Satellite hyperspectral inversion model construction method and system for greenhouse gas point source emission

Abstract

The application discloses a greenhouse gas point source emission satellite hyperspectral inversion model construction method and system. The method constructs a multi-branch deep neural network including a shared feature extraction layer, a physical characterization auxiliary branch and a main task predictor, jointly trains based on data alignment loss and physical residual loss, so that the network directly regresses the emission from the hyperspectral image block end to end; the physical residual loss substitutes the predicted concentration, the mask, the wind field and the emission into a preset physical equation structure, and constrains the network to meet the fluid transport law. The system includes a model construction module and an inversion module. The application avoids the error cascade amplification of the traditional serial inversion, reduces the dependence on external meteorological data, and realizes automatic inversion with high precision and high physical consistency.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing and atmospheric environment monitoring technology, specifically relating to a method and system for constructing satellite hyperspectral inversion models of greenhouse gas point source emissions. Background Technology

[0002] Point source emissions of greenhouse gases (such as methane and carbon dioxide) are a significant driver of global climate change. Stationary point sources, including oil and gas facilities, coal mines, and landfills, account for a very high proportion of these emissions, making accurate monitoring and quantification crucial for scientific and environmental governance. In recent years, with the rapid development of hyperspectral remote sensing technology, satellites equipped with high spatial resolution hyperspectral imagers can acquire continuous spectral information from the visible light to shortwave infrared bands of the Earth's surface, providing a massive data foundation for refined and high-frequency monitoring of greenhouse gas point source emissions. However, how to quickly and accurately invert point source emissions from satellite hyperspectral radiation data remains a key technical challenge in the field of atmospheric remote sensing.

[0003] Currently, satellite hyperspectral inversion of greenhouse gas point source emissions mainly employs the following technical approaches:

[0004] The first method is the traditional multi-step cascaded physical inversion method. This method typically employs an explicit serial pipeline structure: first, it uses a radiative transfer model to perform pixel-by-pixel calculations on the hyperspectral reflectance data acquired from satellites to obtain a greenhouse gas concentration enhancement map; then, it uses a manually set threshold segmentation algorithm to extract the gas plume mask from the concentration map; subsequently, it matches wind field data at the corresponding time and location from an external meteorological reanalysis database; finally, it uses the concentration enhancement, plume mask, and wind field data to estimate point source emissions using flux calculation equations such as the integral mass enhancement method. This method has the following drawbacks: due to the serial connection of each step, upstream steps (such as concentration calculation errors caused by incomplete background spectral removal and mask area errors caused by the subjectivity of threshold segmentation) will be passed downstream and nonlinearly superimposed with external wind field errors, resulting in extremely high uncertainty in the final emission estimation; at the same time, this method is highly dependent on external meteorological data, and if the meteorological data has deviations at the local microscale, the final emission calculation will be seriously inaccurate; in addition, steps such as threshold segmentation and background denoising are highly dependent on human experience, with low automation, making it difficult to meet the requirements of fully automated, near real-time processing of massive satellite data.

[0005] The second approach is a step-by-step, multi-task inversion method based on artificial intelligence. This method breaks down greenhouse gas remote sensing monitoring into three independent deep network subtasks: concentration inversion, plume segmentation, and emission estimation. Typically, synthetic datasets generated by physical simulation software are used for training. During inference, concentration maps and plume masks are output separately, and emissions are estimated by combining them with externally input wind field data. Alternatively, a concatenated approach is used: first, methane concentration increments are extracted through sparse spectral reconstruction, and then a machine learning model is trained based on samples generated by large eddy simulation for emission rate estimation. Although this method incorporates deep learning technology, it still falls within the scope of a step-by-step pipeline. The front and back-end algorithms are completely separate, failing to form a unified global optimization feedback. The reconstruction error of the concentration in the first half is still directly passed to the back-end regression model, resulting in the error cascading problem not being fundamentally solved. Furthermore, the emission rate estimation model often relies entirely on ideal synthetic data generated by atmospheric physics models for training. Because the simulated environment cannot perfectly reproduce the complex surface albedo interference and sensor noise in real remote sensing images, the model exhibits severe cross-domain distribution shifts when facing real satellite data, limiting its generalization ability. Meanwhile, existing schemes either do not incorporate physical constraints or only introduce local physical constraints in a single step, failing to integrate the conservation relationship between concentration, wind speed, and emission rate throughout the entire emission prediction process.

[0006] The third method is a point source emission estimation method based on a Gaussian plume model. This method uses satellite observation data to screen for concentration anomalies, combines this with external meteorological reanalysis datasets to extract wind field data, establishes a two-dimensional Gaussian coordinate system to construct a Gaussian plume model, and substitutes the satellite observation concentrations into the model formula to fit the parameters to estimate emissions. This method relies on strong idealized assumptions such as uniform and constant wind speed, flat terrain, and ideally normal diffusion conditions. However, real greenhouse gas plumes are often irregularly deformed due to complex terrain, atmospheric turbulence, and local micrometeorological conditions, leading to model fitting failure. Furthermore, this method requires directly substituting the wind speed parameters of the external meteorological field into the model equations; the system itself lacks error perception and correction capabilities, and is also highly dependent on prior external meteorological data.

[0007] In summary, existing greenhouse gas point source emission retrieval technologies generally suffer from the following common problems: the retrieval process is serial or multi-step splicing, with single-step errors amplifying in the final emission result; there is an absolute dependence on external meteorological wind field data, with local micro-scale meteorological biases directly transmitted to the emission estimation results; physical constraints and data-driven models are not deeply integrated, resulting in insufficient model generalization ability and physical consistency; and there is a lack of end-to-end retrieval capability that directly maps raw hyperspectral observation data to emissions, making it difficult to meet the operational application requirements of massive satellite data in terms of automation and inference efficiency. Therefore, there is an urgent need for an automated retrieval technology that can embed physical laws into the network, directly output greenhouse gas emissions from real hyperspectral observation data end-to-end, and possess both high retrieval accuracy and strong physical consistency. Summary of the Invention

[0008] The first technical solution of this application discloses a method for constructing a satellite hyperspectral inversion model of greenhouse gas point source emissions, including:

[0009] S1. Obtain satellite hyperspectral image patches containing greenhouse gas point source emission information and their corresponding supervision labels, and construct training sample pairs; the supervision labels include at least emission supervision labels;

[0010] S2. Construct a multi-branch feature-cooperative deep neural network, the network including a shared feature extraction layer, a physical representation auxiliary branch, and a main task predictor:

[0011] The shared feature extraction layer is used to extract shared hidden layer features from the input hyperspectral image patch;

[0012] The physical characterization auxiliary branch includes at least a morphology auxiliary branch, a concentration auxiliary branch, and a wind field implicit estimation branch, which are used to predict intermediate physical quantities related to emission inversion from the shared hidden layer features;

[0013] The main task predictor is used to fuse the shared hidden layer features and the features output by the physical representation auxiliary branch to output the emission prediction value.

[0014] S3. The deep neural network described in S2 is jointly trained based on data alignment loss and physical residual loss, wherein the physical residual loss is constructed by substituting the intermediate physical quantities predicted by the network and the predicted values ​​of point source emissions into a preset physical equation.

[0015] Furthermore, in the physical characterization auxiliary branch described in S2:

[0016] The morphology-assisted branch is used to output the plume mask prediction map;

[0017] The concentration-assisted branch is used to output a concentration-enhanced prediction map;

[0018] The implicit wind field estimation branch is used to output wind field prediction values.

[0019] Furthermore, the data alignment loss described in S3 includes emission loss, plume masking loss, concentration enhancement loss, and wind field loss.

[0020] Furthermore, the calculation of the physical residual loss described in S3 includes:

[0021] The integral mass enhancement is calculated based on the concentration enhancement prediction map and plume mask prediction map obtained from the network prediction.

[0022] Calculate the effective wind speed based on the wind field prediction values ​​obtained from the network prediction;

[0023] Extract the plume feature length based on the plume mask prediction map obtained from the network prediction;

[0024] Substituting the integral mass enhancement, effective wind speed, and plume characteristic length into the fluid transport equation, the emission amount estimated based on the physical equation is obtained;

[0025] The deviation between the predicted point source emissions and the emissions estimated based on the physical equations is constructed as the physical residual loss.

[0026] Furthermore, the training sample pairs also include plume mask supervision labels, concentration enhancement supervision labels, and wind field supervision labels, wherein the wind field supervision labels are calculated from the publicly available wind speed and wind direction fields.

[0027] The second technical solution of this application discloses a satellite hyperspectral inversion method for greenhouse gas point source emissions, including:

[0028] Satellite hyperspectral image patches containing greenhouse gas point source emission information for the target area are input into a trained multi-branch feature collaborative deep neural network. The network includes a shared feature extraction layer, a physical representation auxiliary branch, and a main task predictor.

[0029] The shared hidden layer features of the hyperspectral image patch are extracted by the shared feature extraction layer;

[0030] The physical characterization auxiliary branch predicts intermediate physical quantities related to emission inversion from the shared hidden layer features; wherein the physical characterization auxiliary branch includes a morphology auxiliary branch, a concentration auxiliary branch and a wind field implicit estimation branch, which respectively output plume mask prediction map, concentration enhancement prediction map and wind field prediction value;

[0031] The main task predictor fuses the shared hidden layer features with the features output by the physical representation auxiliary branch to directly output the point source emission prediction results.

[0032] Furthermore, the main task predictor fuses the shared hidden layer features with the features output by the physical representation auxiliary branch to directly output the point source emission prediction results, including:

[0033] The mask prediction map and concentration enhancement prediction map features output by the morphology-assisted branch and the concentration-assisted branch are used to weight and enhance the shared hidden layer features through the spatial attention gating module;

[0034] The enhanced features are then concatenated with the wind field prediction features output by the implicit wind field estimation branch.

[0035] The predicted point source emissions are obtained through direct regression using a multi-layer fully connected network.

[0036] Furthermore, before the main task predictor directly outputs the point source emission prediction results, it also includes:

[0037] Determine whether the plume mask prediction map output by the physical representation auxiliary branch meets the preset no-plume condition;

[0038] If so, output the zero emission value or invalid label directly and terminate the emission calculation;

[0039] If not, the main task predictor will directly output the point source emission prediction result.

[0040] The third technical solution of this application discloses a satellite hyperspectral inversion system for greenhouse gas point source emissions, comprising:

[0041] The model building module is used to execute the model building method described in the first technical solution;

[0042] The inversion module is used to execute the inversion method described in the technical solution and directly output the prediction results of point source emissions.

[0043] Furthermore, before directly outputting the point source emission prediction result, the inversion module also includes determining whether the plume mask prediction map output by the physical characterization auxiliary branch meets the preset no-plume condition; if so, it directly outputs the zero emission value or invalid mark and terminates the emission calculation; if not, the main task predictor directly outputs the point source emission prediction result.

[0044] Compared with the prior art, this application has the following beneficial effects:

[0045] (1) Avoiding error cascading amplification and improving inversion accuracy. This invention adopts an end-to-end network architecture with multi-branch feature collaboration. The intermediate physical features extracted by the auxiliary branches are only used as deep hint information to feed back into the main task predictor, rather than the explicit mathematical multiplication intermediate results in the traditional serial process. This structure enables the model to perform adaptive error compensation based on global features when facing local background noise or blurred pixels, completely avoiding the defect of single-step error being exponentially amplified to the final emission result in the traditional step-by-step inversion.

[0046] (2) Enhancing physical consistency and generalization ability. This invention introduces closed-loop physical residual loss in joint training, and substitutes the concentration enhancement, plume mask, implicit wind field and emission obtained by the network prediction into the fluid transport relationship equation to construct residual constraints. Without interfering with the forward inference of the network, the backpropagation forces the multidimensional features extracted from the latent space to numerically satisfy the mass conservation and fluid transport physics, and constrains the network weights to a reasonable physical manifold, effectively breaking the "pure black box" limitation of deep learning and significantly improving the generalization ability of the model when transferring to complex real satellite observation scenarios.

[0047] (3) Reduce external dependence and improve inference efficiency. This invention completely internalizes the matching process of empirical parameters and external meteorological data that rely on manual interaction in traditional methods into neural network weights. In the actual operational inference stage, only a single forward propagation is needed to directly output the emission scalar from the original hyperspectral image, without the need to explicitly generate intermediate concentration products or call external meteorological data to perform stepwise inversion. This breaks the absolute dependence on external high-frequency meteorological data and meets the needs of fully automated and near real-time processing of massive satellite hyperspectral data. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the model construction process for this application;

[0050] Figure 2 This is a schematic diagram of the inversion inference process. Detailed Implementation

[0051] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0052] The technical terms related to this application will be explained below.

[0053] The term "end-to-end" refers to a model that maps directly from the original input data to the final output, without the need for manual intervention or breaking the task down into multiple independent processing steps.

[0054] The term "EMIT (Earth Surface Mineral Dust Source Investigation)" refers to a scientific project that uses an imaging spectrometer aboard the International Space Station to map the composition and origin of mineral dust around the world, acquiring continuous spectral information from the visible light to the shortwave infrared bands of the Earth's surface.

[0055] The term "LES (Large Eddy Simulation)" refers to a numerical simulation method for turbulence in computational fluid dynamics. Its basic idea is to directly analyze the large-scale eddy structure and use a subgrid model to characterize the motion of small-scale eddies.

[0056] The term "CNN (Convolutional Neural Network)" refers to a convolutional neural network, a deep learning model suitable for processing images, raster data, and other data with local spatial correlations.

[0057] The term "IME (Integrated Mass Enhancement)" refers to the integrated mass enhancement method, a classic method for estimating point source emission rates by spatially integrating the concentration enhancement within a plume region and combining it with effective wind speed and plume characteristic scales.

[0058] The term "hyperspectral image patch" refers to a remote sensing image slice containing multiple continuous spectral bands. In this invention, it specifically refers to a three-dimensional data cube containing information on the characteristic absorption bands of greenhouse gases.

[0059] The term "plume mask" refers to a two-dimensional binary matrix used to identify the spatial distribution range of greenhouse gas plumes in remote sensing images, where 0 represents the background area and 1 represents the plume area.

[0060] The first embodiment of this application discloses a method for constructing a satellite hyperspectral inversion model of greenhouse gas point source emissions, such as... Figure 1 As shown, it includes the following steps:

[0061] S1. Obtain satellite hyperspectral image blocks containing greenhouse gas point source emission information and their corresponding supervision labels, and construct training sample pairs; the supervision labels include at least emission supervision labels.

[0062] The satellite hyperspectral image block containing greenhouse gas point source emission information mentioned in this step refers to a three-dimensional data cube that contains the characteristic absorption spectrum of the target greenhouse gas (such as methane and carbon dioxide) and is acquired by a hyperspectral imager (such as EMIT) mounted on a satellite platform.

[0063] This image patch can be represented as a three-dimensional tensor. ,in, and These represent the height and width of the image block, respectively. This indicates the number of spectral bands involved in the inversion. The image patch is obtained by extracting slices from the original satellite hyperspectral image that contain known point source emission targets and their downwind plume diffusion regions, ensuring that the complete greenhouse gas absorption signal and spatial diffusion pattern are included.

[0064] In a further embodiment, the above steps also include preprocessing the input image blocks. This preprocessing includes at least invalid pixel removal, spatial cropping, spectral band filtering, and band-based normalization. The spectral band filtering is used to retain the characteristic absorption information of the target greenhouse gases and remove bands that are strongly interfered with by water vapor or have excessively low signal-to-noise ratios. Band-based normalization is used to reduce the impact of differences in radiative magnitudes between different scenes on network training, enhancing the model's generalization ability to different surface backgrounds and observation geometries.

[0065] The monitoring labels mentioned in this step include at least emission monitoring labels. In a preferred embodiment, the monitoring tag further includes a plume mask monitoring tag. Concentration Enhancement Monitoring Label and wind farm monitoring labels .

[0066] The monitoring labels are derived from publicly released and quality-controlled point source plume products or source-level emission products, such as the Carbon Mapper public data product system, which includes plume-level and source-level products, publicly providing information such as plume images, concentration maps, integral mass enhancement, plume length, instantaneous emission rate, wind speed and direction, uncertainty, and quality markers. Only plumes that have passed quality control and can be attributed to a reliable point source are publicly released to ensure the reliability of the monitoring labels.

[0067] A correspondence between input image patches and supervised labels is established based on observation time, spatial location, scene identifier, plume identifier, or source identifier to construct training sample pairs. Each training sample pair includes at least one input hyperspectral image patch. and emissions monitoring labels .

[0068] In one specific implementation, emissions monitoring labels Use the plume-level instantaneous emission rate field or source-level emission rate field from publicly available products; wind farm monitoring labels. Calculated from publicly available wind speed and direction fields; plume mask monitoring label and concentration-enhanced monitoring labels Generated from publicly available plume images, concentration raster images, or corresponding boundary information. The wind field component conversion formula is:

[0069] ;

[0070] in, This indicates wind speed (m / s). Indicates wind direction (0-360 degrees).

[0071] Through the above steps, an end-to-end mapping and supervision data foundation from raw hyperspectral observations to emissions was constructed, avoiding the domain offset problem caused by relying on manual annotation or idealized physical simulation to generate pseudo-labels in traditional methods.

[0072] S2. Construct a deep neural network with multi-branch feature collaboration, wherein the network includes a shared feature extraction layer, a physical representation auxiliary branch, and a main task predictor.

[0073] The shared feature extraction layer is used to extract shared hidden layer features from the input hyperspectral image patch. This layer serves as the backbone encoder of the network, receiving the 3D hyperspectral image patch. By transforming multiple layers of features, they are mapped to shared hidden features in a high-dimensional semantic feature space. The shared hidden layer features, after encoding, simultaneously contain high-dimensional coupled information related to plume morphology, concentration enhancement distribution, and local wind field.

[0074] In specific implementations, the shared feature encoder used in the shared feature extraction layer may employ a 3D convolutional structure, a Swin-Transformer architecture based on a self-attention mechanism, or a hybrid architecture combining 3D-CNN and Transformer to achieve depth-adaptive extraction of plume spatial features.

[0075] The physical characterization auxiliary branch is connected to the shared feature extraction layer and is used to predict intermediate physical quantities related to emission inversion from the shared hidden layer features. The physical characterization auxiliary branch includes a morphology auxiliary branch, a concentration auxiliary branch, and a wind field implicit estimation branch, which output plume mask prediction map, concentration enhancement prediction map, and wind field prediction value, respectively.

[0076] The morphology-assisted branch outputs a two-dimensional feather mask prediction map with the same resolution as the input image block through an upsampling decoding network (such as a 2D deconvolution layer or a bilinear interpolation combined with a convolution layer). Furthermore, to ensure the output satisfies a probabilistic interpretation, a sigmoid activation function can be applied to the terminal output of the morphological auxiliary branch, making... , representing the probability that each pixel belongs to the feather region.

[0077] The concentration auxiliary branch outputs a two-dimensional concentration enhancement prediction map through an upsampling decoding network. This branch is used to characterize the enhancement of greenhouse gas concentration relative to background concentration within the target region. The activation function for this branch can be linear or ReLU activation to ensure a non-negative output.

[0078] The implicit wind field estimation branch directly estimates the local wind field components from shared hidden layer features through a global average pooling (GAP) layer and a fully connected layer, outputting a wind field prediction value. Through this branch, the network can learn implicit wind field dynamics from the spatial stretching and diffusion patterns of plumes, thereby reducing reliance on external meteorological reanalysis data during the inference phase.

[0079] The main task predictor is used to fuse the shared hidden layer features and the features output by the physical representation auxiliary branch to directly regress and obtain the predicted value of point source emissions. The main task predictor receives global features output from the shared feature extraction layer, wind field features output from the wind field implicit estimation branch, and deep auxiliary features output from the morphology auxiliary branch and the concentration auxiliary branch, and fuses the above features.

[0080] Specifically, the deep features of the morphology-assisted branch and the concentration-assisted branch are input into the spatial attention gating module to generate attention weights. Then, the attention weights are used to apply the shared features. Weighted enhancement is performed to obtain the enhanced backbone features. Its expression is:

[0081] ;

[0082] in, This indicates element-wise multiplication.

[0083] Enhanced backbone features Wind field characteristics output by the implicit estimation branch of the wind field Feature concatenation is performed to obtain fused features. Its expression is:

[0084] ;

[0085] in, This indicates a tensor splicing operation.

[0086] The fusion features By inputting a multi-layer fully connected network, the predicted values ​​of point source emissions can be obtained through direct regression. Therefore, instead of treating the plume mask, concentration enhancement field, and wind field as intermediate results in an explicit pipeline and calculating them step by step, this invention uses them as auxiliary physical characterizations to feed back into the main task predictor, thereby achieving end-to-end direct inversion of point source emissions. This architecture eliminates explicit numerical errors propagated from upstream stages to the final emissions, fundamentally cutting off the path of error cascading amplification.

[0087] S3. The deep neural network described in S2 is jointly trained based on data alignment loss and physical residual loss, wherein the physical residual loss is constructed by substituting the intermediate physical quantities predicted by the network and the predicted values ​​of point source emissions into a preset physical equation.

[0088] The data alignment loss is used to constrain the consistency between the main task output and the outputs of each auxiliary branch and the corresponding supervision labels. In one specific embodiment, the data alignment loss includes emission loss, plume masking loss, concentration enhancement loss, and wind field loss.

[0089] Emissions loss The mean square error can be defined as:

[0090] ;

[0091] Where N represents the number of samples in a training batch.

[0092] Plume masking loss The binary cross-entropy loss can be defined as:

[0093] ;

[0094] Concentration-increased loss The mean absolute error can be defined as:

[0095] ;

[0096] Wind farm losses The mean square error of the two-dimensional wind field components can be defined as:

[0097] .

[0098] To balance the gradient magnitudes of different task branches, a learnable uncertainty parameter is introduced in a preferred implementation. and As adaptive weights for each sub-loss term, the data alignment loss It can be represented as:

[0099] .

[0100] The physical residual loss This is used to constrain the network output to satisfy mass conservation or fluid transport relationships. Specifically, it involves using the concentration enhancement prediction map output during the network's forward propagation. Flow mask prediction diagram Wind field forecast value and emission forecasts Simultaneously, the preset physical equations are substituted to construct the closed-loop residual loss.

[0101] In one specific implementation, the preset physical equation is a fluid transport relationship equation based on the integral mass enhancement (IME). The calculation of the physical residual loss includes:

[0102] First, the integral mass enhancement is calculated based on the predicted concentration enhancement map and the predicted plume mask. :

[0103] ;

[0104] Where A represents the surface area constant corresponding to a single pixel.

[0105] Then, the effective wind speed is calculated based on the wind field prediction values. :

[0106] ;

[0107] in, A very small positive number introduced to prevent numerical instability.

[0108] plume characteristic length Predicted plume mask Extracted using a differentiable method. Preferably, the second-order central moment is first calculated based on the predicted plume mask, then the principal axis direction is determined, and the projection length of the mask along the principal axis direction is used as the plume feature length to ensure that this quantity can participate in gradient backpropagation. To avoid numerical divergence due to an excessively small denominator, further limitations can be imposed:

[0109] ;

[0110] in, This indicates the plume length extraction operator. This is the lower limit constant for length.

[0111] This allows us to obtain the emissions estimated based on physical equations:

[0112] .

[0113] Accordingly, the physical residual loss is defined as:

[0114] .

[0115] In another embodiment, the preset physical equation can also be a fluid transport equation based on the cross-sectional flux method. Specifically, the flux of the plume cross section is calculated based on the concentration enhancement prediction map and wind field prediction values ​​obtained from the network prediction; the deviation between the predicted point source emissions and the plume cross section flux is constructed as the physical residual loss. This alternative solution can also achieve physical consistency constraints, and those skilled in the art can flexibly choose according to specific application scenarios.

[0116] The total loss function is defined as: ;

[0117] in, This represents the weighting coefficient of the physical residual constraint term. It can be set to a fixed value, or an adjustment strategy that gradually increases with each training round can be adopted, so that the data distribution is learned first in the early stage of training and the physical consistency constraint is gradually strengthened in the middle and later stages of training.

[0118] Through the aforementioned joint training method, the network not only learns the mapping relationship between the input hyperspectral images and emission monitoring labels, but is also constrained in a solution space consistent with mass conservation and fluid transport laws, thereby improving the stability and physical consistency of the emission inversion results.

[0119] The second embodiment of this application discloses a satellite hyperspectral inversion method for greenhouse gas point source emissions.

[0120] Satellite hyperspectral image patches containing greenhouse gas point source emission information for the target area are input into a trained multi-branch feature-cooperative deep neural network. The network includes a shared feature extraction layer, a physical representation auxiliary branch, and a main task predictor.

[0121] The shared hidden layer features of the hyperspectral image patch are extracted by the shared feature extraction layer. The shared hidden layer features simultaneously encode spectral absorption features and plume spatial diffusion features.

[0122] The physical characterization auxiliary branch predicts intermediate physical quantities related to emission inversion from the shared hidden layer features. The physical characterization auxiliary branch includes a morphology auxiliary branch, a concentration auxiliary branch, and a wind field implicit estimation branch, which respectively output plume mask prediction maps. Concentration Enhancement Prediction Chart Wind field forecast .

[0123] The main task predictor fuses the shared hidden layer features with the features output by the physical representation auxiliary branch to directly output the point source emission prediction results.

[0124] Specifically, the process of directly outputting the point source emission prediction result by the main task predictor includes: weighting and enhancing the shared hidden layer features by using the mask prediction map and concentration enhancement prediction map features output by the morphology auxiliary branch and concentration auxiliary branch through a spatial attention gating module; concatenating the enhanced features with the wind field prediction value features output by the wind field implicit estimation branch; and obtaining the point source emission prediction result by direct regression through a multi-layer fully connected network.

[0125] In a further embodiment, before the main task predictor directly outputs the point source emission prediction result, a bypass truncation judgment step is included: determining whether the plume mask prediction map output by the physical characterization auxiliary branch meets a preset plume-free condition, that is: when the plume mask prediction map output by the morphology auxiliary branch satisfies the following formula:

[0126] .

[0127] Then it is determined that there is no valid emission plume in the current image block, where, The preset threshold is used. If the above formula is satisfied, the system directly outputs a zero emission value or an invalid flag, and terminates subsequent emission calculations; if the above formula is not satisfied, the main task predictor continues to output the point source emission prediction results. .

[0128] This bypass truncation logic effectively improves processing efficiency and reduces false alarm rate during the inference phase, making it particularly suitable for batch screening scenarios involving massive amounts of satellite data.

[0129] The third embodiment of this application discloses a satellite hyperspectral inversion system for greenhouse gas point source emissions, comprising:

[0130] The model building module is used to execute the model building method described in the first embodiment. This module is responsible for data preprocessing, network structure construction, loss function calculation, and parameter optimization.

[0131] The inversion module executes the inversion method described in the second embodiment above, directly outputting the point source emission prediction results. This module loads the trained network weights and performs end-to-end inference on the input hyperspectral image patch.

[0132] Before directly outputting the point source emission prediction result, the inversion module also includes determining whether the plume mask prediction map output by the physical characterization auxiliary branch meets the preset no-plume condition; if so, it directly outputs the zero emission value or invalid mark and terminates the emission calculation; if not, the main task predictor directly outputs the point source emission prediction result.

[0133] The system can be implemented by a computing device containing a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned model construction method or inversion method.

[0134] The technical solutions and effects of this application will be described in detail below with reference to specific embodiments.

[0135] Example 1: Model Construction.

[0136] (1) Dataset construction:

[0137] The hyperspectral raw observation data used in this embodiment can be obtained from the L1B-level radiance and geolocation product calibrated by the EMIT sensor aboard the International Space Station. This product has a spatial resolution of 60 meters, a spectral coverage of 381-2493 nm, and includes 285 bands. This data not only covers the characteristic absorption peaks of carbon dioxide and methane, but its visible light band information also provides rich basic features for stripping away surface background interference. Then, a 128×128 pixel three-dimensional image patch is extracted around the known target point source (such as a coal-fired power plant, a large coal mine, or an oil and gas extraction facility), with the image patch input dimensions fixed at 128×128×285.

[0138] Source-level emission data, publicly released on the official Carbon Mapper data portal and rigorously manually quality-checked, were used as supervisory labels for emissions and plume masks. Simultaneously, wind field data from the fifth-generation global climate reanalysis dataset ERA5, released by the European Centre for Medium-Range Weather Forecasts (ECMWF), were extracted as training supervisory labels for the implicit wind field estimation branch. Through temporal and spatial location identifier matching, this embodiment constructed 1200 real hyperspectral training sample pairs (divided into training, validation, and test sets in approximately a 7:1:2 ratio).

[0139] (2) Network parameter settings and hyperparameter adjustment:

[0140] The shared feature extraction layer employs a backbone network combining 3D convolution and self-attention mechanisms. For the input redundant spectral features of up to 285 dimensions, the 3D convolutional layer first performs spectral feature dimensionality reduction and spatial feature extraction, with its kernel size set to 3×3×3 and the number of channels increasing layer by layer to [64, 128, 256], to achieve deep representation of high-dimensional semantic features. The main task predictor consists of three fully connected layers with [512, 128, 1] neurons respectively, ultimately directly regressing to output scalar emission predictions.

[0141] During training, the optimizer used was AdamW (Adam with weight decay), and the initial learning rate was set to 1×10. -4 The cosine annealing learning rate decay strategy was adopted. The batch size was set to 32, the total number of training epochs was 200, and the uncertainty weight parameter in the data alignment loss was initialized to 1.0.

[0142] The weight coefficient λ of the physical residual loss term adopts a progressive adjustment strategy. In the early training phase (0-50 Epochs), the model primarily learns the data distribution patterns, so λ is set to 0. In the middle training phase (51-100 Epochs), to smoothly introduce physical constraints, λ linearly increases from 0 to 0.5. In the later training phase (101-200 Epochs), λ is kept constant at 0.5, ensuring the model strictly converges to the solution space that satisfies the physical laws of fluid transport while fitting the data. In one implementation, the number of convolutional channels, the size of fully connected layers, the batch size, the number of training epochs, the physical residual weight coefficient, and the learning rate scheduling strategy can be set according to the training data scale and hardware conditions.

[0143] Example 2: Model inversion inference.

[0144] like Figure 2 As shown, in one embodiment, a 128×128 pixel hyperspectral image patch can be cropped around the target point source and used as a 128×128×285 three-dimensional input tensor. The dimensions are merely illustrative and can be adjusted according to sensor resolution, plume scale, and memory conditions, but the number of bands must remain consistent with those used during training.

[0145] Then, the preprocessed 128×128×285 image patch is input into the trained network model for forward propagation. The morphological auxiliary branch in the physical representation auxiliary branch first outputs a two-dimensional plume mask prediction probability map. The system presets a bypass truncation threshold τ=50 (i.e., the threshold for the total number of effective plume pixels). The total number of pixels with a value greater than 0.5 in the probability map is counted in real time: when the total number of pixels is less than τ, it is determined that there are no significant greenhouse gas point source emission events in the area where the current image patch is located, the system directly triggers bypass truncation, outputs an emission amount of 0 kg / h and generates a "no effective plume" label, and immediately terminates the subsequent calculation process of the main task predictor. If the total number of effective mask pixels is greater than or equal to 50, the main task predictor continues to fuse the features of each branch for regression prediction. The probability threshold and the effective pixel number threshold are exemplary settings and can be determined through parameter tuning using a validation set.

[0146] The final output is a structured data package, including: core scalar data point source emission rate predictions (e.g., CH4 emission rate prediction for a coal mining area is 1450.5 kg / h), and two-dimensional concentration enhancement maps and plume mask maps (binary images) and implicit wind speed predictions for visualization and verification.

[0147] With hardware acceleration on a single graphics card, the time required to trigger the bypass cutoff of the emission-free area is extremely short, and the time required for a complete forward propagation of a single 128×128×285 image block containing effective emission sources is only about 50 milliseconds. This high inference efficiency greatly meets the needs of fully automated, near real-time, high-frequency operational monitoring of massive satellite hyperspectral data.

[0148] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for constructing a satellite hyperspectral inversion model of greenhouse gas point source emissions, characterized in that, include: S1. Obtain satellite hyperspectral image patches containing greenhouse gas point source emission information and their corresponding supervision labels, and construct training sample pairs; the supervision labels include at least emission supervision labels; S2. Construct a multi-branch feature-cooperative deep neural network, the network including a shared feature extraction layer, a physical representation auxiliary branch, and a main task predictor: The shared feature extraction layer is used to extract shared hidden layer features from the input hyperspectral image patch; The physical characterization auxiliary branch includes at least a morphology auxiliary branch, a concentration auxiliary branch, and a wind field implicit estimation branch, which are used to predict intermediate physical quantities related to emission inversion from the shared hidden layer features; The main task predictor is used to fuse the shared hidden layer features and the features output by the physical representation auxiliary branch to output the emission prediction value. S3. The deep neural network described in S2 is jointly trained based on data alignment loss and physical residual loss, wherein the physical residual loss is constructed by substituting the intermediate physical quantities predicted by the network and the predicted values ​​of point source emissions into a preset physical equation.

2. The method according to claim 1, characterized in that, In the physical characterization auxiliary branch described in S2 The morphology-assisted branch is used to output the plume mask prediction map; The concentration-assisted branch is used to output a concentration-enhanced prediction map; The implicit wind field estimation branch is used to output wind field prediction values.

3. The method according to claim 1 or 2, characterized in that, The data alignment losses described in S3 include emission losses, plume masking losses, concentration enhancement losses, and wind field losses.

4. The method according to claim 1 or 2, characterized in that, The calculation of the physical residual loss mentioned in S3 includes: The integral mass enhancement is calculated based on the concentration enhancement prediction map and plume mask prediction map obtained from the network prediction. Calculate the effective wind speed based on the wind field prediction values ​​obtained from the network prediction; Extract the plume feature length based on the plume mask prediction map obtained from the network prediction; Substituting the integral mass enhancement, effective wind speed, and plume characteristic length into the fluid transport equation, the emission amount estimated based on the physical equation is obtained; The deviation between the predicted point source emissions and the emissions estimated based on the physical equations is constructed as the physical residual loss.

5. The method according to claim 1, characterized in that, The training sample pairs also include plume mask supervision labels, concentration enhancement supervision labels, and wind field supervision labels, wherein the wind field supervision labels are calculated from the publicly available wind speed and wind direction fields.

6. A satellite hyperspectral inversion method for greenhouse gas point source emissions, characterized in that, include: A satellite hyperspectral image patch containing greenhouse gas point source emission information of the target area is input into a trained multi-branch feature collaborative deep neural network, which includes a shared feature extraction layer, a physical representation auxiliary branch, and a main task predictor. The shared hidden layer features of the hyperspectral image patch are extracted by the shared feature extraction layer; The physical characterization auxiliary branch predicts intermediate physical quantities related to emission inversion from the shared hidden layer features; wherein the physical characterization auxiliary branch includes a morphology auxiliary branch, a concentration auxiliary branch and a wind field implicit estimation branch, which respectively output plume mask prediction map, concentration enhancement prediction map and wind field prediction value; The main task predictor fuses the shared hidden layer features with the features output by the physical representation auxiliary branch to directly output the point source emission prediction results.

7. The method according to claim 6, characterized in that, The main task predictor fuses the shared hidden layer features with the features output by the physical representation auxiliary branch to directly output the point source emission prediction results, including: The mask prediction map and concentration enhancement prediction map features output by the morphology-assisted branch and the concentration-assisted branch are used to weight and enhance the shared hidden layer features through the spatial attention gating module; The enhanced features are then concatenated with the wind field prediction features output by the implicit wind field estimation branch. The predicted point source emissions are obtained through direct regression using a multi-layer fully connected network.

8. The method according to claim 7, characterized in that, Before the main task predictor directly outputs the point source emission prediction results, the following steps are also included: Determine whether the plume mask prediction map output by the physical representation auxiliary branch meets the preset no-plume condition; If so, output the zero emission value or invalid label directly and terminate the emission calculation; If not, the main task predictor will directly output the point source emission prediction result.

9. A satellite hyperspectral inversion system for greenhouse gas point source emissions, characterized in that, include: The model building module is used to execute the model building method according to any one of claims 1-5; The inversion module is used to execute the inversion method according to any one of claims 6-8 and directly output the point source emission prediction results.

10. The system according to claim 9, characterized in that, Before directly outputting the point source emission prediction result, the inversion module also includes determining whether the plume mask prediction map output by the physical characterization auxiliary branch meets the preset no-plume condition; if so, it directly outputs the zero emission value or invalid mark and terminates the emission calculation; if not, the main task predictor directly outputs the point source emission prediction result.

Images