Carbon monoxide plume identification and quantification method based on domestic hyperspectral satellite

Abstract

The invention provides a carbon monoxide plume identification and quantification method based on a domestic hyperspectral satellite, and the method comprises the steps: carrying out the radiometric calibration and atmospheric correction of the data of a visible short-wave infrared hyperspectral camera AHS I carried by the domestic hyperspectral observation satellite (GF-5B, Gaofen-5 B satellite); generating a carbon monoxide unit absorption spectrum by using a medium-resolution atmospheric transmission model MODTRAN; inverting a carbon monoxide enhancement value by adopting a matched filtering method; generating a binary plume mask through threshold segmentation; and constructing a deep learning model based on a residual neural network ResNet network, and realizing plume pixel-level identification and carbon monoxide column concentration enhancement value prediction. The method can identify the carbon monoxide plume with high precision and high efficiency, and is suitable for atmospheric pollution monitoring and pollution source quantification.

Description

Technical Field

[0001] This application relates to the field of atmospheric environment remote sensing image processing and quantitative analysis technology, specifically to a method for identifying and quantifying carbon monoxide plumes based on domestically produced hyperspectral satellites. Background Technology

[0002] Carbon monoxide (CO), as a significant air pollutant and indirect greenhouse gas, has a substantial impact on human health and climate change. Its emissions mainly originate from human activities such as the incomplete combustion of fossil fuels and the burning of biomass. While traditional ground-based monitoring methods can provide precise single-point data, their coverage is limited by the sparse distribution of monitoring stations, making it difficult to achieve a continuous and comprehensive characterization of emissions over large areas. Against this backdrop, satellite remote sensing technology, with its advantages of global coverage and periodic observation, has made it possible to track and quantify CO emissions.

[0003] However, existing satellite CO inversion methods still face bottlenecks. Most methods rely on low spatial resolution sensor data, which, while suitable for large-scale background concentration analysis, struggles to effectively identify the fine plume structures generated by emission point sources such as industrial areas and power plants. Regarding identification algorithms, traditional threshold segmentation methods, while simple and easy to implement, are susceptible to interference from changes in land surface type, clouds, and aerosols, leading to frequent false positives and false negatives and insufficient reliability. Furthermore, while physical inversion methods based on radiative transfer models offer high accuracy, their computational process is complex, relies on numerous auxiliary parameters, and has low computational efficiency, making it difficult to meet the demands of near real-time processing of large-scale data. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this application provides a method for identifying and quantifying carbon monoxide plumes based on domestically produced hyperspectral satellites, which solves the problems of low detection accuracy and high computational complexity of carbon monoxide plumes.

[0005] To achieve the aforementioned objectives, this application proposes a method for identifying and quantifying carbon monoxide plumes based on domestically developed hyperspectral satellites. This method includes the following steps:

[0006] Step 1: Radiometric calibration of GF5B AHSI (Advanced Hyperspectral Imager) data, converting DN values ​​into radiance values; obtaining surface reflectance data using the FLAASH atmospheric correction model to eliminate the effects of atmospheric absorption and scattering; and dividing the remote sensing dataset into training and test sets according to a set ratio after data preprocessing.

[0007] Step 2: Use the MODTRAN (MODerate resolution atmospheric TRANsmission) atmospheric radiative transfer model to generate the CO unit absorption spectrum s to describe the absorption characteristics of CO; multiply the unit absorption spectrum s element-wise with the mean μ of the image pixel spectrum to obtain the target spectrum t.

[0008] Step 3: Apply matched filtering to calculate the similarity between the spectrum of each pixel and the target spectrum to obtain CO enhancement; perform threshold segmentation on the CO enhancement image to generate a binary mask for the CO plume;

[0009] Step 4: Construct a deep learning model for predicting CO column concentration enhancement based on the ResNet (Residual Neural Network) network architecture. Use the bands where CO has significant absorption in shortwave infrared, the RGB band, and the binary mask from Step 3 as input features. Use the CO enhancement calculated in Step 3 as labels for training. Select mean squared error as the loss function. Update the network parameters to obtain the optimal parameter model.

[0010] Step 5: Apply the optimal model to the test set processed in Step 1, and finally output whether each pixel contains a CO plume and its enhancement value.

[0011] Further, step 1 includes:

[0012] Radiometric calibration was performed on the raw GF5B AHSI data, converting the DN values ​​into radiance values. The calculation formula is as follows:

[0013] L = Gain × DN + Offset

[0014] Where L is radiance (unit: W / (m²)). 2 ·sr·μm)), Gain and Offset are scaling coefficients, obtained from the Radiance Gains and Radiance Offsets variables in the satellite's metadata, respectively;

[0015] FLAASH atmospheric correction was performed on the calibrated data, with the following parameters set: sensor height set to 705 km; average ground elevation of the corresponding area in the image set to 0.194 km; pixel size set to 30 m; atmospheric model for each image scene determined based on the image center latitude and acquisition month; water vapor inversion was performed, with the WaterAbsorption Feature option kept at the default 1135 nm; the aerosol model was determined based on the image, and the aerosol inversion method used the dark pixel reflectance model to estimate the aerosol content and average visibility of the image, selecting 2-Band(KT); the initial visibility was set based on the atmospheric conditions at the time of image acquisition; if aerosols could not be inverted, this value would be used as the initial value for atmospheric correction, and this default value was kept in this method; spectral smoothing was performed, with the spectral smoothing window size kept at the default value of 9; the channel parameter source for the hyperspectral parameter settings was kept at the default setting Automatic Selection, which FLAASH would automatically allocate based on the spectral characteristics of the data; the application was selected, and FLAASH atmospheric correction was executed.

[0016] The GF5B AHSI remote sensing dataset, processed according to the above steps, is randomly divided proportionally to obtain the training set and the test set.

[0017] Further, step 2 includes:

[0018] The MODTRAN atmospheric radiative transfer model was used to generate CO unit absorption spectra to describe the absorption characteristics of CO; the target spectra were obtained by scaling the scalar unit absorption spectra with the average radiance at each wavelength, as calculated by the following formula:

[0019] t s (μ)=μ·s

[0020] Among them, t s (μ) represents the target spectrum, μ represents the mean radiance of the image, and s represents the unit absorption spectrum of CO.

[0021] Further, step 3 includes:

[0022] Matched filtering is used to calculate the similarity between the spectrum of each pixel in the training set image and the spectrum of the target image. Let L0 be the radiation observed under the background concentration without CO enhancement. According to the Lambert-Beer absorption law, L(α,s)=L0e -αs This refers to the CO enhancement effect. The Beer-Lambert absorption law in the Gaussian model is linearized using a first-order Taylor series expansion:

[0023] L0e -αs ≈L0-αt s (L0)

[0024] The formula for calculating the CO enhancement value is:

[0025]

[0026] Where α is the CO enhancement value, t s To consider the target spectrum of CO absorption spectrum; L i For observed radiation; μ is the mean background radiation; C -1 is the inverse of the background radiation covariance matrix; t(μ) represents the spectral change caused by a unit CO enhancement value under the background radiation mean μ, i.e., the target spectrum; To obtain the CO enhancement value.

[0027] An adaptive thresholding method was used to determine the threshold for CO plumes in the training set images. First, the histogram of the CO enhancement map was calculated, and the gray values ​​were normalized to the [0,1] interval and divided into 256 gray levels. Then, the optimal segmentation threshold T was determined based on the Otsu algorithm. opt The calculation formula is as follows:

[0028] T opt =argmax T [ω0(T)(μ0(T)-μ T ) 2 +ω1(T)(μ1(T)-μ T ) 2 ]

[0029] Where ω0(T) and ω1(T) are the pixel weights of the foreground and background regions segmented by threshold T, respectively, and μ0(T) and μ1(T) are the gray-scale mean values ​​of the corresponding regions, μ T The global mean is used. T is determined. opt Then, the threshold segmentation formula is obtained:

[0030]

[0031] Where mask(i,j) is the mask value of the pixel in the i-th row and j-th column, and α(i,j) is the CO enhancement value of the pixel in the i-th row and j-th column calculated by the matched filtering method.

[0032] Further, step 4 includes:

[0033] Using the bands where CO has significant absorption in shortwave infrared, the RGB band, and the binary mask obtained in step 3 as input features, the features are first fed into a 1×1 convolutional layer for linear weighting. Batch normalization is applied, and non-linearity is introduced through the ReLU activation function, resulting in 64 output channels. The output of this layer is then fused with the original input using the concat function. The fused features are fed into a 3×3 convolutional layer, subjected to batch normalization, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. The output of this layer is then fused with the original input using the concat function. The fused features are then fed into a 3×3 convolutional layer, subjected to batch normalization, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. The output features are then fed into another 3×3 convolutional layer, subjected to batch normalization, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. Finally, the output features are fed into a 1×1 convolutional layer, mapping the feature map from the previous layer to the CO concentration enhancement value of each pixel. The mean squared error (MSE) is selected as the loss function.

[0034]

[0035] Where n is the sum of all pixels in the current training batch, y i Let i be the true value of the i-th pixel. Let be the predicted value of the i-th pixel. The network is trained using the Adam optimizer, with backpropagation of the loss function to update the network parameters. The optimal parameter model is obtained and saved by minimizing the loss function.

[0036] This application also provides an electronic device, including a processor and a memory, wherein the memory stores program code that, when executed by the processor, causes the processor to perform the steps of the method described.

[0037] This application also provides a storage medium storing a computer program or instructions that, when run on a computer, execute the steps of the method described.

[0038] This application offers at least the following advantages: By combining physical models with deep learning techniques, it significantly improves detection performance and practicality. It utilizes MODTRAN to generate unit absorption spectra, employs matched filtering for rapid initial screening, and generates physically meaningful CO enhancement maps and binary masks. Subsequently, a ResNet deep learning model is constructed to fuse features from the CO absorption band, RGB background band, and plume binary mask, effectively utilizing spectral absorption features and spatial context information to accurately identify weak plume signals in complex backgrounds, significantly reducing false positives and false negatives compared to traditional thresholding methods. The method balances accuracy and efficiency, avoids the high computational cost of pure physical inversion, and the trained model enables rapid end-to-end quantitative concentration inversion, providing a reliable and efficient technical means for atmospheric pollution monitoring applications based on domestically developed high-resolution satellites. Attached Figure Description

[0039] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, 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.

[0040] Figure 1 A flowchart illustrating the steps of a carbon monoxide plume identification and quantitative inversion method based on Gaofen-5B satellite data, provided in one embodiment of this application.

[0041] Figure 2 This is a diagram of the ResNet network architecture provided in the embodiment. Detailed Implementation

[0042] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the solutions in the embodiments of this application are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0043] It should be noted that the embodiments and features in the embodiments of this application can be combined with each other, as long as there is no conflict between the combinations.

[0044] The method for identifying and quantitatively retrieving carbon monoxide plumes based on Gaofen-5B satellite data provided in this application will be further described below with reference to the accompanying drawings and specific embodiments.

[0045] Please see Figure 1In the embodiments of this application, a method for identifying and quantitatively retrieving carbon monoxide plumes based on Gaofen-5B satellite data is disclosed. The method includes the following steps:

[0046] Step 1: Obtain the high-resolution 5B AHSI remote sensing dataset and perform preprocessing operations such as radiometric calibration and atmospheric correction; divide the processed data into training and testing sets according to the set ratio;

[0047] Step 2: Simulate the radiance at the AHSI sensor using the MODTRAN atmospheric transport model and convert it into the unit absorption spectrum of CO; then multiply it with the background radiance to obtain the target spectrum;

[0048] Step 3: Apply the matched filtering algorithm to compare the image pixels in the training set with the target spectrum obtained in Step 2, calculate the similarity between the spectrum of each pixel and the target spectrum, i.e. the CO column concentration enhancement value; then generate a binary feather mask for each image based on the adaptive threshold segmentation method.

[0049] Step 4: Construct a CO column concentration enhancement prediction model based on ResNet architecture by combining spectral absorption features and spatial context information; train the model, optimize network parameters, and save the optimal parameter model;

[0050] Step 5: Apply the best model obtained in Step 4 to the test set.

[0051] The data preprocessing in step 1 specifically includes: reading AHSI metadata, radiometrically calibrating the GF5B AHSI data, converting DN values ​​into radiance values, and calculating them using the following formula:

[0052] L = Gain × DN + Offset

[0053] Where L is radiance (unit: W / (m²)). 2 ·sr·μm)), Gain and Offset are scaling coefficients, obtained from the Radiance Gains and Radiance Offsets variables in the satellite's metadata, respectively;

[0054] FLAASH atmospheric correction was performed on the calibrated data, with the following parameters set: sensor height set to 705 km; average ground elevation of the corresponding area in the image set to 0.194 km; pixel size set to 30 m; atmospheric model for each image scene determined based on the image center latitude and acquisition month; water vapor inversion was performed, with the WaterAbsorption Feature option kept at the default 1135 nm; the aerosol model was determined based on the image, and the aerosol inversion method used the dark pixel reflectance model to estimate the aerosol content and average visibility of the image, selecting 2-Band(KT); the initial visibility was set based on the atmospheric conditions at the time of image acquisition; if aerosols could not be inverted, this value would be used as the initial value for atmospheric correction, and this default value was kept in this method; spectral smoothing was performed, with the spectral smoothing window size kept at the default value of 9; the channel parameter source for the hyperspectral parameter settings was kept at the default setting Automatic Selection, which FLAASH would automatically allocate based on the spectral characteristics of the data; the application was selected, and FLAASH atmospheric correction was executed.

[0055] The GF5B AHSI remote sensing dataset, processed according to the above steps, is randomly divided proportionally to obtain the training set and the test set.

[0056] Step 2, calculating the CO unit absorption spectrum, includes: simulating hyperspectral radiance using the MODTRAN atmospheric radiative transfer model, with specific parameters set as follows: sensor altitude to GF-5B orbital altitude 705km; atmospheric profile selected according to the observation area; rural aerosol scattering model used for aerosol type; and visibility set to a typical value of 23km; reading AHSI metadata and calculating the standard deviation for each band.

[0057]

[0058] Where, σ i Let FWHM be the standard deviation of the Gaussian function for the i-th band. i Let be the full width at half maximum (FWHM) of the i-th band; the value of the Gaussian response function at wavelength λ is:

[0059]

[0060] Among them, c i Let be the center wavelength of the i-th band; then normalize the Gaussian response function of each band so that the sum of the response functions of each band is 1:

[0061]

[0062] in, For the normalized Gaussian response value, g i(λ) represents the Gaussian response value of the i-th band at wavelength λ. The hyperspectral radiance obtained by running the MODTRAN model based on the above parameter settings is convolved with the corresponding band of the AHSI data, and the calculation formula is as follows:

[0063]

[0064] Among them, resampled i Let λ be the low-resolution radiance value after convolution of the i-th band, and rads(λ) be the high-resolution radiance value obtained by running the MODTRAN model; for each band, multiple simulations are performed at different concentrations, assuming the concentration vector is λ. The corresponding convolutional radiance value is The natural logarithm of the convolutional radiance is fitted using a least-squares method:

[0065] ln(rad i )=k·c+b

[0066] The slope k of the regression line is the CO unit absorption of the i-th band. The calculated CO unit absorption of each band constitutes the desired CO unit absorption spectrum s.

[0067] The target spectrum is obtained by scaling the scalar unit absorption spectrum using the average radiance at each wavelength, and the calculation formula is as follows:

[0068] t s (μ)=μ·s

[0069] Among them, t s (μ) represents the target spectrum, μ represents the mean radiance of the image, and s represents the unit absorption spectrum of CO.

[0070] Step 3, which uses matched filtering to calculate the similarity between the spectrum of each pixel in the training set image and the target spectrum, includes: Let L0 be the radiation observed under the environmental background concentration without CO enhancement; according to the Lambert-Beer absorption law, L(α,s)=L0e -αs This refers to the CO enhancement effect. The Beer-Lambert absorption law in the Gaussian model is linearized using a first-order Taylor series expansion:

[0071] L0e -αs ≈L0-αt s (L0)

[0072] The formula for calculating the CO enhancement value is:

[0073]

[0074] Where α is the CO enhancement value, t sTo consider the target spectrum of CO absorption spectrum; L i To observe radiation; The value is the CO enhancement obtained by solving; μ is the mean background radiation; C is the background radiation covariance; t is the target spectrum.

[0075] Step 3, which uses an adaptive threshold segmentation method to determine the threshold for CO plumes in the training set images, includes: calculating the CO enhancement histogram, normalizing the gray values ​​to the [0,1] interval and dividing it into 256 gray levels; and then determining the optimal segmentation threshold T based on the Otsu algorithm. opt The calculation formula is as follows:

[0076] T opt =arg max T [ω0(T)(μ0(T)-μ T ) 2 +ω1(T)(μ1(T)-μ T ) 2 ]

[0077] Where ω0(T) and ω1(T) are the pixel weights of the foreground and background regions segmented by threshold T, respectively, and μ0(T) and μ1(T) are the gray-scale mean values ​​of the corresponding regions, μ T The global mean is used. T is determined. opt Then, the threshold segmentation formula is obtained:

[0078]

[0079] Where mask(i,j) is the mask value of the pixel in the i-th row and j-th column, and α(i,j) is the CO enhancement value of the pixel in the i-th row and j-th column calculated by the matched filtering method. A value of 1 for mask(i,j) indicates that the current pixel contains a CO plume; otherwise, it does not contain a CO plume.

[0080] Please see Figure 2The model constructed in step 4 for predicting CO column concentration enhancement in GF5B AHSI data includes: using the bands where CO has significant absorption in shortwave infrared, the RGB band, and the binary mask obtained in step 3 as input features, first entering a 1×1 convolutional layer for linear weighting; using batch normalization and introducing nonlinearity through the ReLU activation function, resulting in 64 output channels; using the concat function to fuse the output of this layer with the original input features; the fused features then entering a 3×3 convolutional layer; performing batch normalization and introducing nonlinearity through the ReLU activation function... The output has 32 channels. The concat function is used to fuse the output of this layer with the original input. The fused features are then fed into a 3×3 convolutional layer. Batch normalization is performed, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. The output features are then fed into a 3×3 convolutional layer again. Batch normalization is performed, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. Finally, the output features are fed into a 1×1 convolutional layer, which maps the feature map from the previous layer to the CO concentration enhancement value of each pixel.

[0081] Loss function selection is often used to evaluate the mean squared error (MSE) of regression models.

[0082]

[0083] Where n is the sum of all pixels in the current training batch, y i To enhance the true value of CO column concentration at the i-th pixel, This is the model prediction value for CO enhancement at the i-th pixel. The network is trained using the Adam optimizer, with backpropagation of the loss function to update the network parameters. The optimal parameter model is obtained and saved by minimizing the loss function.

[0084] To address the issues of low detection accuracy, high complexity of traditional radiative transfer models, and poor recognition performance of simple threshold segmentation in existing CO plume identification methods, this application introduces an atmospheric radiative transfer model and a deep learning algorithm. This approach considers both CO absorption spectral characteristics and spatial context information, enabling accurate spectral simulation based on Gaofen-5B satellite data. Furthermore, by combining matched filtering and adaptive threshold segmentation, a high-quality plume mask is generated. Finally, a deep neural network based on the ResNet architecture is constructed to fuse spectral and spatial features, ultimately achieving accurate CO plume identification and efficient inversion of concentration enhancement values.

[0085] The foregoing has shown and described the main features and advantages of this application. It will be apparent to those skilled in the art that this application is limited to the details of the exemplary embodiments described above, and that it can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within this application.

[0086] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in the embodiments can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for identifying and quantifying carbon monoxide plumes based on domestically produced hyperspectral satellites, characterized in that, Includes the following steps: Step 1: Radiometrically calibrate the GF5B AHSI data to convert DN values ​​into radiance values; use the FLAASH atmospheric correction model to obtain surface reflectance data to eliminate the effects of atmospheric absorption and scattering; after preprocessing the remote sensing dataset, divide it into training and test sets according to a set ratio. Step 2: Use the MODTRAN atmospheric radiative transfer model to generate CO unit absorption spectra s to describe the absorption characteristics of CO; The target spectrum t is obtained by element-wise multiplying the unit absorption spectrum s with the mean μ of the image pixel spectrum. Step 3: Apply the matched filtering method to calculate the similarity between the spectrum of each pixel and the target spectrum to obtain CO enhancement; Threshold segmentation is performed on the CO enhancement map to generate a binary mask for the CO plume; Step 4: Construct a deep learning model for predicting CO column concentration enhancement based on the ResNet network architecture. Use the bands where CO has significant absorption in shortwave infrared, the RGB band, and the binary mask from Step 3 as input features. Use the CO enhancement calculated in Step 3 as labels for training. Select mean squared error as the loss function. Update the network parameters to obtain the optimal parameter model. Step 5: Apply the optimal model to the test set processed in Step 1, and finally output the CO column concentration enhancement value containing the plume pixels.

2. The method for atmospheric carbon monoxide plume identification and quantitative inversion based on domestically produced hyperspectral satellites according to claim 1, characterized in that, Step 1 includes: Radiometric calibration was performed on the raw GF5B AHSI data, converting the DN values ​​into radiance values. The calculation formula is as follows: L = Gain × DN + Offset where L is radiance (unit: W / (m 2 ·sr·μm)), Gain and Offset are scaling factors, which are obtained from Radiance Gains and Radiance Offsets variables in the metadata provided by the satellite, respectively; FLAASH atmospheric correction was performed on the calibrated data, with the following parameters: sensor height set to 705 km; average ground elevation of the corresponding area in the image set to 0.194 km; pixel size set to 30 m; atmospheric model for each image scene determined based on the image center latitude and acquisition month; water vapor inversion was performed, with the WaterAbsorption Feature option kept at the default 1135 nm; the aerosol model was determined based on the image, and the aerosol inversion method used the dark pixel reflectance model to estimate the aerosol content and average visibility of the image, selecting 2-Band(KT); the initial visibility was set based on the atmospheric conditions at the time of image acquisition; if aerosols could not be inverted, this value would be used as the initial value for atmospheric correction, and this default value was kept in this method; spectral smoothing was performed, with the spectral smoothing window size kept at the default value of 9; the channel parameter source for the hyperspectral parameter settings was kept at the default setting AutomaticSelection, which FLAASH would automatically allocate based on the spectral characteristics of the data; the application was selected, and FLAASH atmospheric correction was executed. The GF5B AHSI remote sensing dataset, processed according to the above steps, is randomly divided proportionally to obtain the training set and the test set.

3. The method for atmospheric carbon monoxide plume identification and quantitative inversion based on domestic hyperspectral satellites according to claim 1, characterized in that, Step 2 includes: The MODTRAN atmospheric radiative transfer model was used to generate CO unit absorption spectra to describe the absorption characteristics of CO; the target spectra were obtained by scaling the scalar unit absorption spectra with the average radiance at each wavelength, as calculated by the following formula: t s (μ) = μ·s Among them, t s (μ) represents the target spectrum, μ represents the mean radiance of the image, and s represents the unit absorption spectrum of CO.

4. The method for atmospheric carbon monoxide plume identification and quantitative inversion based on domestic hyperspectral satellites according to claim 1, characterized in that, Step 3 includes: Matched filtering is used to calculate the similarity between the spectrum of each pixel in the training set image and the spectrum of the target image. Let L0 be the radiation observed under the background concentration without CO enhancement. According to the Lambert-Beer absorption law, L(α,s)=L0e -αs This refers to the CO enhancement effect, which is linearized using a first-order Taylor series expansion of the Beer-Lambert absorption law in the Gaussian model: L0e -αs ≈L0-αt s (L0) The formula for calculating the CO enhancement value is: Where α is the CO enhancement value, t s To consider the target spectrum of CO absorption spectrum; L i To observe radiation; The obtained CO enhancement value is given by the solution; μ is the mean background radiation value; C -1 is the inverse of the background radiation covariance matrix; t(μ) represents the spectral change caused by a unit CO enhancement value under the background radiation mean μ, i.e., the target spectrum; To obtain the CO enhancement value, An adaptive thresholding method was used to determine the threshold for CO plumes in the training set images. First, the histogram of the CO enhancement image was calculated, and the gray values ​​were normalized to the [0,1] interval and divided into 256 gray levels. Then, the optimal segmentation threshold T was determined based on the Otsu algorithm. opt The calculation formula is as follows: T opt =arg max T [ω0(T)(μ0(T)-μ T ) 2 +ω1(T)(μ1(T)-μ T ) 2 ] Where ω0(T) and ω1(T) are the pixel weights of the foreground and background regions segmented by threshold T, respectively, and μ0(T) and μ1(T) are the gray-scale mean values ​​of the corresponding regions, μ T To determine the global mean, T opt Then, the threshold segmentation formula is obtained: Where mask(i,j) is the mask value of the pixel in the i-th row and j-th column, and α(i,j) is the CO enhancement value of the pixel in the i-th row and j-th column calculated by the matched filtering method.

5. The method for atmospheric carbon monoxide plume identification and quantitative inversion based on domestically produced hyperspectral satellites according to claim 1, characterized in that, Step 4 includes: Using the bands where CO has significant absorption in shortwave infrared, the RGB band, and the binary mask obtained in step 3 as input features, the features are first fed into a 1×1 convolutional layer for linear weighting. Batch normalization is applied, and non-linearity is introduced through the ReLU activation function, resulting in 64 output channels. The output of this layer is then fused with the original input using the concat function. The fused features are fed into a 3×3 convolutional layer, subjected to batch normalization, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. The output of this layer is then fused with the original input using the concat function. The fused features are then fed into a 3×3 convolutional layer, subjected to batch normalization, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. The output features are then fed into another 3×3 convolutional layer, subjected to batch normalization, and non-linearity is introduced through the ReLU activation function, resulting in 32 output channels. Finally, the output features are fed into a 1×1 convolutional layer, mapping the feature map from the previous layer to the CO concentration enhancement value of each pixel. The mean squared error (MSE) is selected as the loss function. Where n is the sum of all pixels in the current training batch, y i Let i be the true value of the i-th pixel. For the predicted value of the i-th pixel, the Adam optimizer is used for training. The loss is backpropagated to update the network parameters. The optimal parameter model is obtained and saved with the loss function as the optimization objective.

6. A computer device, characterized in that, It includes a memory, a processor, and program instructions stored in the memory that are executable by the processor, wherein the processor executes the program instructions to implement the steps in the method of any one of claims 1 to 5.

Images