Hyperspectral methane detection method based on joint space spectrum

Abstract

This application relates to the field of hyperspectral remote sensing application technology, and provides a hyperspectral methane detection method based on spatial-spectral joint analysis. First, a sample dataset is constructed, and simultaneously, a methane detection neural network is built to extract the spatial-spectral features of the methane plume from the sample dataset. Then, a preprocessing module based on fundamental laws is established to form a hard constraint mechanism, outputting standardized spatial-spectral features of the methane plume to obtain a predicted methane concentration distribution map. A physical information penalty term is constructed to form a soft physical constraint, minimizing the error between the predicted methane concentration distribution map and the pseudo-true label, thus incentivizing the network to generate a full-resolution methane concentration distribution map. Through a deeply fused spatial-spectral joint strategy, relying on the synergistic effect of the hard constraint mechanism and the soft physical constraint, effective suppression of background noise can be achieved in complex atmospheric environments and various background interference scenarios, significantly improving the accuracy and reliability of signal detection. The detection efficiency is high, making it suitable for large-scale and efficient processing of massive hyperspectral data.

Description

Technical Field

[0001] This application relates to the field of hyperspectral remote sensing application technology, and in particular to a hyperspectral methane detection method based on spatial-spectral joint method. Background Technology

[0002] Hyperspectral remote sensing technology, with its unique advantage of "integrating image and spectrum," can simultaneously acquire spatial distribution information and continuous narrow-band spectral information of a target area. By capturing the characteristic absorption peaks of methane molecules in the near-infrared to short-wave infrared bands (such as 1642-1672nm and 2169-2378nm), this technology can achieve long-distance, large-scale "visual" monitoring of methane, and has now become the mainstream technology in the field of methane remote sensing monitoring.

[0003] Existing hyperspectral remote sensing methods for methane detection can be mainly divided into three categories: First, spectral-based inversion methods, which rely on analyzing parameters such as the depth and area of ​​methane characteristic absorption peaks to invert concentration. Typical techniques include linear spectral unmixing, matched filtering algorithms, and empirical index methods. Second, spatial-based localization methods, which utilize the spatial texture and morphological features of methane plumes (such as diffusion profiles and concentration gradients) to screen suspected areas. These methods often combine with UAV platforms for close-range inspection and locate leak points through threshold segmentation or clustering algorithms. Third, preliminary spatial-spectral combination methods, which mostly adopt a simple approach of "spectral feature extraction + spatial feature stitching" to attempt to balance inversion accuracy and source localization efficiency. These methods suffer from weak identification capabilities, low detection accuracy, and overall low detection efficiency in complex scenarios, small leaks, and situations where plume features are not obvious. They are also difficult to adapt to the large-scale and efficient processing of massive hyperspectral data. Summary of the Invention

[0004] In view of this, this application provides a hyperspectral methane detection method based on spatial-spectral integration to solve the problems of weak identification ability, low detection accuracy, and low overall detection efficiency of existing hyperspectral remote sensing methane detection methods in complex scenarios, small leaks, and when plume characteristics are not obvious, making it difficult to adapt to the large-scale and efficient processing of massive hyperspectral data.

[0005] This application provides a hyperspectral methane detection method based on spatial-spectral combination, including:

[0006] Obtain the raw hyperspectral data for methane detection, and preprocess the raw hyperspectral data to obtain the initial dataset;

[0007] A preliminary methane-enhanced spectrum was generated based on the original hyperspectral data using a standard matched filter. A methane plume mask was obtained by thresholding the preliminary methane-enhanced spectrum. A pseudo-true label for the methane concentration distribution was determined based on the methane plume mask. The pseudo-true label for the methane concentration distribution was added to the initial dataset to obtain the sample dataset.

[0008] A methane detection neural network was established, and the spatial-spectral features of methane plumes were extracted from the sample dataset using a supervised training method. The spatial-spectral features of methane plumes in the sample dataset were optimized based on a hard constraint mechanism through a basic law preprocessing module. A dedicated convolutional fusion network was used to integrate the data-driven features with the optimized spatial-spectral features of methane plumes, and a methane concentration distribution prediction map was output based on the integration result.

[0009] A loss function is constructed based on a physical information penalty term and a data-driven L1 loss term. The methane detection neural network is trained until the difference between the output methane concentration distribution prediction map and the pseudo-true label is less than a preset threshold, thus obtaining the trained methane detection neural network.

[0010] The trained methane detection neural network was used to detect methane in the hyperspectral data, resulting in a full-resolution methane concentration distribution map.

[0011] The beneficial effects of the embodiments in this application compared with the prior art are:

[0012] This application's embodiments construct a joint spatial-spectral detection framework integrating physical information and deep learning. First, an initial dataset is built using raw hyperspectral data from methane detection. A preliminary methane-enhanced spectrum is generated using a standard matched filter, and a methane plume mask is obtained through threshold filtering. This provides physically meaningful pseudo-realistic labels for model training, addressing the scarcity of labeled data in the field of hyperspectral methane detection and enabling supervised training. Second, a preprocessing module based on fundamental laws optimizes spatial-spectral features using a hard constraint mechanism. This adapts to local differences in background characteristics and atmospheric conditions, effectively suppressing background noise and enhancing weak features at the methane plume edges. This ensures effective output of standardized methane plume physical features in complex real-world scenarios, resulting in a predicted methane concentration distribution map. Finally, a physical information penalty term is introduced into the loss function, forming a soft physical constraint. This minimizes the direct error between the predicted methane concentration distribution map and the pseudo-realistic label, dynamically adjusting network training. This allows the network to maintain high detection accuracy and stability even under challenging scenarios such as complex background interference and low-concentration leakage, generating a highly accurate full-resolution methane concentration distribution map. This method simplifies the multi-step manual intervention process in traditional methods through end-to-end network design, improves the overall detection efficiency, and can adapt to the needs of large-scale and efficient processing of massive hyperspectral data, providing strong technical support for large-scale and high-precision methane remote sensing monitoring. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic flowchart of a hyperspectral methane detection method based on spatial-spectral combination provided in an embodiment of this application.

[0015] Figure 2 This is an architecture diagram of the methane detection neural network provided in the embodiments of this application.

[0016] Figure 3 This is a schematic diagram of the channel enhancement module provided in the embodiments of this application.

[0017] Figure 4 This is a schematic diagram of the channel attention module structure provided in an embodiment of this application.

[0018] Figure 5 This is a schematic diagram of the residual network module structure provided in the embodiments of this application.

[0019] Figure 6 This is a schematic diagram of the Transformer architecture for the sequence modeling module provided in this application embodiment.

[0020] Figure 7 This is a schematic diagram of the Mamba architecture of the sequence modeling module provided in this application embodiment.

[0021] Figure 8 This is a schematic diagram of the structure of the basic law preprocessing module provided in the embodiments of this application.

[0022] Figure 9 This is an architecture diagram of the methane detection neural network provided in this application embodiment after introducing a basic law preprocessing module to form a hard constraint mechanism.

[0023] Figure 10 This paper compares the effectiveness of traditional methods and the method provided in this application in a methane leak incident at an oil and gas facility. Detailed Implementation

[0024] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0025] The following describes in detail, with reference to the accompanying drawings, a hyperspectral methane detection method based on spatial-spectral combination according to an embodiment of this application.

[0026] As mentioned above, existing hyperspectral remote sensing methods for methane detection mainly include inversion methods based on the spectral dimension, positioning methods based on the spatial dimension, and preliminary spatial-spectral combination methods based on the "spectral feature extraction + spatial feature stitching" model.

[0027] Among them, the inversion method based on the spectral dimension can achieve high-precision inversion under ideal conditions. For example, the correlation coefficient between the methane index (MI) based on the data of the Airborne Visible / Infrared Imaging Spectrometer (AVIRIS) and the measured concentration can reach 0.9942. However, it has poor adaptability in complex scenarios such as unstable atmospheric conditions and severe background interference from ground objects. The characteristic signal is easily affected by noise, making it difficult to accurately separate the methane signal in the mixed spectrum, resulting in a high false negative rate for low-concentration methane.

[0028] Spatial-dimensional location methods are insufficient for accurately identifying suspected areas and locating leak points in cases of minor leaks or when plume characteristics are not obvious.

[0029] The preliminary spatial-spectral combination method failed to fully explore the potential of spatial-spectral combination and had low computational efficiency. In practical applications, it could not well balance the concentration inversion accuracy and source localization efficiency.

[0030] In view of this, this application provides a hyperspectral methane detection method based on spatial-spectral joint analysis. By deeply extracting and optimizing spatial-spectral features, and relying on the synergistic coupling of hard constraint mechanisms and soft physical constraints, this method effectively suppresses background noise and enhances the weak features of methane plumes by adapting to local differences in background characteristics and atmospheric conditions, thereby achieving accurate methane detection. This method effectively overcomes the limitations of existing methods in complex scenarios and provides an efficient and accurate solution for methane detection. This method is suitable for the large-scale and efficient processing of massive hyperspectral data in complex scenarios.

[0031] refer to Figure 1 The hyperspectral methane detection method based on spatial-spectral combination provided in this application includes the following steps:

[0032] In step S101, the raw hyperspectral data of methane detection is obtained, and the raw hyperspectral data is preprocessed to obtain the initial dataset.

[0033] In step S102, a preliminary methane-enhanced spectrum is generated based on the original hyperspectral data using a standard matched filter. The preliminary methane-enhanced spectrum is then thresholded to obtain a methane plume mask. Based on the methane plume mask, a pseudo-true label for the methane concentration distribution is determined. This pseudo-true label for the methane concentration distribution is then added to the initial dataset to obtain the sample dataset.

[0034] In step S103, a methane detection neural network is established, and the spatial-spectral features of the methane plume are extracted from the sample dataset using a supervised training method. The spatial-spectral features of the methane plume in the sample dataset are optimized based on a hard constraint mechanism through a basic law preprocessing module. A dedicated convolutional fusion network is used to integrate the data-driven features with the optimized spatial-spectral features of the methane plume, and a methane concentration distribution prediction map is output based on the integration result.

[0035] In step S104, a loss function is constructed based on the physical information penalty term and the data-driven L1 loss term to train the methane detection neural network until the difference between the output methane concentration distribution prediction map and the pseudo-true label is less than a preset threshold, thus obtaining the trained methane detection neural network.

[0036] In step S105, the trained methane detection neural network is used to detect methane in the hyperspectral data to obtain a full-resolution methane concentration distribution map.

[0037] In some embodiments of this application, the method may be executed by a server or by a terminal device with certain processing capabilities.

[0038] In some embodiments of this application, raw hyperspectral data for methane detection can be acquired, and preprocessed to obtain an initial dataset. Furthermore, a standard matched filter can be used to generate a preliminary methane-enhanced spectrum based on the raw hyperspectral data. This preliminary spectrum can be thresholded to obtain a methane plume mask. The pseudo-true labels for the dataset can be determined based on the methane plume mask, and these pseudo-true labels can be added to the initial dataset to obtain a sample dataset.

[0039] For example, the methane plume mask can be screened again, retaining the concentration of the high-confidence plume mask region and assigning it a first-class pseudo-real label, while treating the remaining regions as background and assigning a second-class pseudo-real label to the background region.

[0040] In some embodiments of this application, a methane detection neural network is used with a supervised training method to extract spatial-spectral features of methane plumes from a sample dataset. Based on this, the spatial-spectral features of the sample dataset can be optimized using a hard constraint mechanism in the fundamental law preprocessing module.

[0041] Specifically, a preprocessing module for fundamental laws can be established to form a hard constraint mechanism, and then the spatial-spectral characteristics of methane plumes can be optimized through batch-level and scenario-level parallel paths to obtain higher-precision spatial-spectral characteristics.

[0042] In some embodiments of this application, a dedicated convolutional fusion network can be used to integrate data-driven features and optimized spatial-spectral features, and output a methane concentration distribution prediction map based on the integration result. The data-driven features consist only of the network's learned features and are not derived from prior physical knowledge.

[0043] In other words, a standardized methane plume physical feature can be output by integrating the extracted multi-granularity physical features through a dedicated convolutional fusion network, and then a methane concentration distribution prediction map can be obtained based on the standardized methane plume physical feature.

[0044] In some embodiments of this application, a loss function can be constructed based on a physical information penalty term, and a pre-trained methane detection neural network can be trained based on this loss function until the difference between the output methane concentration distribution prediction map and the pseudo-true label of the dataset is less than a preset threshold, thereby obtaining the trained methane detection neural network.

[0045] Finally, the trained methane detection neural network can be used to detect the methane concentration distribution in the hyperspectral data, thereby obtaining a full-resolution methane concentration distribution map.

[0046] According to the technical solution provided in the embodiments of this application, by using a deeply integrated spatial-spectral joint strategy and combining it with a methane detection neural network, and relying on the synergistic effect of hard constraint mechanism and soft physical constraint, effective background noise suppression is achieved in complex atmospheric conditions and background interference scenarios, thus realizing accurate detection of methane signals.

[0047] In some embodiments of this application, preprocessing the raw hyperspectral data to obtain an initial dataset may include:

[0048] The original hyperspectral data was cropped and stitched into fixed-size image patches using a non-overlapping extraction strategy;

[0049] Based on the spatiotemporal separation strategy, the image patches are divided into several datasets according to different acquisition times and geographical locations;

[0050] Class balancing is performed on each dataset to make the distribution of image patches containing feather patches and pure background image patches more even in the balanced dataset;

[0051] The balanced dataset is then normalized to obtain the initial dataset.

[0052] In other words, when preprocessing raw hyperspectral data, a large-scale raw dataset can be first divided into blocks. A non-overlapping image patch extraction strategy can be used to crop and stitch the raw hyperspectral dataset into fixed-size (e.g., 128×128) image patches. This increases the number of training samples to capture subtle spectral-spatial features while preserving image patch contextual information, optimizing GPU memory management, avoiding data omissions, and maintaining sample independence.

[0053] Then, a spatiotemporal separation dataset is performed. Based on different acquisition times and geographical locations, the image patches are divided into several datasets. These datasets may include a training set, a validation set, and an independent test set. The training set is used for model training, the validation set for model optimization, and the test set for unbiased evaluation, effectively supporting the model's generalization ability and reducing the risk of data omission. The overall principle of the spatiotemporal separation strategy is complete isolation between time and geographical location.

[0054] Next, we perform class balancing on the dataset. To address the class imbalance caused by the extremely low proportion of methane plumes in hyperspectral images, we perform class balancing on the dataset to make the distribution of image patches containing plumes and pure background patches more balanced. This ensures that the model can effectively learn the subtle features of methane plumes and reduces background prediction bias.

[0055] Finally, the dataset is normalized to the [0,1] interval before being input into the network model. This ensures that all input features are at a comparable scale, stabilizes the model training process, and accelerates convergence.

[0056] In some embodiments of this application, a methane plume mask is obtained by threshold screening of a preliminary methane enhancement spectrum, including:

[0057] Using the methane-enhanced preliminary spectrum as an initial mask;

[0058] Statistical feature analysis was performed on the methane plume edges in the matched filter output, and the screening threshold was determined based on the statistical law of the fluctuation of methane concentration enhancement at the edges.

[0059] The methane-enhanced preliminary spectrum was screened using a threshold. Spectral signals with methane concentration enhancement greater than the threshold were identified as plume region signals and retained in the initial mask. Conversely, signals with methane concentration enhancement less than the threshold were removed from the initial mask, thus obtaining the methane plume mask.

[0060] In other words, to address the difficulty of obtaining pixel-level ground-based true labels for methane plumes in certain scenarios, such as space methane plumes, this application establishes a standardized process for generating pseudo-true labels: First, a standard matched filter is used to generate a preliminary methane-enhanced spectrum. A threshold screening strategy is then employed, and a threshold is determined by analyzing the plume edge features and the statistical characteristics of the matched filter output. The preliminary spectrum is then subjected to threshold screening to separate the methane signal from the background noise, resulting in a clear plume mask as a pseudo-true label, which assists the network model in accurately capturing subtle methane signals.

[0061] In some embodiments of this application, the methane detection neural network can be built using a U-shaped encoder-decoder architecture, and skip connections between different network depths can be used to preserve the fine-grained relationship between encoder-decoder layers.

[0062] The encoder extracts features from the dataset through convolutional layers to obtain methane plume features; at the same time, the encoder also optimizes the methane plume features through convolutional downsampling layers to obtain optimized methane plume features.

[0063] The decoder integrates and reconstructs the optimized methane plume features through transposed convolution upsampling and skip connections to obtain a predicted methane concentration distribution map; wherein, the transposed convolution upsampling and convolution downsampling have the same size.

[0064] In some embodiments of this application, the U-shaped encoder-decoder architecture incorporates a layered embedding of a channel enhancement module, a channel attention module, a residual network module, and a sequence modeling module. Layered embedding refers to adding a corresponding module to each layer.

[0065] The channel enhancement module has N feature channels, which process data independently and in parallel. Each feature channel adopts a standardized "convolution-Gelu-convolution" operation sequence structure. Each feature channel extracts the spatial-spectral feature representation of the target band corresponding to the methane plume in the dataset in parallel. Different feature channels are responsible for extracting the spatial-spectral feature representation corresponding to different target bands.

[0066] The channel attention module adopts a fixed operation process of "global average pooling - fully connected layer operation - sigmoid activation function" and dynamically and adaptively learns the weight coefficients of each feature channel.

[0067] The residual network module includes dedicated residual skip connections with addition operations, which adaptively integrate the original input with the spatial-spectral features extracted by the convolutional network to achieve efficient transmission of methane plume features in the network, while ensuring the effective propagation of methane plume features.

[0068] The sequence modeling module uses a Transformer architecture and employs a windowed multi-head self-attention mechanism to capture long-range spatial-spectral dependencies, extracting deep spatial-spectral physical features strongly correlated with methane plumes. Alternatively, the sequence modeling module uses a Mamba architecture, employing state-space model units to handle long-range spatial-spectral dependencies with linear complexity, extracting deep spatial-spectral physical features strongly correlated with methane plumes.

[0069] refer to Figure 2 The methane detection neural network provided in this application includes a channel enhancement module 1, a channel attention module 2, a 3×3 convolutional and Rectified Linear Unit (ReLU) layer 3, a residual network module 4, a sequence modeling module 5, a 4×4 convolutional layer with a stride of 2 6, a transposed 4×4 convolutional layer with a stride of 2 7, and a 3×3 convolutional layer 8. In the figure, gray blocks represent images, and the letters and numbers on the right side of the image, such as B×C×128×128, represent the image size.

[0070] The encoder uses 3×3 convolutional layer 8 and 4×4 convolutional layer 7 to downsample and extract and optimize the multi-scale, multi-dimensional and multi-source features of the methane plume layer layer by layer. The decoder uses 4×4 transposed convolutional upsampling and skip connections to integrate and reconstruct all the optimized features of the downsampled methane plume layer ...

[0071] refer to Figure 3 The channel enhancement module provided in this application embodiment adopts independent parallel processing of multiple feature channels. Each channel is configured with a standardized "convolution-Gelu-convolution" operation sequence structure, where ×2 indicates that the module is serially superimposed twice. Here, Z is the number of feature channels, the convolutional layer adopts a 3×3 convolutional layer, and "" indicates that it contains a sequence of network layers.

[0072] Each channel's parallel path precisely extracts the physical characteristics of the methane plume corresponding to a specific spectral band, focusing on different spectral properties and maximizing the preservation of physical feature information of the methane plume in each spectral band. The Gelu activation function is used to provide a smooth nonlinear transformation, enhancing feature discriminativeness while preserving gradient flow.

[0073] refer to Figure 4 The channel attention module provided in this application includes a one-dimensional average pooling layer, two linear layers, a ReLU layer and a Sigmoid layer. It adopts a fixed operation process of "global average pooling - fully connected layer operation - Sigmoid activation function" and can dynamically and adaptively learn the channel weights of each feature channel.

[0074] This module first performs one-dimensional average pooling along the spatial dimension, compressing the spatial information of each frequency band into a single numerical value, which represents the aggregated features of the frequency band characteristics. These features are then input into a fully connected neural network consisting of two linear layers, with the intermediate layer using the ReLU activation function. The output is processed by the Sigmoid activation function to generate a weight vector in the range of 0 to 1, with each element corresponding to the learned importance of the spectral band. These adaptive weights are then multiplied by the features of their respective frequency bands. This adaptive weighting method fundamentally overcomes the limitations of pre-set or uniform weighting in the Born-Oppenheimer approximation. The network learns the optimal weights directly from the feature interaction patterns, effectively highlighting the effective information frequency bands of methane plume physical characteristics and accurately suppressing noise interference in low-correlation frequency bands.

[0075] refer to Figure 5 The residual network module provided in this application includes two 3×3 convolutional layers, two LeakyReLU layers, and a layer normalization layer (without hard constraints). The residual network module introduces dedicated residual skip connections with addition operations, enabling efficient propagation of gradient flow in deep networks, thereby completely solving the gradient vanishing problem in deep networks while ensuring the effective propagation of methane plume characteristics.

[0076] The sequence modeling module provided in this application can be either a Transformer architecture or a Mamba architecture. Both architectures can efficiently capture long-range spatial-spectral dependencies, making up for the shortcomings of pure convolutional networks in capturing global information and extracting deep physical features strongly correlated with methane plumes. Specifically, the Transformer architecture can employ a windowed multi-head self-attention mechanism to capture long-range spatial-spectral dependencies.

[0077] refer to Figure 6 The sequence modeling module Transformer architecture provided in this application includes a layer normalization layer, a windowed multi-head self-attention mechanism layer, as well as a layer normalization layer and a linear layer, which can effectively improve the ability to capture local features of methane plumes and optimize computational efficiency.

[0078] refer to Figure 7 The Mamba architecture of the sequence modeling module provided in this application includes branches of linear layers, 4×1 convolutional layers, and SiLU layers, as well as branches of linear layers, 4×1 convolutional layers, SiLU layers, and a state space model (SSM). The two branches are concatenated and then pass through a linear layer. The concatenation method C represents concentrated concatenation.

[0079] By employing SSM to handle long-distance spatial-spectral dependencies with linear complexity, computational efficiency for high-dimensional datasets can be optimized, resulting in lower computational costs and adaptability to massive amounts of high-dimensional pixel data.

[0080] In some embodiments of this application, optimizing the spatial-spectral features of the dataset using the hard constraint mechanism of the fundamental law preprocessing module may include: acquiring the physical information of the sample dataset, and performing targeted optimization of the spatial-spectral features of the methane plume through batch-level and scene-level parallel paths, as follows:

[0081] Configure a batch-level computational processing path, and perform batch normalization processing on the physical information of the sample dataset according to the information category; use the LeakyRelu activation function to perform batch-level feature optimization on the physical information after batch normalization and correction processing in order to suppress the overall noise level of the background region;

[0082] Configure a scenario-level computational processing path, and perform application instance normalization processing on the physical information of the sample dataset according to a single scenario; use the LeakyRelu activation function to perform scenario-level feature optimization on the physical information after application instance normalization processing, so as to enhance the methane plume feature information in the corresponding scenario;

[0083] The physical information includes detection results obtained using physical methods.

[0084] The fundamental law preprocessing module provided in this application's embodiments enforces physical laws with machine precision by introducing a hard constraint P(in,o)=0 between the input (in) and output (o). P(in,o)=0 is a hard constraint formula. (Reference) Figure 8 The basic law preprocessing module provided in this application includes a batch-level processing branch and a scene-level processing branch. The batch-level processing branch includes a batch-level normalization layer, a LeakyReLU layer, and a residual network module 4. The scene-level processing branch includes a scene-level normalization layer, a LeakyReLU layer, and a residual network module 4. The outputs of the two branches are added together and then input into a residual network module 4.

[0085] refer to Figure 9 The methane detection neural network provided in this application embodiment with hard constraint mechanism introduces a basic law preprocessing module 9. The input of this model is the methane concentration enhancement after physical methods such as radiometric calibration, spectral calibration and atmospheric correction processing, and the output is connected to the subsequent methane deep feature extraction network.

[0086] The batch processing of the basic law preprocessing module 9 can be achieved by batch normalizing the physical information of the input dataset and then applying the LeakyRelu activation function during network model training. Since methane plume enhancement usually produces a signal higher than the background noise, this batch processing can suppress the background noise and amplify the real methane plume characteristics.

[0087] The process of the basic law preprocessing module 9 performing scene-level processing can be as follows: at the individual scene level of each batch of the dataset, application instance normalization processing is performed, and then the LeakyRelu activation function is used for nonlinear transformation. The physical characteristics of the methane plume are optimized for each unique environment, adapting to the local differences between background characteristics and atmospheric conditions, and strengthening the weak features of the methane plume edge.

[0088] In some embodiments of this application, using a dedicated convolutional fusion network to integrate data-driven features and optimized spatial-spectral features may include:

[0089] A dedicated convolutional fusion network is used to stitch together and fuse batch-optimized spatial-spectral features, scene-optimized spatial-spectral features, and data-driven features along the channel dimension, and adjust the dimensions to achieve seamless integration of different information sources, generate a unified representation of fusion physical principles and data-driven learning, and output standardized methane plume spatial-spectral physical features; among which, the data-driven features are network-learned features.

[0090] In other words, a dedicated convolutional fusion network splices and fuses batch-level and scene-level optimized methane plume physical features with data-driven features along the channel dimension and adjusts the dimensions, achieving seamless integration of different information sources. This generates a unified representation of fused physical principles and data-driven learning, outputting standardized methane plume physical features. This fusion network successfully achieves a unified representation of multi-granularity physical features of methane, effectively overcoming the limitations of traditional methods—in complex real-world scenarios, traditional methods often generate background noise interference, and the methane absorption characteristics may be ambiguous in the original signal.

[0091] In some embodiments of this application, the loss function may include a physical information penalty term and a data-driven L1 loss term. The physical information penalty term is... ;in, These are preset coefficients; For the residual vector, ; Indicates the number of pixels in the spectral band of the sensor. The radiance measured at that location, Pixel serial number; Indicates the pixel points during the optimization process. The methane concentration at that location increased; This represents the unattenuated radiation reaching the sensor in the absence of a target methane plume; This represents the transmittance function of the methane plume gas in the spectral band. It is the noise covariance matrix, used to characterize the spectral noise properties; This indicates transpose.

[0092] The data-driven L1 loss term is ;in, Represents pixels methane concentration at each pixel Represents pixels The methane concentration distribution corresponding to the pseudo-real label at that location. This represents the total number of pixels in the original hyperspectral image data.

[0093] In other words, the embodiments of this application employ a loss function based on physical information. To achieve soft constraints, where the coefficients... and Used to balance data fidelity and physical consistency This represents the norm of the input *in* and the output *o*. In some implementations, the physical information penalty term can be the standard objective function in matched filtering methods, penalizing outputs that deviate from the pseudo-true label of the methane concentration distribution. ;in, This indicates that the final predicted methane concentration has increased. This indicates that the methane concentration increased during the optimization process. Indicates when When the minimum is reached, the corresponding That is ; For the residual vector, .

[0094] In implementation, C can be approximated as a diagonal matrix, where the diagonal elements represent the noise variance of each band. By minimizing this term during training, the neural network is motivated to generate physically plausible methane concentration distributions, even without a directly provided explicit physical model as a label.

[0095] The physical information penalty term is combined with the data-driven L1 loss term to form the overall loss function. ;in, The data-driven L1 loss term represents the mean absolute error, which enables the model to achieve accurate data fitting by minimizing the direct error between the predicted methane concentration distribution and the pseudo-true label.

[0096] To verify the technical effectiveness of the technical solution provided in the embodiments of this application, the following experiment was designed: Hyperspectral satellites such as Gaofen-5 02 and Ziyuan-1 02D / 02E obtained coverage of 440 billion pixels (337 scenes × 1.3 × 10⁻⁶). 9 (pixels / scene) and 1.2×10 6 Rich spatial-spectral information covering 337 scenes × 3600 square kilometers per scene. The 337 scenes are divided into a total area of ​​3.1 × 10⁻⁶ square kilometers. 7 Each image patch was allocated 2.4 × 10⁻⁶ pixels for model optimization. 7 and 2.9×10 6 The samples (approximately 77.3% and 9.4%) were used for training and validation, with the remaining 4.1 × 10⁻⁶ samples used for training and validation, respectively. 6 The sample (approximately 13.3%) serves as an independent test set, providing experimental data for verifying the technical solutions provided in the embodiments of this application.

[0097] All model training and validation were performed on a computing cluster equipped with NVIDIA V100 graphics processors. The deep learning framework was implemented using PyTorch, and all training and validation runs used a fixed random seed to ensure repeatability. After training and validation, the model was deployed to extract methane concentration distribution maps from new images in the test set. The inference process begins with a preprocessing stage, which involves segmenting the original image into overlapping 128×128 pixel regions to ensure complete coverage and provide contextual information for boundaries, enabling seamless reconstruction. Each preprocessed region is fed into the trained framework for forward computation, generating predicted maps for each region. Individual predictions are stitched together based on the original spatial coordinates, and a weighted averaging strategy is used to merge the predicted pixel values ​​of overlapping regions. This method smooths discontinuities at patch boundaries, ensuring the reconstruction of full-resolution methane concentration distribution maps suitable for downstream analysis, including plume visualization, quantitative methane emission estimation, and anomaly detection.

[0098] In addition, to comprehensively evaluate the effectiveness of the methane detection method provided in the embodiments of this application, seven evaluation indicators were adopted: frequency domain energy ( Contrast ratio (CON), Root Mean Square Error (RMSE), Coefficient of Determination (R), Pearson Correlation Coefficient (PCC), Jaccard Coefficient (Jaccard), Methane Emission Rate of a Single Plume ( This paper provides a comprehensive evaluation of the background suppression and plume region accuracy of the method provided in the embodiments of this application. From the perspective of background noise evaluation, effective suppression of background noise is crucial to the reliability of methane plume detection, as it directly affects the false alarm rate and signal clarity.

[0099] Among them, frequency domain energy ( The spectral characteristics are quantified by analyzing the frequency components of the background region.

[0100] ;

[0101] in, and This indicates the width and height of the image region (in pixels). Frequency coordinates Fourier transform coefficients at the point, The square modulus of the complex coefficients is expressed by the following formula: ,in The values ​​are complex conjugates. Lower energy values ​​(especially in high-frequency components) indicate a more uniform background and less noise, as high-frequency energy typically corresponds to rapid intensity changes and noise artifacts. The background region standard deviation quantifies the dispersion of pixel values ​​within the identified background region in the reconstructed density map; a lower value indicates a more uniform background and better noise suppression.

[0102] Contrast ratio (CON) is calculated using the gray-level co-occurrence matrix (GLCM), and its intensity variation is measured based on the following:

[0103] ;

[0104] in, Indicates the number of gray levels. For GLCM elements, The grayscale value representing the reference pixel; Represents the grayscale value of a neighboring pixel at a specific distance and direction from the reference pixel; The step size is the distance between two pixels (e.g., ...). =1 indicates adjacent, =2 indicates a one-pixel separation); The direction is typically chosen as 0°, 45°, 90°, or 135°, which determines the direction in which neighboring pixels are searched starting from the reference pixel. A lower background contrast value indicates reduced texture variation and suppression of high-frequency noise, thus avoiding the simulation of plume features. From the perspective of plume region evaluation, accurate quantification and spatial definition of methane plumes are core framework objectives evaluated through multiple metrics.

[0105] Root mean square error (RMSE) is used to measure the error of all pixels within the plume region. Predicted methane concentration values Compared with the actual concentration value The average difference between them:

[0106] ;

[0107] The coefficient of determination (R) is used to quantify the proportion of variation in actual methane concentration that can be explained by the model's predictions.

[0108] ;

[0109] in, This represents the mean of the actual concentration. The closer the value is to 1, the better the match between the predicted value and the actual concentration change.

[0110] The Pearson correlation coefficient (PCC) is used to quantify the linear correlation between predicted and actual concentrations.

[0111] ;

[0112] in, This represents the mean of the predicted concentrations. A value close to 1 indicates a strong positive linear relationship and a consistent concentration trend.

[0113] The Jaccard coefficient (Jaccard) assesses the spatial accuracy of plume segmentation by quantifying the degree of overlap between the predicted plume and the actual binary plume mask.

[0114] ;

[0115] in, A true positive (a pixel correctly identified as a plume). This indicates a false positive (background pixels are incorrectly identified as plumes). This indicates a false negative (the actual plume pixel was not detected).

[0116] The methane emission rate of a single plume was determined using the integrated mass enhancement method. The rate is based on the total mass (t / h). (kg), wind speed ( (m / s) and plume length ( The formula for calculating m is as follows:

[0117] ;

[0118] First, a threshold-based segmentation technique was employed to accurately distinguish the methane plume region from the background region. During the binary mask construction stage, pixels with concentrations higher than the 95th percentile threshold were selected, and then a 3×3 median filter was used for mask smoothing. An adaptive mask algorithm was further employed to remove outliers deviating from the mean by more than one standard deviation, effectively filtering out background noise. The square root of the mask region area was set as the feature dimension. Wind speed parameters were derived based on the empirical relationship between ERA5 reanalysis data and the output of the WRF-LES66 model. For the WRF-LES model set, a 5-hour numerical simulation was conducted within a 3×3 km simulation domain at a spatial resolution of 20×20 meters. The first hour was used for model startup, and the remaining 4 hours were used for calibration and optimization of the empirical wind speed relationship. The evaluation results are shown in Table 1.

[0119] Table 1. Comparison of evaluation results between the methane detection method provided in this application and the traditional matched filter-based method.

[0120]

[0121] In summary, the method provided in this application significantly reduces both frequency domain energy (E) and contrast based on the gray-level co-occurrence matrix (CON), indicating superior background uniformity and efficient noise suppression capabilities. Furthermore, the method demonstrates superior performance across all evaluation metrics, including root mean square error (RMSE), Pearson correlation coefficient (PCC), coefficient of determination (R²), and Jaccard coefficient, fully validating its technical advantages in methane plume region detection.

[0122] To further and more intuitively verify the application effect of the method provided in this application in the methane detection task of hyperspectral remote sensing imagery, this application selects a methane leak incident at an oil and gas facility on July 1, 2025, as a real-world case for analysis. (Reference) Figure 10 A comparison of the application effects of the traditional method and the method of the present invention in a methane leak incident at an oil and gas facility shows that the traditional method, due to interference from complex observation environments and weather conditions, produces noisy detection results and is completely unable to detect the plume area; while the method of the present invention significantly reduces background noise, making the real methane plume area clearly visible, and the approximate methane concentration (in ppb) can also be determined.

[0123] In summary, the methane detection method provided in this application is adaptable to local differences in background characteristics and atmospheric conditions, effectively suppressing background noise and enhancing the weak features of the methane plume edge. It can not only accurately locate the methane leak area but also effectively identify the turbulent structure of the methane plume and capture individual vortex features and energy cascade phenomena within the plume. This method is suitable for processing and analyzing massive amounts of hyperspectral remote sensing data in complex scenarios, providing a reliable technical solution for high-precision remote sensing methane detection. It has significant practical value and guiding significance for promoting the quantitative and engineering applications of methane remote sensing detection.

[0124] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0125] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application, and should all be included within the protection scope of this application.

Claims

1. A hyperspectral methane detection method based on spatial-spectral integration, characterized in that, include: Obtain the raw hyperspectral data for methane detection, and preprocess the raw hyperspectral data to obtain an initial dataset; A preliminary methane-enhanced spectrum is generated based on the original hyperspectral data using a standard matched filter. A methane plume mask is obtained by thresholding the preliminary methane-enhanced spectrum. A pseudo-true label for the methane concentration distribution is determined based on the methane plume mask. The pseudo-true label for the methane concentration distribution is added to the initial dataset to obtain the sample dataset. A methane detection neural network was established, and the spatial-spectral features of the methane plume were extracted from the sample dataset using a supervised training method. The spatial-spectral characteristics of the methane plume in the sample dataset are optimized using the fundamental law preprocessing module based on a hard constraint mechanism. A dedicated convolutional fusion network is used to integrate data-driven features with optimized methane plume spatial-spectral features, and a predicted methane concentration distribution map is output based on the integration result. A loss function is constructed based on a physical information penalty term and a data-driven L1 loss term. The methane detection neural network is trained until the difference between the output methane concentration distribution prediction map and the pseudo-real label is less than a preset threshold, thus obtaining the trained methane detection neural network. The trained methane detection neural network was used to detect methane in the hyperspectral data, resulting in a full-resolution methane concentration distribution map.

2. The hyperspectral methane detection method based on spatial-spectral combination according to claim 1, characterized in that, The initial dataset is obtained by preprocessing the raw hyperspectral data, including: The hyperspectral raw data is cropped and stitched into fixed-size image blocks using a non-overlapping extraction strategy; Based on the spatiotemporal separation strategy, the image patch is divided into several datasets according to the differences in acquisition time and geographical location; Class balancing is performed on each dataset to make the distribution of image patches containing methane plume patches and pure background image patches more even in the balanced dataset; The balanced dataset is then normalized to obtain the initial dataset.

3. The hyperspectral methane detection method based on spatial-spectral combination according to claim 1, characterized in that, The methane plume mask is obtained by threshold screening of the preliminary methane enhancement spectrum, including: The methane-enhanced preliminary spectrum was used as an initial mask; Statistical feature analysis was performed on the methane plume edges in the matched filter output, and the screening threshold was determined based on the statistical law of the fluctuation of methane concentration enhancement at the edges. The methane-enhanced preliminary spectrum is subjected to threshold screening using the aforementioned screening threshold. Spectral signals with methane concentration enhancement greater than the threshold are identified as plume region signals and retained in the initial mask. Conversely, signals with methane concentration enhancement less than the threshold are removed from the initial mask, thus obtaining the methane plume mask.

4. The hyperspectral methane detection method based on spatial-spectral combination according to claim 1, characterized in that, The methane detection neural network is built using a U-shaped encoder-decoder architecture and employs skip connections between different network depths to preserve the fine-grained relationship between encoder-decoder layers. The encoder extracts features from the dataset through convolutional layers to obtain methane plume features; The encoder optimizes the methane plume features through a convolutional downsampling layer to obtain optimized methane plume features; The decoder integrates and reconstructs the optimized methane plume features through transposed convolutional upsampling and skip connections to obtain a predicted methane concentration distribution map. Wherein, the transposed convolution upsampling and the convolution downsampling have the same size.

5. The hyperspectral methane detection method based on spatial-spectral combination according to claim 4, characterized in that, The U-shaped encoder-decoder architecture incorporates a layered embedded channel enhancement module, a channel attention module, a residual network module, and a sequence modeling module.

6. The hyperspectral methane detection method based on spatial-spectral combination according to claim 5, characterized in that, The channel enhancement module has N feature channels, which process data independently and in parallel. Each feature channel adopts a standardized "convolution-Gelu-convolution" operation sequence structure. Each feature channel extracts the spatial-spectral feature representation of the target band corresponding to the methane plume in the dataset in parallel. Different feature channels are responsible for extracting the spatial-spectral feature representation corresponding to different target bands. The channel attention module adopts a fixed operation process of "global average pooling - fully connected layer operation - sigmoid activation function" and dynamically and adaptively learns the weight coefficients of each feature channel. The residual network module includes dedicated residual skip connections with addition operations, which adaptively integrate the original input with the spatial-spectral features extracted by the convolutional network to achieve efficient transmission of methane plume features in the network.

7. The hyperspectral methane detection method based on spatial-spectral combination according to claim 5, characterized in that, The sequence modeling module is based on the Transformer architecture and uses a windowed multi-head self-attention mechanism to capture long-range spatial-spectral dependencies and extract deep spatial-spectral features that are strongly correlated with methane plumes. Alternatively, the sequence modeling module is based on the Mamba architecture, which uses state-space model units to handle long-distance spatial-spectral dependencies with linear complexity, and extracts deep spatial-spectral features strongly correlated with methane plumes.

8. The hyperspectral methane detection method based on spatial-spectral combination according to claim 1, characterized in that, The spatial-spectral characteristics of the methane plume in the sample dataset are optimized using a fundamental law preprocessing module based on a hard constraint mechanism, including: The physical information of the sample dataset is obtained, and the spatial-spectral characteristics of the methane plume are optimized in a targeted manner through batch-level and scene-level parallel paths. Specifically, the spatial-spectral characteristics of methane plumes are optimized through a batch-level approach, including: Configure a batch-level computational processing path, and perform batch normalization processing on the physical information of the sample dataset according to the information category; use the LeakyRelu activation function to perform batch-level feature optimization on the physical information after batch normalization and correction processing in order to suppress the overall noise level of the background region; Targeted optimization of the spatial-spectral characteristics of methane plumes is performed through a scenario-level approach, including: Configure a scenario-level computational processing path, and perform application instance normalization processing on the physical information of the sample dataset according to a single scenario; use the LeakyRelu activation function to perform scenario-level feature optimization on the physical information after application instance normalization processing, so as to enhance the methane plume feature information in the corresponding scenario; The physical information includes detection results obtained using physical methods.

9. The hyperspectral methane detection method based on spatial-spectral combination according to claim 8, characterized in that, A dedicated convolutional fusion network is used to integrate data-driven features and optimized methane plume spatial-spectral features, including: A dedicated convolutional fusion network is used to stitch and fuse spatial-spectral features optimized by batch-level paths, spatial-spectral features optimized by scene-level paths, and data-driven features along the channel dimension, and adjust the dimensions to achieve seamless integration of different information sources, outputting standardized methane plume spatial-spectral features. The data-driven features are network learning features.

10. The hyperspectral methane detection method based on spatial-spectral combination according to claim 1, characterized in that, The loss function includes a physical information penalty term and a data-driven L1 loss term; The physical information penalty item is ;in, These are preset coefficients; For the residual vector, ; Indicates the number of pixels in the spectral band of the sensor. The radiance measured at that location, Pixel serial number; Indicates the pixel points during the optimization process. The methane concentration at that location increased; This represents the unattenuated radiation reaching the sensor in the absence of a target methane plume; This represents the transmittance function of the methane plume gas in the spectral band. It is the noise covariance matrix, used to characterize the spectral noise properties; Indicates transpose; The data-driven L1 loss term is ;in, Represents pixels methane concentration at each pixel Represents pixels The methane concentration distribution corresponding to the pseudo-real label at that location. This represents the total number of pixels in the original hyperspectral image data.

Images