Only successful detection samples with an IoU greater than 0.1 are used for training to ensure the quality of the input mask.

[0133] The models were comprehensively evaluated using the following five metrics, and the model with the best overall performance was selected: Intersection over Union (IoU), Dice coefficient, Precision, Recall, and F1 score. Typical segmentation performance using U-Net as an example is shown in Table 2.

[0134] Table 2 Performance Comparison of Typical Segmentation Models (Synthetic Test Set)

[0135]

[0136] Based on the comprehensive evaluation results, this embodiment selects SegFormer as the optimal segmentation model to generate a plume binary mask for the target satellite scene, which can be used in the subsequent quantization stage.

[0137] IV. Implementation of CNN Regression Module (Step S4 - CNN Part)

[0138] The CNN regression module employs a three-channel input architecture, fully utilizing methane-enhanced images, plume spatial distribution, and prior wind speed information for emission rate regression estimation.

[0139] Channel 1: Enhanced image of raw methane column concentration (EMIT L2B product, unit: mol·m -2 ), containing information on the spatial distribution and concentration of plumes;

[0140] Channel 2: The binary plume mask generated in step S3 (0 represents the background, 1 represents the plume), used to guide the CNN to focus on the plume region;

[0141] Channel 3: ERA5 reanalysis of wind speed at 10 meters (uniform value within the scene, unit: m·s) -1 ), providing prior information on wind speed.

[0142] The unique feature of the three-channel input combination in this invention lies in the fact that existing remote sensing methane quantification methods use single-channel or dual-channel input, employing only the original radiance image, without explicitly guiding convolutional feature extraction using a semantic segmentation mask as an independent channel, and without inputting the wind speed field as a spatial field (rather than a scalar). This invention uses a binary mask as an independent channel to guide the CNN to focus on the plume region, avoiding the impact of background interference on regression accuracy; simultaneously, it inputs wind speed as a spatial field, preserving the spatial heterogeneity of wind speed information rather than simplifying it to a scene-uniform value. Each channel in the three-channel combination undertakes an independent physical function: channel 1 (concentration image) provides a direct signal of emission intensity, channel 2 (mask) provides spatial attention guidance, and channel 3 (wind speed field) provides meteorological prior constraints. These three channels work together to form a complete physical-data dual-constraint input system.

[0143] The network architecture is based on the U-Net encoder path, truncating and connecting fully connected regression layers before the decoder. The encoder contains four downsampling stages, each containing two 3×3 convolutional layers (including batch normalization (BN) layers and ReLU activation), followed by 2×2 max pooling downsampling. The initial number of filters is 16, increasing in increments of 2, with a maximum channel depth of 256. The encoder output is compressed into a 256-dimensional feature vector by global average pooling (GAP), then passed through two fully connected layers (32 and 16 nodes respectively, both using ReLU activation), and finally output by a single-node linear output layer at the source rate Q. CNN (Unit: kg·h) -1 ).

[0144] Training uses the mean squared error (MSE) loss function:

[0145] (2)

[0146] In equation (2), n is the batch size, and Q is the batch size. predicted,i and Q true,i The predicted and actual source rates for the i-th sample are respectively (unit: kg·h). -1 Using the Adam optimizer (learning rate 10) -4 (β1 = 0.9, β2 = 0.999), batch size 32, training for 10 epochs, with early stopping strategy enabled to prevent overfitting.

[0147] V. IME Physical Model Module Implementation (Step S4 - IME Section)

[0148] The IME (Integrated Mass Enhancement) module, based on the principle of mass conservation, calculates point source emission rates by integrating the methane column concentration enhancement value within the plume region. The basic formula for IME is:

[0149] (3)

[0150] The variables in equation (3) are defined as follows:

[0151] Q IME Source rate estimated by the IME method (kg·h) -1 );

[0152] IME is the integral mass enhancement (mol·m²), calculated using the following formula:

[0153] (4)

[0154] In equation (4), M is the set of pixels within the feather mask, and ΔX CH4,i Increase the methane column concentration at pixel i (mol·m -2 A pixel For EMIT data with a spatial resolution of 60 m, where A is the area of ​​a single pixel (m²), pixel = 60 × 60 = 3600 m²;

[0155] L is the plume length scale (m), defined as the square root of the mask area:

[0156] (5)

[0157] In equation (5), N pixel L represents the total number of pixels within the mask, and physically represents the characteristic diffusion distance of the plume.

[0158] U eff Effective wind speed (m·s) -1 ), by analyzing the training dataset (U 10 U eff The data pairs were calibrated using linear regression to obtain:

[0159] (6)

[0160] In equation (6), U 10 Reanalysis of wind speed at 10 meters height (m·s) for ERA5 -1 The coefficients a and b are determined as follows: for each known Q in the training set... true , IME and U 10 The sample, U is deduced from equation (3). eff = Q true × L / IME, then for all (U 10 U eff The best-fit coefficients were obtained by performing least-squares linear regression on the data pairs. For EMIT satellite data, the typical value of the intercept α is approximately 0.7 m·s.-1 (This represents the minimum turbulent diffusion velocity of the atmospheric boundary layer at low wind speeds), and the typical value of the slope b is approximately 0.23 (dimensionless, representing the wind speed scaling factor). Note: Calibration factors a and b are instrument and platform dependent and must be recalibrated for different satellite platforms.

[0161] VI. Adaptive Fusion Strategy (Step S4 - Fusion Section)

[0162] 6.1 Construction of Feature Vector Fusion

[0163] To achieve dynamic allocation of fusion weights based on scene features, a fusion feature vector F is constructed, comprising the following five dimensions:

[0164] ① Mask quality index IoU mask Jaccard score of the plume mask, ranging from [0, 1], reflects the detection accuracy;

[0165] ② Wind speed uncertainty σ wind Standard error of ERA5 wind speed data (m·s) -1 The typical value is approximately 2 m·s. -1 This reflects the wind speed dependence risk of the IME method;

[0166] ③Source rate range Q range The source rate range is categorized as low (<500 kg·h). -1 ), medium (500-2000 kg·h) -1 High (>2000 kg·h) -1 Three levels; the physical basis is: under low emission intensity, the IME physical model has a higher signal-to-noise ratio advantage, and under extremely high emission intensity, CNN regression has better linear extrapolation ability; by adjusting the factor f(Q) range ) participates in the adaptive weight calculation in formula (8).

[0167] ④ Background noise level σ B : Standard deviation of background noise (%), reflecting the complexity of the scene;

[0168] ⑤ Method consistency C consistency The relative difference between the estimates from the two methods:

[0169] (7)

[0170] C consistency The value ranges from [0, 1]. The smaller the value, the more consistent the estimates from the two methods are, and the higher the fusion confidence.

[0171] 6.2 Adaptive Weight Function

[0172] The adaptive weight function α(·) controls the weights of the CNN method in the fusion process, and takes the following piecewise function form:

[0173] High-confidence IME pattern (α = 0.2): when IoU mask > 0.5 and σ wind < 1.5 m·s -1 And C consistency When the value is less than 0.3, it mainly relies on the IME method (IME weight 0.8), which is suitable for scenarios with high mask quality and reliable wind speed data.

[0174] High-confidence CNN mode (α = 0.8): when IoU mask < 0.3 or σ wind > 3.0 m·s -1 Or C consistency When the value is greater than 0.5, it mainly relies on the CNN method (CNN weight 0.8), which is suitable for scenarios with complex background noise or unreliable wind speed data;

[0175] Adaptive transition mode: In other cases, the sigmoid function is used to achieve a smooth transition.

[0176] (8)

[0177] In equation (8), f(Q) range ) represents the source rate range adjustment factor; w1 = 2.0, w2 = 0.5, w3 = 1.0, sigmoid(x) = 1 / (1+e -x ), α adaptive The value range is [0.2, 0.8], achieving a smooth and continuous change in weights. Final fusion output:

[0178] (9)

[0179] 6.3 Basis for determining weighting coefficients and thresholds

[0180] The selection of parameters in the adaptive weighting function is based on the following technical analysis and optimization criteria:

[0181] Determining the threshold for high / low confidence patterns:

[0182] IoU mask Thresholds of 0.5 and 0.3: IoU=0.5 is the generally accepted threshold for "qualified detection" in the field of computer vision (based on benchmarks such as PASCAL VOC); IoU=0.3 is the lower tolerance limit for "candidate detection". Using these thresholds directly for pattern judgment has clear technical basis, rather than being arbitrarily specified.

[0183] σwind Thresholds of 1.5 and 3.0 m·s -1 The root mean square error of wind speed in ERA5 reanalysis is approximately 1.0–1.5 m / s. -1 , at 1.5 m·s -1 The threshold for determining the reliability of wind speed data has a meteorological basis; 3.0 m·s -1 Approximately twice the ERA5 error, considered a conservative threshold for "unreliable wind speed data".

[0184] C consistency Thresholds 0.3 and 0.5: When the relative difference between the two methods exceeds 30%, it indicates that at least one method has a significant bias, and trust in the low-quality method should be reduced; when it exceeds 50%, it indicates that the results of the two methods seriously conflict, and a single dominant method should be switched.

[0185] The determination of the sigmoid weight coefficients w1=2.0 / w2=0.5 / w3=1.0 / w4=1.0:

[0186] The relative magnitudes of the weighting coefficients are based on the following physical analysis: IoU mask The influence on the fusion weights is most direct, therefore w1 is the largest (2.0); σ wind The impact on the weights is achieved through IME error propagation, which is indirect and weak, hence w2 is the smallest (0.5); C consistency As a consistency test item, its impact is moderate (w3=1.0); f(Q) range The auxiliary correction term has the same impact (w4=1.0). The above relative weight relationship was confirmed by grid search of the synthetic validation set (10000+ samples). Within the parameter space of w1∈[1.5,2.5], w2∈[0.3,0.8], and w3∈[0.7,1.3], the validation set RMSE is the lowest under the current configuration. The corresponding parameter sensitivity analysis shows that the RMSE is most sensitive to changes in w1 (each change of 0.5 leads to an RMSE change of about 3%), and least sensitive to w2 (about 0.8%), indicating that the weight ranking is reasonable.

[0187] VII. Inverse Variance Weighted Optimization (Step S5)

[0188] To further refine the fusion results, an inverse variance weighted optimization mechanism based on method uncertainty is introduced.

[0189] Define the relative uncertainty of CNN methods as σCNN (dimensionless), taking into account the effects of mask quality and background noise:

[0190] (10)

[0191] In equation (10), IoU maskA higher σ indicates better mask quality and more reliable CNN estimation. B The background noise level (%) is represented by σ, where 10% in the denominator is a normalized reference value (corresponding to a moderate noise level). CNN,rel This represents the dimensionless relative uncertainty, with a value range of (0,1]. The smaller the value, the more reliable the CNN estimation.

[0192] Define the relative uncertainty σ of the IME method IME (Dimensionless), taking into account the relative uncertainty of wind speed and the influence of mask quality:

[0193] (11)

[0194] In equation (11), σ wind / U 10 The relative uncertainty of wind speed (dimensionless) is the main source of error in the IME method. Both equations have the same dimension, representing dimensionless relative uncertainty, and can be directly used for inverse variance weighted calculations.

[0195] Based on the principle of inverse variance weighting, the final fusion weights are determined by the following formula:

[0196] (12)

[0197] Equation (12) ensures that the method with lower uncertainty automatically obtains higher fusion weights, and the final source rate output is: (13)

[0198] In this invention, σ CNN,rel and σ IME,rel The construction method has three special characteristics: (1) Both equations use IoU mask As a common factor, physically this is because mask quality affects the reliability of both CNN and IME: for CNN, a low-quality mask means the neural network is focusing on the wrong region; for IME, a low-quality mask means the integration region is inaccurate, σ wind / U 10 The error propagation factor is amplified accordingly; (2) the ratio σ of the two equations CNN,rel / σ IME,rel = (σ B / 10%) / (σ wind / U 10 When the background noise σ B Higher and the relative error of wind speed σ wind / U 10When the IME is lower, it automatically gets a higher weight, and vice versa, which reflects the physical rationality of “noise-driven method selection”; (3) Both formulas are dimensionless relative uncertainties with unified dimensions, which can be directly substituted into the inverse variance weighting formula without any unit conversion, thus avoiding the weight distortion problem caused by dimension mismatch.

[0199] VIII. Implementation Example – Quantitative Accuracy Assessment

[0200] Typical numerical examples are given below to verify the effectiveness of the method of the present invention.

[0201] 8.1 Evaluation Indicators

[0202] The following six quantitative accuracy indicators are used to comprehensively compare and evaluate the three methods:

[0203] (1) Root Mean Square Error (RMSE): Measures the overall level of prediction error and is more sensitive to large errors.

[0204] (14)

[0205] in, The predicted source rate for the i-th sample (unit: kg h) -1 ), The true source rate of the i-th sample (unit: kg h) -1 ), where n is the total number of samples. A smaller RMSE value indicates higher quantization accuracy, and the unit is kg / h. -1 RMSE can effectively reflect the overall level of prediction error and is particularly suitable for evaluating the overall performance of methods under different emission intensities.

[0206] (2) Mean Absolute Percentage Error (MAPE): Eliminating differences in magnitude, assessing the relative error under different emission intensities:

[0207] (15)

[0208] MAPE reflects the percentage of prediction error relative to the true value; a smaller value indicates higher quantification accuracy. MAPE is comparable across emission sources of different magnitudes and can objectively evaluate the relative error level of a method under different emission intensities. The unit of MAPE is percentage (%), and it is commonly used to compare the relative performance of different methods. In the formula, n is the total number of samples, and Q... predicted,i The predicted source rate (kg·h) for the i-th sample -1 ), Q true,i The true source rate (kg·h) for the i-th sample -1 MAPE values ​​range from [0, +∞), with smaller values ​​indicating higher quantization precision, and the unit is percentage (%).

[0209] (3) Pearson correlation coefficient (R): assesses the degree of linear correlation between predicted and actual values.

[0210] (16)

[0211] in, and These are the mean values ​​of the predicted and actual values, respectively. The Pearson correlation coefficient ranges from -1 to 1. A value closer to 1 indicates a stronger linear correlation between the predicted and actual values, meaning the method accurately reflects the trend of emission rate changes. When R approaches 1, it indicates a high correlation between the predicted and actual values, demonstrating strong predictive ability of the method.

[0212] (4) Mean Absolute Error (MAE): Evaluates the average level of prediction error and is insensitive to outliers.

[0213] (17)

[0214] A smaller MAE value indicates higher quantization accuracy, and the unit is kg h. -1 MAE is insensitive to outliers and is suitable for assessing the robustness of quantification tasks. Compared to RMSE, MAE penalizes large errors less and can more objectively reflect the average level of prediction errors, making it particularly suitable for assessing the performance of methods under normal emission intensities.

[0215] (5) Systematic bias: assessing whether there is a systematic overestimation or underestimation:

[0216] (18)

[0217] Bias reflects the systematic bias of the predicted values. The closer the value is to 0, the smaller the systematic bias, and the more accurate the prediction results. Positive bias indicates that the predicted values ​​are generally too high, suggesting a systematic overestimation by the method; negative bias indicates that the predicted values ​​are generally too low, suggesting a systematic underestimation by the method. The unit of bias is kg / h. -1 By analyzing deviations, we can understand the systematic error characteristics of the method and provide a basis for method improvement.

[0218] (6) Coefficient of determination (R²): assesses the model's explanatory power for data variation.

[0219] (19)

[0220] R² ranges from -∞ to 1. A value closer to 1 indicates a better fit to the data, meaning the predicted values ​​better explain the variation in the true values. When R² is close to 1, it means the predicted values ​​explain the changes in the true values ​​well, and the method has strong predictive power. When R² is close to 0 or negative, it means the predicted values ​​are less correlated with the true values, and the method has poor predictive power. R² is a dimensionless indicator that comprehensively reflects the overall fit of the method.

[0221] In the above formulas, n is the total number of samples, and Q predicted,i The predicted source rate (kg·h) for the i-th sample -1 ), Q true,i The true source rate (kg·h) for the i-th sample -1 ), and These are the mean of the predicted value and the actual value, respectively.

[0222] 8.2 Numerical Calculation Example

[0223] Taking a validation set containing 5 test samples as an example, the true source rate Q of each sample is... true (kg·h) -1 ) and the predicted source rate Q of the three methods pred (kg·h) -1 As shown in Table 3:

[0224] Table 3 Quantitative results of typical test samples (unit: kg·h) -1 )

[0225]

[0226] RMSE statistics: CNN=68.0, IME=48.9, fusion=37.3; the fusion method reduces RMSE by 45% compared to CNN and by 24% compared to IME.

[0227] 8.3 Example of Calculating Fusion Weights

[0228] Taking sample 1 as an example, Q CNN =128 kg·h -1 Q IME =93 kg·h -1 Q true =100 kg·h -1

[0229] Let IoU mask = 0.75, σ wind = 1.0 m·s -1 U 10 =5.0 m·s -1 , σ B =3%

[0230] Step 1: Consistency of Calculation Methods

[0231] C consistency = |128-93| / max(128,93) = 35 / 128 = 0.273.

[0232] Step 2: Determine the fusion mode

[0233] Due to IoU mask = 0.75 > 0.5, σ wind = 1.0 m·s -1 < 1.5 m·s -1 C consistency = 0.273 < 0.3, all three conditions are met, enter high confidence IME mode, α = 0.20.

[0234] Step 3: Initial Integration

[0235] Q hybrid = 0.20×128 + 0.80×93 = 25.6 +74.4 = 100.0 kg\cdotph -1

[0236] Step 4: Calculate the relative uncertainty

[0237] ;

[0238] ;

[0239] σ IME,rel (0.1143) < σ CNN,rel (0.1714) The IME method has lower uncertainty and is more reliable in this scenario.

[0240] Step 5: Calculate α using inverse variance weighting final

[0241] , ,

[0242]

[0243] Step 6: Final Output

[0244] Q final = 0.1000×128 +0.9000×93 = 12.80 +83.70 = 96.5=97 kg\cdotph -1

[0245] Error = |97-100| = 3 kg·h -1 The accuracy is better than the preliminary fusion results (Q hybrid When the value is 100.0, the error is 0, but in practice, inverse variance weighting further increases the IME weight for Q. IME In scenarios where Q is slightly lower than the true value final It will be slightly lower than Q. true This falls under the category of statistically reasonable fluctuations.

[0246] This example shows that when σ IME,rel (0.1143) < σ CNN,rel At (0.1714), the inverse variance weighting mechanism will reduce α final The fusion weight of the IME method is automatically increased (from 0.80 to 0.90) by further reducing α from 0.20 to 0.1000, reflecting the intelligent adaptive characteristic that the lower the uncertainty, the higher the weight of the method.

[0247] IX. Verification of Real Data

[0248] The method of this invention was cross-validated using real methane plume observation data (CarbonMapper satellite) from an independent hyperspectral satellite platform for a specific region. First, an optimal segmentation model was applied to the CarbonMapper data to generate a plume mask, and detection metrics (IoU, Dice, Precision, Recall, F1) were calculated. Then, three quantization methods were applied to the detected plumes, and the quantization results were compared with the reference emission rates provided by the CarbonMapper platform to calculate quantization accuracy metrics (RMSE, MAPE, Pearson R, MAE, Bias, R²) to verify the cross-platform generalization capability of this invention from synthetic data training to real-world data application.

[0249] Preliminary cross-validation results show that the plume detection model of this invention achieves an IoU of 0.71–0.78 on independent scenes of the CarbonMapper satellite (comparable to the synthetic test set), and a Dice coefficient of 0.82–0.87, indicating that the model has good cross-platform generalization ability. The Pearson correlation coefficient R between the quantization results and the CarbonMapper reference values ​​is >0.90, and the MAPE is <25%, verifying the effective transfer of the fusion quantization method from synthetic data to real satellite data. Complete cross-validation datasets and statistical results will be added in the formal implementation.

[0250] 10. Extended Applications of the Method

[0251] The technical framework of the method of this invention has good versatility and can be extended to the following application scenarios:

[0252] (1) Multi-satellite platform adaptation: This method framework can be adapted to other hyperspectral satellite platforms such as CarbonMapper, EnMAP, and PRISMA. It only requires reconfiguring the WRF-LES simulation parameters and effective wind speed calibration coefficients for different platforms based on their spatial resolution and spectral characteristics.

[0253] (2) Detection of other greenhouse gases: The method framework can be extended to the detection and quantification of other greenhouse gases such as CO2 and N2O, simply by replacing the corresponding satellite spectral products and physical quantification models;

[0254] (3) Online monitoring system integration: This method supports automated batch processing and can be integrated into the satellite data on-orbit processing system to achieve near real-time monitoring and early warning of methane point source emissions.

[0255] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Acquire methane column concentration enhancement product data from hyperspectral imaging satellites, perform quality control on the data, remove scenes with cloud cover exceeding 5%, pixels with missing data or abnormal quality indicators, and perform geometric and radiometric corrections on valid data; simultaneously acquire reanalysis wind speed data and interpolate it to the satellite data spatial resolution. S2. Construction of synthetic training dataset: A three-dimensional methane plume concentration field is generated based on the large eddy simulation method. The simulated plume is superimposed on real satellite observation data of a background scene with no emissions to construct a large-scale synthetic training dataset containing binary mask labels for the plume. The background scenes are categorized into three types based on the pixel-level methane column concentration enhancement standard deviation: low noise, medium noise, and high noise. The method for constructing a synthetic training dataset includes the following steps: (1) Enhance the standard deviation σ using methane column concentration. B The ratio to the global average background concentration is used as a classification index to divide real emission-free satellite background scenes into low noise (σ) categories. B <2%), moderate noise (2%≤σ) B <5%) and high noise (σ B ≥5%) are three categories; (2) Using the WRF-LES large eddy simulation method, covering a range of 1–10 m·s -1 Wind speed, three atmospheric stability levels (stable / neutral / unstable), 1–1000 kg·h -1 A three-dimensional methane plume concentration field is generated within the parameter space of emission intensity; (3) The simulated plume is superimposed onto the corresponding category of emission-free background scene through four data augmentation operations: random placement, random rotation, random scaling and noise matching, according to the background noise level, to generate a large-scale labeled synthetic dataset with real noise distribution characteristics; the synthetic training dataset construction method solves the core problem of the scarcity of real labeled data in methane plume detection, and ensures the matching of the synthetic data distribution with the real complex scene through background noise layering. S3. Feather Detection and Mask Generation: Using the synthetic training dataset constructed in S2, a deep learning semantic segmentation model is trained. The optimal model is selected by comparing segmentation performance indicators, and a binary feather mask is generated for the target scene. S4. Adaptive Fusion Source Rate Quantization: Based on the plume binary mask and the original methane-enhanced image generated in S3, the source rate is independently estimated using the CNN regression module and the IME physics model module, respectively, and a fusion source rate quantization is constructed, including ① mask quality IoU. mask ② Wind speed uncertainty σ wind ③ Source rate range Q range ④ Background noise level σ B Consistency with method ⑤ C consistency =|Q CNN -Q IME | / max(Q CNN Q IME The five-dimensional fusion feature vector; through IoU mask σ wind and C consistency The piecewise adaptive weight function, used as the criterion, dynamically switches between a high-confidence IME mode (α=0.2), a high-confidence CNN mode (α=0.8), and an adaptive mode with continuous sigmoid transitions, where f(Q) range The adjustment factor is physically constrained to correct the weights based on the emission intensity range; the fused output Q hybrid = α·Q CNN + (1-α)·Q IME ; S5, Q output from S4 hybrid As a result, based on scene observable variables (IoU) mask σ B σ wind U 10 ) Calculate the relative uncertainty σ of the CNN method CNN,rel =1 / (1+IoU mask )×σ B / 10% and the relative uncertainty of the IME method σ IME,rel =σ wind / U 10 ×1 / (1+IoU mask ), using the inverse variance weighting formula α final = α·(1 / σ² CNN,rel ) / [α·(1 / σ² CNN,rel ) + (1-α)·(1 / σ² IME,rel The weights are refined and integrated to output the final source rate Q. final = α final ·Q CNN + (1-α final )·Q IME And its uncertainty assessment.

2. The method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data according to claim 1, characterized in that, The large eddy simulation described in S2 uses the WRF-LES method, based on the Navier-Stokes equations and scalar transport equations. The simulation parameters are configured as follows: horizontal spatial resolution 60 m, simulation domain size 10.5 km × 10.5 km × 2.4 km, vertical grid spacing 15 m, time step 0.25 s, and sampling interval 30 s. The point source is located at one-third of the way upwind of the simulation domain to ensure that the plume extends fully downwind and is completely contained within the simulation scene. The simulation covers a wind speed range of 1–10 m / s. -1 At least 10 different wind speed scenarios, and three atmospheric stability conditions (stable, neutral, and unstable), with emission intensity ranging from 1 to 1000 kg·h. -1 ; and / or The construction of the synthetic training dataset described in S2 includes the following operations: performing four data augmentation operations on the simulated plume—random placement, random rotation, random scaling, and noise matching—and then overlaying it onto the background scene; The noise level σ in the background scene B Calculate using the following formula: （1） In the formula, x i Let methane column concentration be the value of pixel i. σ represents the average column density within the scene, where n is the total number of pixels; B Expressed as global average background concentration x b = 0.011 kg·m -2 The percentage; low-noise scenarios satisfy σ B < 2%, in moderate noise scenarios, 2% ≤ σ B < 5%, high-noise scenarios satisfy σ B ≥ 5%; the synthetic dataset should have at least 10,000 training samples, 2,000 validation samples, and 2,000 test samples.

3. The method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data according to claim 1, characterized in that, The deep learning semantic segmentation model described in S3 is one or more of U-Net, DeepLabV3+, SegFormer, or PSPNet; the optimal model is selected from the various segmentation models by comprehensively comparing and evaluating IoU, Dice coefficient, precision, recall, and F1 score; during training, only successful detection samples with a Jaccard score greater than 0.1 are used to ensure the quality of the input mask.

4. The method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data according to claim 1, characterized in that, The CNN regression module described in S4 employs a three-channel explicit functionally decoupled input architecture: Channel 1 is the original methane column concentration enhanced image to provide a direct signal of emission intensity; Channel 2 is the binary plume mask generated in S3 for spatial attention guidance; and Channel 3 is the spatially distributed 10-meter-high wind speed field to provide meteorological prior constraints. The binary mask in the three-channel input explicitly guides convolutional feature extraction to focus on the plume region in an independent channel manner, avoiding the impact of background interference on regression accuracy. The network structure is based on U-Net encoder path truncation and connection to fully connected regression layers. The encoder has 16 initial filters and a maximum channel depth of 256, containing four downsampling stages. Each stage contains two 3×3 convolutional layers and a batch normalization layer. The encoder output is compressed into a 256-dimensional feature vector by global average pooling, then passes through two fully connected layers, and finally outputs the source rate estimate in kg·h by a single-node linear output layer. -1 The training loss function is the mean squared error. （2） In the formula, n is the batch size, Q predicted,i and Q true,i Let $\frac{i}{i}$ be the predicted and true source rates for the $i$-th sample, respectively; the Adam optimizer is used with a learning rate of 10^25. -4 The training rounds consist of 10 epochs, and an early stopping strategy (patience = 3 epochs) and cosine annealing learning rate scheduling are used. Training is terminated when the validation set loss no longer decreases for 3 consecutive epochs. For large datasets (>50,000 samples), the number of training rounds can be appropriately increased to 50-100 epochs to ensure sufficient convergence, and the batch size is 32.

5. The method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data according to claim 1, characterized in that, The IME physical model module described in S4 calculates the source rate according to the following formula: （3） In the formula, IME represents integral quality enhancement, calculated as the methane column concentration enhancement ΔX for all pixels within the mask. CH4,i With pixel area A pixel The sum of their products: （4） For satellite data with a spatial resolution of 60 m, A pixel = 3600 m²; L is the plume length scale, defined as the square root of the mask area: （5） Where N pixel U represents the total number of pixels within the mask. eff To determine the effective wind speed, the training dataset was used to measure the wind speed U at a height of 10 meters. 10 Linear regression calibration yielded the following: （6） The calibration coefficients a and b are instrument and platform dependent; for EMIT satellite data, the typical value for intercept a is approximately 0.7 m·s. -1 The typical value of the slope b is about 0.

23.

6. The method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data according to claim 1, characterized in that, The adaptive weight function α(·) described in S4 is in piecewise form: When IoU mask > 0.5 and σ wind < 1.5 m·s -1 And C consistency When α < 0.3, α = 0.2, which is a high-confidence IME pattern; When IoU mask < 0.3 or σ wind > 3.0 m·s -1 Or C consistency When the confidence level is greater than 0.5, α = 0.8, which is a high-confidence CNN mode; In other cases: （8） Where w1 = 2.0, w2 = 0.5, w3 = 1.0, sigmoid(x) = 1 / (1+e^(-1 / 2)) -x ), α adaptive The value range is [0.2, 0.8]; where f(Q) range ) is the source rate range adjustment factor: ； Weight w4 = 1.0 The fused output is: （9） In the five-dimensional fused feature vector Method consistency (7), The value range is [0, 1].

7. The method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data according to claim 1 or 6, characterized in that, The inverse variance weighted optimization described in S5 is implemented through the following steps: The relative uncertainty (dimensionless) of the CNN method is defined by the following physical basis: the main source of error in CNN estimation is the blurring of the mask boundaries (due to IoU). mask Quantization) and background noise confusion (by σ) B Both quantization and quantization can lead to CNNs underestimating or overestimating the enhancement of integral quality. （10）； Where (1+IoU) mask Ensure IoU in the denominator position mask When σ approaches 1 CNN,rel Approaching 0 (CNN is highly reliable), IoU mask When σ = 0 CNN,rel =σ B / 10% (degraded to pure noise level); σ B / 10% normalizes the background noise to the range [0,∞); where IoU mask A higher σ indicates better mask quality and more reliable CNN estimation; B The background noise level is expressed as a percentage, with 10% in the denominator being a normalized reference value (corresponding to a moderate noise level); σ CNN,rel This represents a dimensionless relative uncertainty, with a value range of (0,1]. The smaller the value, the more reliable the CNN estimation. The relative uncertainty (dimensionless) of the IME method is defined based on the following physical basis: the IME method uses Q... IME =IME×U eff / L calculation, U eff Directly linearly dependent on U 10 Therefore, the relative error of wind speed σ wind / U 10 The first-order approximate direct propagation is used as the relative error of the emission rate; mask quality also affects IME (erroneous mask regions change the IME integral), with 1 / (1+IoU) as the relative error. mask Factor correction: （11）； The key innovation of equations (10) and (11) lies in: unifying the uncertainty of the two methods into an observable scenario variable (IoU). mask σ B σ wind / U 10 The analytical function enables the inverse variance weighting mechanism to respond to scene changes in real time without additional training; In the formula σ wind / U 10 The relative uncertainty of wind speed (dimensionless) is the main source of error in the IME method; both equations have the same dimension, being dimensionless relative uncertainties, and can be directly used for inverse variance weighted calculation; The final fusion weights are determined using the inverse variance weighting method: （12）； The final source rate is: (13) Methods with low uncertainty automatically receive higher weights.

8. The method for automatic detection and source rate quantification of methane plumes based on hyperspectral satellite data according to claim 1, characterized in that, The method also includes a real data verification step S6: using observation data of methane plumes in the same region from an independent satellite platform, the detection results of step S3 and the quantification results of step S5 are cross-verified respectively. Detection performance was evaluated using IoU, Dice coefficient, precision, recall, and F1 score. Quantization accuracy was evaluated using RMSE, MAPE, Pearson correlation coefficient R, MAE, systematic bias, and coefficient of determination R². （14） （15） （16） （18） In the formula, n is the total number of samples, Q predicted,i The predicted source rate (kg·h) for the i-th sample -1 ), Q true,i The true source rate (kg·h) for the i-th sample -1 ), The mean of the true values; MAPE values ​​range from [0, +∞), and the smaller the value, the higher the quantization accuracy, and the unit is percentage (%).

9. A system for automatic methane plume detection and source rate quantification implementing the method of any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire hyperspectral satellite methane column concentration enhancement product data and reanalysis wind speed data, and to perform quality control and preprocessing. The synthetic data generation module is used to run large eddy simulation to generate a three-dimensional methane plume distribution, and superimpose the simulated plume onto a zero-emission background scene to construct a large-scale synthetic training dataset according to the background noise level. The plume detection module is used to train multiple deep learning semantic segmentation models based on a synthetic training dataset, select the optimal model through comparison and evaluation, and generate a binary plume mask for the input scene. The CNN quantization submodule is used to perform regression estimation of methane point source emission rates based on a convolutional neural network with three-channel input. The IME quantization submodule is used to perform physical inversion of methane point source emission rates based on an integral mass-enhanced physical model and effective wind speed parameters calibrated with training data. The adaptive fusion module is used to construct a five-dimensional fused feature vector. Through piecewise adaptive weight function and inverse variance weighted optimization, it dynamically fuses the estimation results of the CNN quantization submodule and the IME quantization submodule, and outputs the final source rate quantization result and its uncertainty.

10. An adaptive uncertainty-driven fusion method for quantifying methane point source emission rates, applicable to any method providing mask quality IoU and background noise σ. B A satellite remote sensing platform for reanalyzing wind speed data, characterized in that: For the estimated value Q1 from the first quantization method and the estimated value Q2 from the second quantization method, an uncertainty-driven adaptive fusion is performed through the following steps: (a) Constructing the relative uncertainty σ of the first method based on scene-observable variables 1,rel The relative uncertainty σ of the second method 2,rel , where σ 1,rel and σ 2,rel All are dimensionless quantities and are all analytically derived from the same set of observable scenario variables; (b) Construct a preliminary fusion weight α based on a multi-dimensional scene criterion, wherein the multi-dimensional scene criterion includes at least three independent indicators reflecting different physical mechanisms; (c) Through the inverse variance weighting mechanism α final = α·(1 / σ² 1,rel ) / [α·(1 / σ² 1,rel ) + (1-α)·(1 / σ² 2,rel Refine α; In the method described, when σ 1,rel < σ 2,rel α final If α < α, automatically increase the weight of the first method; if σ < α, automatically increase the weight of the first method. 1,rel > σ 2,rel α final > α, automatically increase the weight of the second method; the weight adjustment does not require additional training, but only relies on the real-time calculation of observable physical quantities in the current scene.

Images