Adaptive physical sensing methane plume hyperspectral retrieval method

Abstract

The application discloses a self-adaptive physical perception methane plume hyperspectral inversion method, belongs to the technical field of hyperspectral remote sensing data processing and atmospheric environment monitoring, is used for methane plume detection and concentration inversion, and comprises the following steps: obtaining spectral radiance data and methane plume data of a hyperspectral satellite, constructing a methane plume inversion data set, preprocessing the spectral radiance data and the methane plume data, constructing a methane plume detection model and training, constructing a methane plume concentration increment inversion model and training; inputting a to-be-predicted image into the trained methane plume detection model, inputting the to-be-predicted image into the trained methane plume concentration increment inversion model, and generating a methane concentration distribution map of the whole image. The application greatly reduces false positive false reports under a complex background through a binary classification neural network, and significantly improves the inversion accuracy of the methane plume concentration increment through the methane plume concentration increment inversion model.

Description

Technical Field

[0001] This invention discloses an adaptive physical sensing-based hyperspectral inversion method for methane plumes, belonging to the field of hyperspectral remote sensing data processing and atmospheric environment monitoring technology. Background Technology

[0002] Currently, the automatic detection and concentration inversion of methane plumes mainly rely on physical modeling and end-to-end deep learning methods, but existing technologies still have significant limitations in practical applications.

[0003] Existing technologies often employ physical models such as matched filters and multi-band / multi-temporal phase difference methods. While these methods offer strong physical interpretability, they still suffer from the following drawbacks: Weak anti-interference capability: They are extremely sensitive to surface type, background noise, and atmospheric conditions. In cloud edges, complex surface spectral interference, or weak plume scenarios, they are highly prone to false positives or missed detections. Reliance on manual intervention, resulting in low inversion efficiency: Taking EMIT's official data products as an example, the generated methane plume products typically require manual verification to remove artifacts. This manual verification process leads to significant data processing delays, severely limiting the scalability of large-scale real-time monitoring applications.

[0004] In recent years, data-driven deep learning segmentation networks (such as MPSUNet and FSDINet) have begun to be applied to automatically segment plume regions. However, existing deep learning methods face the following problems when retrieving methane plumes: **Difference between simulation and measured data (domain gap):** Due to the scarcity of real plume samples, existing models typically rely on samples generated by atmospheric dynamics simulations for training. However, the inherent differences between simulation and measured data lead to a significant increase in false alarm rates when processing complex meteorological and surface conditions in real remote sensing imagery. **False positives due to imbalanced data distribution:** In actual monitoring scenarios, obvious methane plumes are extremely rare events (data is extremely sparse); however, during model training, a large number of plume-containing samples are typically used to ensure a good fit. This significant difference between the training data and the data distribution in real-world applications makes the model prone to generating numerous false positives when processing massive amounts of blank background imagery.

[0005] In summary, existing technical solutions cannot simultaneously achieve a high level of automation in monitoring, a low false alarm rate (reliability), and inversion accuracy. In particular, when processing massive amounts of hyperspectral remote sensing data, how to effectively suppress background interference and accurately quantify weak plume signals is a technical challenge that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive physical sensing hyperspectral inversion method for methane plumes, in order to solve the problem in the prior art that it is difficult to simultaneously achieve a high level of automation in monitoring, a low false alarm rate, and inversion accuracy.

[0007] Adaptive physical sensing hyperspectral inversion methods for methane plumes include: S1. Construct a methane plume inversion dataset, including acquiring spectral radiance data and methane plume data from hyperspectral satellites, using spectral radiance data as input data and methane plume data as ground truth labels. S2. Data preprocessing, including replacing non-finite values ​​in the spectral radiance data with 0, extracting the absorption characteristic bands of methane and the reference background bands from the hyperspectral satellite, performing random horizontal or vertical flipping with a preset flipping probability, and injecting Gaussian noise with a preset scale. S3. Construct a methane plume detection model. Use spectral radiance data as input and methane plume data as ground truth labels for neural network training. Use weighted random sampling to balance positive and negative samples. Use weighted cross-entropy as the classification loss function. When the classification loss function reaches a predetermined threshold, output the trained methane plume detection model and output the predicted image containing the plume. S4. Construct a methane plume concentration increment inversion model. Use spectral radiance data as input and methane concentration increment in the methane plume data as the true label for neural network training. Use SmoothL1 loss as the loss function. When the loss function reaches a predetermined threshold, output the trained methane plume concentration increment inversion model. S5. Input the image to be predicted into the trained methane plume detection model, set the detection probability. If the predicted image contains a plume greater than the set detection probability, input the image to be predicted into the trained methane plume concentration increment inversion model, adopt a semi-overlapping sliding window strategy, traverse the entire image with a set step size for prediction, and perform pixel-by-pixel accumulation and average fusion on the overlapping areas to generate a methane concentration distribution map of the entire image.

[0008] S1 includes a methane plume inversion dataset, comprising a methane plume detection model dataset and a methane plume concentration increment inversion model dataset. The methane plume detection model dataset uses L1B spectral radiance data from a hyperspectral satellite as input data. Using methane plume data L2B as the ground truth label, L1B that spatiotemporally matches L2B is labeled as a positive sample, and data retrieved from the plume-free region in L2B is used as a negative sample, maintaining a positive-to-negative ratio of 1:1 in the training set. The methane plume concentration increment inversion model dataset includes L1B data that is spatiotemporally matched with L2B as input data, and methane concentration increments in L2B as true labels.

[0009] S2 includes, S2.1, replacing non-finite values ​​in L1B with 0; S2 includes S2.2, calculating the median and interquartile range of the pixel value matrix for each hyperspectral satellite image's spectral bands, and normalizing the data. The normalized data is then cropped to a preset range using the normalization formula: ; In the formula, For normalized data, This is the original data. To take the median, To obtain the maximum value, for interquartile range, To prevent division by zero of constants; S2 includes S2.3, based on the radiative transfer characteristics of methane molecules, extracting the bands corresponding to the absorption valley positions in the methane spectral transmittance curve from hyperspectral satellite data as absorption characteristic bands, and extracting the continuous bands adjacent to both sides of the absorption characteristic bands as reference background bands. S2 includes S2.4, and the preset flip probability. Preset scale With probability Randomly flip horizontally or vertically, on a scale of Inject Gaussian noise.

[0010] S3 includes constructing a methane plume detection model, which sequentially includes a physical sensing spectral gating module, a ResNet-18 neural network, a pooling layer, a fully connected layer, and a binary classifier.

[0011] Will After weighted random sampling, the data is input into the physical sensing spectral gating module, which then obtains the spectral vector through global average pooling. : ; Will Input two fully connected layers and calculate the unnormalized score of the fully connected layer output. : ; In the formula, This is the weight matrix for the second fully connected layer. This is the weight matrix of the first fully connected layer. It is the ReLU activation function; Using the addition operation layer Methane injection absorption prior bias : ; In the formula, To inject unnormalized scores with prior bias; The weights are obtained through the Sigmoid activation function. ; ; The final output processed by the physical sensing spectral gating module is calculated using an element-wise multiplier. : .

[0012] Will The ResNet-18 neural network, pooling layer, fully connected layer and binary classifier are input sequentially, and the output image predicts whether the plume is present. Neural network training uses weighted cross-entropy as a constraint.

[0013] S4 includes, S4.1, and... For each sample, randomly generate A candidate cropping window, using Calculate the sum of the true values ​​of the plume concentration within each clipping window. For methane concentration data in L2B, select the cropping window with the largest total concentration as the image patch for training. S4 includes, S4.2, performing a logarithmic transformation on the truth labels: ; In the formula, These are the truth labels for the methane plume concentration increment inversion model structure.

[0014] S4 includes S4.3, which constructs a methane plume concentration increment inversion model, including stem convolution, encoder, decoder, skip connections and residual and group normalization modules. The residual and group normalization modules are stacked convolutional blocks, each of which introduces residual shortcuts and group normalization.

[0015] S4 includes S4.4, setting dynamic weights based on pixel density, with the weight calculation formula as follows: ; In the formula, These are the concentration values ​​after quantile normalization. and For hyperparameters; S4 includes S4.5, the loss function. Weighted Smooth L1 loss is used: ; In the formula, The set of all pixel locations involved in the loss calculation. is the x-coordinate of the pixel. y is the ordinate of the pixel. For in pixels place , Pixels predicted by the methane plume concentration increment inversion model The methane concentration value at that location, For real pixels The methane concentration value at that location, This describes the processing procedure for the SmoothL1 loss function.

[0016] S5 includes employing a semi-overlapping sliding window strategy to traverse the entire image for prediction with a set step size, performing pixel-by-pixel accumulation and averaging fusion on the overlapping areas, and generating a methane concentration distribution map of the entire image.

[0017] Compared with existing technologies, this invention has the following advantages: By constructing a cascaded network architecture that integrates physical priors and deep learning, this invention significantly improves the accuracy and efficiency of methane plume detection and concentration inversion; this invention first utilizes a binary classification network with a physically-aware spectral gating module for efficient plume screening, and then achieves high-precision concentration inversion by combining a methane plume concentration increment inversion model with intelligent pruning, weighted loss function and semi-overlapping sliding window strategy. At the same time, with the addition of robust data preprocessing and enhancement techniques, the model has stronger robustness, generalization ability and adaptability to multi-source satellite data. Attached Figure Description

[0018] Figure 1 This is a flowchart of the technology of this invention; Figure 2 This is a flowchart of the methane plume detection model technology of the present invention; Figure 3 This is a flowchart of the methane plume concentration increment inversion model technology of the present invention; Figure 4 This is the ROC curve of the methane plume detection model; Figure 5 This is the PR curve of the methane plume detection model; Figure 6 This is a true label plot of methane plume concentration for the first set of test samples; Figure 7 This is the inversion prediction diagram of the methane plume concentration increment for the first set of test samples; Figure 8 This is a true value label plot of methane plume concentration for the second set of test samples; Figure 9 This is the inversion prediction graph of the methane plume concentration increment for the second set of test samples; Figure 10 This is a true value label plot of methane plume concentration for the third set of test samples; Figure 11This is the inversion prediction graph of the methane plume concentration increment for the third set of test samples; Figure 12 This is a true value label plot of methane plume concentration for the fourth set of test samples; Figure 13 This is the inversion prediction diagram of the methane plume concentration increment for the fourth group of test samples. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Adaptive physical sensing hyperspectral inversion methods for methane plumes include: S1. Construct a methane plume inversion dataset, including acquiring spectral radiance data and methane plume data from hyperspectral satellites, using spectral radiance data as input data and methane plume data as ground truth labels. S2. Data preprocessing, including replacing non-finite values ​​in the spectral radiance data with 0, extracting the absorption characteristic bands of methane and the reference background bands from the hyperspectral satellite, performing random horizontal or vertical flipping with a preset flipping probability, and injecting Gaussian noise with a preset scale. S3. Construct a methane plume detection model. Use spectral radiance data as input and methane plume data as ground truth labels for neural network training. Use weighted random sampling to balance positive and negative samples. Use weighted cross-entropy as the classification loss function. When the classification loss function reaches a predetermined threshold, output the trained methane plume detection model and output the predicted image containing the plume. S4. Construct a methane plume concentration increment inversion model. Use spectral radiance data as input and methane concentration increment in the methane plume data as the true label for neural network training. Use SmoothL1 loss as the loss function. When the loss function reaches a predetermined threshold, output the trained methane plume concentration increment inversion model. S5. Input the image to be predicted into the trained methane plume detection model, set the detection probability. If the predicted image contains a plume greater than the set detection probability, input the image to be predicted into the trained methane plume concentration increment inversion model, adopt a semi-overlapping sliding window strategy, traverse the entire image with a set step size for prediction, and perform pixel-by-pixel accumulation and average fusion on the overlapping areas to generate a methane concentration distribution map of the entire image.

[0021] S1 includes a methane plume inversion dataset, comprising a methane plume detection model dataset and a methane plume concentration increment inversion model dataset. The methane plume detection model dataset uses L1B spectral radiance data from a hyperspectral satellite as input data. Using methane plume data L2B as the ground truth label, L1B that spatiotemporally matches L2B is labeled as a positive sample, and data retrieved from the plume-free region in L2B is used as a negative sample, maintaining a positive-to-negative ratio of 1:1 in the training set. The methane plume concentration increment inversion model dataset includes L1B data that is spatiotemporally matched with L2B as input data, and methane concentration increments in L2B as true labels.

[0022] S2 includes, S2.1, replacing non-finite values ​​in L1B with 0; S2 includes S2.2, calculating the median and interquartile range of the pixel value matrix for each hyperspectral satellite image's spectral bands, and normalizing the data. The normalized data is then cropped to a preset range using the normalization formula: ; In the formula, For normalized data, This is the original data. To take the median, To obtain the maximum value, for interquartile range, To prevent division by zero of constants; S2 includes S2.3, based on the radiative transfer characteristics of methane molecules, extracting the bands corresponding to the absorption valley positions in the methane spectral transmittance curve from hyperspectral satellite data as absorption characteristic bands, and extracting the continuous bands adjacent to both sides of the absorption characteristic bands as reference background bands. S2 includes S2.4, and the preset flip probability. Preset scale With probability Randomly flip horizontally or vertically, on a scale of Inject Gaussian noise.

[0023] S3 includes constructing a methane plume detection model, which sequentially includes a physical sensing spectral gating module, a ResNet-18 neural network, a pooling layer, a fully connected layer, and a binary classifier.

[0024] Will After weighted random sampling, the data is input into the physical sensing spectral gating module, which then obtains the spectral vector through global average pooling. : ; Will Input two fully connected layers and calculate the unnormalized score of the fully connected layer output. : ; In the formula, This is the weight matrix for the second fully connected layer. This is the weight matrix of the first fully connected layer. It is the ReLU activation function; Using the addition operation layer Methane injection absorption prior bias : ; In the formula, To inject unnormalized scores with prior bias; The weights are obtained through the Sigmoid activation function. ; ; The final output processed by the physical sensing spectral gating module is calculated using an element-wise multiplier. : .

[0025] Will The ResNet-18 neural network, pooling layer, fully connected layer and binary classifier are input sequentially, and the output image predicts whether the plume is present. Neural network training uses weighted cross-entropy as a constraint.

[0026] S4 includes, S4.1, and... For each sample, randomly generate A candidate cropping window, using Calculate the sum of the true values ​​of the plume concentration within each clipping window. For methane concentration data in L2B, select the cropping window with the largest total concentration as the image patch for training. S4 includes, S4.2, performing a logarithmic transformation on the truth labels: ; In the formula, These are the truth labels for the methane plume concentration increment inversion model structure.

[0027] S4 includes S4.3, which constructs a methane plume concentration increment inversion model, including stem convolution, encoder, decoder, skip connections and residual and group normalization modules. The residual and group normalization modules are stacked convolutional blocks, each of which introduces residual shortcuts and group normalization.

[0028] S4 includes S4.4, setting dynamic weights based on pixel density, with the weight calculation formula as follows: ; In the formula, These are the concentration values ​​after quantile normalization. and For hyperparameters; S4 includes S4.5, the loss function. Weighted Smooth L1 loss is used: ; In the formula, The set of all pixel locations involved in the loss calculation. is the x-coordinate of the pixel. y is the ordinate of the pixel. For in pixels place , Pixels predicted by the methane plume concentration increment inversion model The methane concentration value at that location, For real pixels The methane concentration value at that location, This describes the processing procedure for the SmoothL1 loss function.

[0029] S5 includes employing a semi-overlapping sliding window strategy to traverse the entire image for prediction with a set step size, performing pixel-by-pixel accumulation and averaging fusion on the overlapping areas, and generating a methane concentration distribution map of the entire image.

[0030] The following description, in conjunction with the accompanying drawings, further illustrates the process of this invention. Figure 1 As shown, EMIT remote sensing L1B spectral data and EMIT remote sensing L2B plume data products were acquired. Then, a methane plume inversion dataset was constructed, and the EMIT remote sensing data was preprocessed and enhanced. A methane plume detection model was constructed for coarse screening to determine whether a methane plume exists. If not, the methane plume detection model was returned for further coarse screening. If it exists, a methane plume concentration increment inversion model was constructed for fine measurement, and finally the methane plume concentration distribution was obtained. The specific operation process for negative sample retrieval is as follows: First, observation scenes marked as having no plume were screened from the L2B methane plume products. Then, non-overlapping sampling was performed in the above scenes using the same sliding window size as the positive samples.

[0031] The technical process of methane plume detection model is as follows: Figure 2As shown, the loss function threshold for the methane plume detection model is set to 0.01. EMIT hyperspectral data is input into the model, and weighted random sampling is performed during the training phase before being input into the Physical Sensing Spectral Gating Module. For the trained model, EMIT hyperspectral data is input into the model, first into the Physical Sensing Spectral Gating Module, and then sequentially processed through a ResNet-18 neural network, global mean pooling, and a binary classifier to obtain the methane plume filtering result (present / absent methane plume). The Physical Sensing Spectral Gating Module internally contains one global average pooling layer, two fully connected layers, one ReLU activation function layer, one addition operation layer, and one Sigmoid activation function layer. Finally, the generated weight coefficients are applied to the input spectral radiance data through an element-wise multiplier. The ResNet-18 neural network contains an initial 7×7 convolutional layer (with a stride of 2), a 3×3 max pooling layer, and four cascaded residual stages (each stage contains two basic residual blocks, each residual block consisting of two 3×3 convolutional layers, a normalization layer, and a ReLU layer). The binary classifier contains a global average pooling layer, a Dropout layer (with a dropout rate of 0.5), a fully connected linear layer with an output dimension of 2, and a Softmax layer.

[0032] The technical process of methane plume concentration increment inversion model is as follows: Figure 3 As shown, the threshold of the loss function for the methane plume concentration increment inversion model was set to 0.01. EMIT hyperspectral data was input into the model, and plume-focused patch sampling was performed during the training phase. Then, the data was input... Stem convolution; during the inference phase, EMIT hyperspectral data undergoes semi-overlapping sliding window inference, skipping... Stem convolution directly introduces residual shortcuts and group normalization in the U-Net neural network; the trained model is then processed using EMIT hyperspectral data input. Stem convolution is applied, followed by residual shortcut and group normalization using a U-Net neural network to obtain the methane plume concentration distribution regression prediction result. During the training phase, a concentration-aware pixel-level loss is also output. The stem convolution consists of a 1×1 convolutional layer, a group normalization layer, and a SiLU activation function layer. In this embodiment, each residual and group normalization module consists of two stacked convolutional blocks. The internal structure of each convolutional block sequentially includes a 3×3 depthwise separable convolutional layer (reducing the number of parameters), a group normalization layer, a SiLU activation function layer, a 3×3 point convolutional layer, and a residual shortcut. The residual shortcut ensures the stability of deep gradients by performing pixel-level addition operations on the block input and block output. The encoder contains four consecutive downsampling units, each consisting of a 3×3 convolutional layer with a stride of 2 connected to the aforementioned residual and group normalization modules. The decoder contains four consecutive upsampling units, each consisting of a bilinear interpolation upsampling layer, a 1×1 convolutional layer, a skip connection fusion layer, and a residual and group normalization module. The output layer consists of a 1×1 convolutional layer appended to the decoder, mapping the feature map to a single-channel methane concentration increment distribution map.

[0033] The following is a further explanation with reference to an example. EMITL1B spectral data and corresponding L2B methane plume products were collected from August 2022 to June 2024. Due to missing and outlier values ​​in the original spectral data, to prevent numerical instability in subsequent calculations, non-finite values ​​(such as NaN, Inf) were replaced with 0 to avoid risks such as gradient explosion or loss function failure caused by non-finite values ​​in gradient calculations. The L1B data contains 285 bands (380-3500nm). The preset range for cropping normalized data is: Flipping probability Gaussian noise injection at a preset scale Preprocessing is performed. The bands corresponding to the absorption valleys in the methane spectral transmittance curves are extracted from hyperspectral satellite data as absorption characteristic bands. Adjacent continuous bands on either side of the absorption characteristic bands are extracted as reference background bands. The absorption characteristic bands refer to the continuous band regions where the transmittance of the satellite spectral signal through methane gas exhibits a significant concave shape with wavelength. In this invention, the methane absorption characteristic bands selected are from 2250 nm to 2350 nm. The reference background bands refer to the band regions located on either side of the absorption characteristic bands, which are weakly affected by methane absorption, and where the satellite spectral signal mainly reflects the surface reflectance characteristics. In this invention, the reference background bands selected are 400 nm, 2100 nm, 2490 nm, 1900 nm–1950 nm, and 2400 nm–2450 nm.

[0034] A methane plume detection model was constructed (with ResNet-18 as the backbone network), and 6410 samples were prepared (641 positive samples and 5769 negative samples). During training, the positive and negative ratio was forced to be 1:1. The output determines whether an image patch contains a plume.

[0035] A methane plume concentration increment inversion model was constructed, with EMIT hyperspectral remote sensing data as input. The ingress layer used a 1×1 stem convolution to compress 285 channels to 64 channels. This was followed by 4 downsampling layers and 4 upsampling layers, each containing "convolution + GroupNorm + SiLU" and residual connections. During training, the following settings were configured... , For each image, 10 candidate boxes are randomly generated. The sum of the L2B true values ​​within each box is calculated, and the box with the largest sum is selected as the input for this iteration. The SmoothL1 Loss between the predicted value and the logarithmically transformed true value is calculated and multiplied by the weight W.

[0036] A large slice of the image to be tested is fed into a trained methane plume detection model. If the detection probability is greater than 0.5, it is fed into a trained methane plume concentration increment inversion model. The inversion model uses a sliding window prediction with a step size of half the window size (i.e., a semi-overlapping sliding window strategy) and outputs the average value of the overlapping area. Finally, the methane concentration distribution map of the region is output (unit: ppm·m).

[0037] The above implementation method was validated on a test set. First, the methane plume detection model was compared with the traditional method using a matched filter and the Swin-TransformerV2 model. The comparison results are shown in Table 1. Table 1. Performance Comparison of Methane Plume Detection Models ; Experimental results show that the methane plume detection model significantly outperforms the benchmark method in ROC-AUC (area under the receiver operating characteristic curve), PR-AUC (area under the precision-recall curve), and F1 score, demonstrating the effectiveness of the method of this invention. When processing hyperspectral data, the methane plume detection model captures more methane-specific features, effectively reducing false positives. Figure 4 and Figure 5 The ROC (Receiving Operating Characteristic) curve and PR (Precision) curve of the methane plume detection model were further demonstrated. The curves show that the methane plume detection model maintains higher precision at a high recall rate, making it suitable for the low false positive requirements in actual monitoring.

[0038] Since most current methane plume inversion methods are modeled and validated using simulated images, they are not comparable to the hyperspectral remote sensing data of this invention. Therefore, this invention still compares with the classic methane plume inversion method, matched filter, and the standard U-Net data-driven pixel-level regression method. Table 2 shows the performance comparison of the inversion models: Table 2. Performance Comparison of Methane Plume Concentration Increment Inversion Models ; Experimental results show that, in actual remote sensing monitoring data training and validation scenarios, the methane plume concentration increment inversion model proposed in this invention reduces the MAE (mean absolute error) by approximately 20% compared to the standard U-Net method before improvement, while increasing the positive pixel Pearson correlation coefficient to 0.85. Although the traditional matched filter method maintains a good level of Pearson correlation coefficient, its MAE at the pixel level is significantly lower than that of data-driven methods due to the limitations of the method's accuracy. These results demonstrate that the methane plume concentration increment inversion model constructed according to this invention has higher inversion accuracy for methane plume concentration increments.

[0039] Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 , Figure 11 , Figure 12 , Figure 13 The visualization of the results from different methods for processing four sets of data shows a comparison between the prediction results of the methane plume concentration increment inversion model on real EMIT remote sensing data and the ground truth labels. As can be seen in the figure, the method of this invention accurately captures the plume shape and overall concentration change trend, with errors mainly distributed at the low-concentration edges and a very small number of high-concentration points.

[0040] Figure 6 A true value label plot of methane plume concentration for the first set of test samples; Figure 7 This is the inversion prediction diagram of the methane plume concentration increment for the first set of test samples. Figure 6 and Figure 7 This demonstrates the model's inversion capability for typical high-concentration point source plumes. A comparison shows that the inversion prediction diagram of this invention (…) Figure 7 Not only did it accurately reproduce the true value ( Figure 6 The spatial morphology of the plume and the diffusion direction of the plume head were observed, and the numerical fitting of the core high-concentration region showed extremely high consistency.

[0041] Figure 8 A true label plot of methane plume concentration for the second set of test samples; Figure 9This is an inversion prediction diagram of the methane plume concentration increment for the second set of test samples. Figure 8 and Figure 9 The model's inversion performance in handling diffuse, weak-signal plumes is demonstrated. Although Figure 8 The plume signal in the middle is weak and the edges are blurred. Figure 9 The prediction results show that the model can still effectively suppress background noise and clearly depict the boundary contour of the plume, verifying the model's ability to capture weak targets.

[0042] Figure 10 A true value label plot of methane plume concentration for the third group of test samples; Figure 11 This is an inversion prediction diagram of the methane plume concentration increment for the third set of test samples. Figure 10 and Figure 11 This demonstrates the robustness of the inversion in complex surface contexts. (Comparison) Figure 10 The truth value of , Figure 11 The prediction results accurately present the plume concentration while maintaining a clean zero-value output in complex background areas without plumes, proving that the physical sensing module effectively reduces false positives caused by surface interference.

[0043] Figure 12 A true value label plot of methane plume concentration for the fourth group of test samples; Figure 13 This is the inversion prediction diagram of the methane plume concentration increment for the fourth group of test samples. Figure 12 and Figure 13 This demonstrates a refined inversion for plumes with complex geometries. The prediction results of this invention ( Figure 13 It can keenly capture the truth value. Figure 12 The subtle changes in the plume concentration gradient demonstrate the effectiveness of the loss function based on dynamic pixel concentration weights in solving the long-tailed distribution problem.

[0044] The above 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 or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive physics-sensing hyperspectral inversion method for methane plumes, characterized in that, include: S1. Construct a methane plume inversion dataset, including acquiring spectral radiance data and methane plume data from hyperspectral satellites, using spectral radiance data as input data and methane plume data as ground truth labels. S2. Data preprocessing, including replacing non-finite values ​​in the spectral radiance data with 0, extracting the absorption characteristic bands of methane and the reference background bands from the hyperspectral satellite, performing random horizontal or vertical flipping with a preset flipping probability, and injecting Gaussian noise with a preset scale. S3. Construct a methane plume detection model. Use spectral radiance data as input and methane plume data as ground truth labels for neural network training. Use weighted random sampling to balance positive and negative samples. Use weighted cross-entropy as the classification loss function. When the classification loss function reaches a predetermined threshold, output the trained methane plume detection model and output the predicted image containing the plume. S4. Construct a methane plume concentration increment inversion model. Use spectral radiance data as input and methane concentration increment in the methane plume data as the true label for neural network training. Use SmoothL1 loss as the loss function. When the loss function reaches a predetermined threshold, output the trained methane plume concentration increment inversion model. S5. Input the image to be predicted into the trained methane plume detection model, set the detection probability. If the predicted image contains a plume greater than the set detection probability, input the image to be predicted into the trained methane plume concentration increment inversion model, adopt a semi-overlapping sliding window strategy, traverse the entire image with a set step size for prediction, and perform pixel-by-pixel accumulation and average fusion on the overlapping areas to generate a methane concentration distribution map of the entire image.

2. The adaptive physical sensing hyperspectral inversion method for methane plumes according to claim 1, characterized in that, S1 includes a methane plume inversion dataset, comprising a methane plume detection model dataset and a methane plume concentration increment inversion model dataset. The methane plume detection model dataset uses L1B spectral radiance data from a hyperspectral satellite as input data. Using methane plume data L2B as the ground truth label, L1B that spatiotemporally matches L2B is labeled as a positive sample, and data retrieved from the plume-free region in L2B is used as a negative sample, maintaining a positive-to-negative ratio of 1:1 in the training set. The methane plume concentration increment inversion model dataset includes L1B data that is spatiotemporally matched with L2B as input data, and methane concentration increments in L2B as true labels.

3. The adaptive physical sensing hyperspectral inversion method for methane plumes according to claim 2, characterized in that, S2 includes, S2.1, replacing non-finite values ​​in L1B with 0; S2 includes S2.2, calculating the median and interquartile range of the pixel value matrix for each hyperspectral satellite image's spectral bands, and normalizing the data. The normalized data is then cropped to a preset range using the normalization formula: ； In the formula, For normalized data, This is the original data. To take the median, To obtain the maximum value, for interquartile range, To prevent division by zero of constants; S2 includes S2.3, based on the radiative transfer characteristics of methane molecules, extracting the bands corresponding to the absorption valley positions in the methane spectral transmittance curve from hyperspectral satellite data as absorption characteristic bands, and extracting the continuous bands adjacent to both sides of the absorption characteristic bands as reference background bands. S2 includes S2.4, and the preset flip probability. Preset scale With probability Randomly flip horizontally or vertically, on a scale of Inject Gaussian noise.

4. The adaptive physical sensing methane plume hyperspectral inversion method according to claim 3, characterized in that, S3 includes constructing a methane plume detection model, which sequentially includes a physical sensing spectral gating module, a ResNet-18 neural network, a pooling layer, a fully connected layer, and a binary classifier.

5. The adaptive physical sensing methane plume hyperspectral inversion method according to claim 4, characterized in that, Will After weighted random sampling, the data is input into the physical sensing spectral gating module, which then obtains the spectral vector through global average pooling. : ； Will Input two fully connected layers and calculate the unnormalized score of the fully connected layer output. : ； In the formula, This is the weight matrix for the second fully connected layer. This is the weight matrix of the first fully connected layer. It is the ReLU activation function; Using the addition operation layer Methane injection absorption prior bias : ; In the formula, To inject unnormalized scores with prior bias; The weights are obtained through the Sigmoid activation function. ; ； The final output processed by the physical sensing spectral gating module is calculated using an element-wise multiplier. : 。 6. The adaptive physical sensing hyperspectral inversion method for methane plumes according to claim 5, characterized in that, Will The ResNet-18 neural network, pooling layer, fully connected layer and binary classifier are input sequentially, and the output image predicts whether the plume is present. Neural network training uses weighted cross-entropy as a constraint.

7. The adaptive physical sensing methane plume hyperspectral inversion method according to claim 6, characterized in that, S4 includes, S4.1, and... For each sample, randomly generate A candidate cropping window, using Calculate the sum of the true values ​​of the plume concentration within each clipping window. For methane concentration data in L2B, select the cropping window with the largest total concentration as the image patch for training. S4 includes, S4.2, performing a logarithmic transformation on the truth labels: ； In the formula, These are the truth labels for the methane plume concentration increment inversion model structure.

8. The adaptive physical sensing methane plume hyperspectral inversion method according to claim 7, characterized in that, S4 includes S4.3, which constructs a methane plume concentration increment inversion model, including stem convolution, encoder, decoder, skip connections and residual and group normalization modules. The residual and group normalization modules are stacked convolutional blocks, each of which introduces residual shortcuts and group normalization.

9. The adaptive physical sensing hyperspectral inversion method for methane plumes according to claim 8, characterized in that, S4 includes S4.4, setting dynamic weights based on pixel density, with the weight calculation formula as follows: ； In the formula, These are the concentration values ​​after quantile normalization. and For hyperparameters; S4 includes S4.5, the loss function. Weighted Smooth L1 loss is used: ； In the formula, The set of all pixel locations involved in the loss calculation. is the x-coordinate of the pixel. y is the ordinate of the pixel. For in pixels place , Pixels predicted by the methane plume concentration increment inversion model The methane concentration value at that location, For real pixels The methane concentration value at that location, This describes the processing procedure for the SmoothL1 loss function.

10. The adaptive physical sensing hyperspectral inversion method for methane plumes according to claim 9, characterized in that, S5 includes employing a semi-overlapping sliding window strategy to traverse the entire image for prediction with a set step size, performing pixel-by-pixel accumulation and averaging fusion on the overlapping areas, and generating a methane concentration distribution map of the entire image.

Images