Gas remote sensing image feature extraction method based on diffusion model

Abstract

The invention discloses a gas remote sensing image feature extraction method based on a diffusion model, and relates to the technical field of remote sensing image processing, and the method specifically comprises the steps: obtaining a gas remote sensing image data set; preprocessing the gas remote sensing image data set; constructing a basic CNN network model of the fusion diffusion model; inputting the preprocessed gas remote sensing image into a CNN network model, and obtaining a shallow feature map through a CNN shallow network; inputting the shallow feature map into a pre-trained diffusion module, and carrying out denoising and enhancement processing; and inputting the enhanced feature map into a CNN deep network to obtain a final gas remote sensing image feature. According to the method, the diffusion model is used as an independent module to be embedded between the shallow layer and the deep layer of the CNN network, the diffusion model is used for denoising the feature map extracted by the CNN network, details are enhanced, the problems of noise interference and difficulty in capturing weak signals in the remote sensing image are solved, related feature data and analysis conclusions of the gas remote sensing image are output, and powerful support is provided for subsequent analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, in particular to a method for extracting gas remote sensing image features based on a diffusion model. Background Art

[0002] Gas remote sensing technology has been widely used in fields such as environmental monitoring and industrial emission detection. During the processing of gas remote sensing images, accurate image feature extraction is crucial for subsequent tasks such as gas concentration inversion and pollution source identification. Traditional gas remote sensing image feature extraction methods based on convolutional neural networks (CNNs) can automatically learn image features. However, due to the low signal-to-noise ratio, complex background interference, and fuzzy gas target features of gas remote sensing images, CNN networks struggle to extract sufficiently accurate and robust features. Furthermore, traditional CNN networks are unable to fully tap into the potential information in the image during feature extraction, resulting in insufficient representation of gas targets using the extracted features, which impacts the accuracy and reliability of subsequent gas remote sensing image analysis.

[0003] Based on this, a gas remote sensing image feature extraction method based on a diffusion model is now provided, which can eliminate the drawbacks of existing technical solutions. Summary of the Invention

[0004] The purpose of the present invention is to provide a gas remote sensing image feature extraction method based on a diffusion model to solve the problem that the traditional method for extracting gas remote sensing image features based on CNN in the background technology is insufficient.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The gas remote sensing image feature extraction method based on the diffusion model is applied to the gas remote sensing image feature extraction system based on the diffusion model. The specific steps include:

[0007] S1. Obtain gas remote sensing image dataset;

[0008] S2. Preprocessing the gas remote sensing image dataset;

[0009] S3, build the basic CNN network model of the fusion diffusion model;

[0010] S4, inputting the preprocessed gas remote sensing image into the CNN network model, performing feature extraction through the shallow network of CNN, and obtaining a shallow feature map;

[0011] S5. Input the shallow feature map into the pre-trained diffusion module, and perform denoising and enhancement processing on the shallow feature map using the inverse process of the diffusion model;

[0012] S6. Input the enhanced feature map into the deep network of CNN for classification and regression operations to obtain the final gas remote sensing image features.

[0013] Preferably, the gas remote sensing image dataset includes remote sensing images of different times, different geographical information, and different gas types, and the gas remote sensing image dataset is grouped according to time, geographical information, and gas type.

[0014] Preferably, the step S2 specifically includes: performing noise reduction processing on the gas remote sensing image, and performing normalization processing on the noise reduced gas remote sensing image.

[0015] Preferably, step S3 specifically includes: constructing a basic CNN network, the CNN network includes several convolutional layers, pooling layers, fully connected layers and activation function layers, fusing the diffusion model with the CNN network, the diffusion model is used to enhance the feature map output by the basic CNN network, the shallow network of the CNN network includes multiple convolutional layers and pooling layers, which are used to extract the basic features and local features of the gas remote sensing image, the deep network of the CNN network includes several convolutional layers, global average pooling layers, fully connected layers and activation function layers, which are used to classify and regress the feature map processed by the diffusion module, and output the gas information of the gas remote sensing image, the diffusion model is embedded in the CNN network structure as an independent diffusion module, and is connected to the convolutional layer and pooling layer of the CNN network, and information fusion is achieved through feature map splicing.

[0016] Preferably, the step S4 specifically includes:

[0017] S41, inputting the preprocessed gas remote sensing image into the shallow network of the CNN network model, wherein the shallow network includes at least 3 convolution layers and 2 pooling layers, each convolution layer uses a 3×3 convolution kernel, and extracts edge and texture underlying features in the image through convolution operation;

[0018] S42, connect the ReLU activation function layer after each convolution layer;

[0019] S43. Use the pooling layer to downsample the convolutional feature map, reduce the size and computational complexity of the feature map, retain the main features, and generate a shallow feature map containing multi-scale information.

[0020] Preferably, the step S5 specifically includes:

[0021] S51. During the training phase, Gaussian noise is gradually added to the shallow feature map according to a preset noise scale sequence to construct a noisy feature map sequence;

[0022] S52, using the U-Net structure to build a diffusion model, input the noisy feature map and the corresponding time step, and add noise through the diffusion model prediction;

[0023] S53. After adding Gaussian noise of the maximum noise scale to the shallow feature map, the map is input into the pre-trained diffusion module, and the feature map is updated through the reverse iterative denoising process to obtain the denoised and enhanced feature map.

[0024] Preferably, the pre-training process of the diffusion model includes: after constructing a CNN network model based on the diffusion model enhancement, using the pre-processed gas remote sensing image data set to train the diffusion model, learning the image denoising and enhancement mode, and the loss function used in the training process is the mean square error loss function.

[0025] Preferably, the step S6 specifically includes:

[0026] S61, inputting the enhanced feature map into the deep network of the CNN network model, wherein the deep network comprises at least 3 convolutional layers, 1 global average pooling layer and 2 fully connected layers;

[0027] S62, further extract high-level semantic features of gas through the convolutional layer of the deep network;

[0028] S63, using a global average pooling layer to reduce the dimension of the convolution feature map to obtain a global feature vector;

[0029] S64. Input the global feature vector into the fully connected layer, output the probability distribution of gas types through the classification head, and output the gas concentration value and concentration distribution heat map through the regression head to obtain the final gas remote sensing image features.

[0030] Preferably, the gas remote sensing image feature extraction system based on the diffusion model includes:

[0031] Database, used to store multi-source gas remote sensing image data, and supports data addition, deletion, modification, query and version management operations;

[0032] A data acquisition module, which is used to acquire gas remote sensing images in real time and in batches in accordance with image formats and simultaneously acquire auxiliary data, including but not limited to meteorological parameters and geographic information;

[0033] The preprocessing module is used to perform noise reduction and normalization on the collected data;

[0034] The feature extraction and enhancement module is used to extract the underlying features through the shallow CNN network, perform inverse denoising operations on the shallow feature maps using the pre-trained diffusion module, and merge feature maps at different levels;

[0035] Deep semantic analysis module, used to integrate enhanced feature maps through CNN deep network to perform classification and regression tasks;

[0036] The result output module is used to display gas remote sensing image data, related feature data and analysis conclusions in a visual way.

[0037] Preferably, the system also includes a memory and a processor, the memory is used to store a program for implementing the gas remote sensing image feature extraction method based on the diffusion model, and the processor is used to execute the program for implementing the gas remote sensing image feature extraction method based on the diffusion model to implement the steps of the gas remote sensing image feature extraction method based on the diffusion model.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention provides a gas remote sensing image feature extraction method based on a diffusion model. The diffusion model is embedded as an independent module between the shallow and deep layers of a CNN network. The diffusion model is used to denoise and enhance the details of the feature maps extracted by the CNN, solving the problems of noise interference and difficulty in capturing weak signals in remote sensing images. The method outputs relevant feature data and analysis conclusions of the gas remote sensing images, providing strong support for subsequent analysis. Through the progressive optimization of the diffusion model, the CNN network learns more robust feature representations and adapts to complex imaging conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the step flow structure of the present invention.

[0041] Figure 2 Schematic diagram of the system structure of the present invention.

[0042] Notes on the accompanying drawings: database 10, data acquisition module 20, preprocessing module 30, feature extraction and enhancement module 40, deep semantic analysis module 50, result output module 60. DETAILED DESCRIPTION

[0043] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0044] In this embodiment, if Figure 1-Figure 2 As shown in FIG, the gas remote sensing image feature extraction method based on the diffusion model is applied to the gas remote sensing image feature extraction system based on the diffusion model. The specific steps include:

[0045] S1. Obtain gas remote sensing image dataset;

[0046] Multispectral / hyperspectral images of gas distribution are collected using satellites, drones, or ground-based remote sensing equipment in conjunction with sensors and cameras, covering characteristic bands of gases such as CO2, CH4, and O3. The data must include samples under different weather conditions, at different times, and in different regions to improve the generalization ability of the model. The images are presented in standard remote sensing formats such as TIFF and HDR, which include spatial and spectral dimensions. Labels such as the true value of gas concentration and the boundaries of polluted areas can be annotated as needed. The gas remote sensing image dataset is divided into training, validation, and test sets in a ratio of 7:2:1 to ensure balanced sample distribution.

[0047] S2. Preprocessing the gas remote sensing image dataset: performing noise reduction processing on the gas remote sensing image and normalizing the noise-reduced gas remote sensing image;

[0048] Convert pixel values ​​to actual radiance values, eliminate sensor noise, remove the effects of atmospheric scattering and absorption, and restore the true surface reflectance. Perform denoising on the above image, remove salt and pepper noise or Gaussian noise through median filtering and wavelet denoising, and then perform normalization to scale pixel values ​​to the range of [-1, 1] or [0, 1] to accelerate model convergence.

[0049] S3, build the basic CNN network model of the fusion diffusion model;

[0050] The ResNet classic architecture is used as the backbone network. The shallow network is used to extract low-level features such as edges and textures, and the deep network is used to extract high-level semantic features such as gas distribution patterns and concentration gradients. The diffusion model is embedded as an independent module between the shallow and deep layers of the CNN network, forming a network structure of CNN shallow layer → diffusion module → CNN deep layer. During the forward diffusion process of the diffusion module, the shallow feature map F is forwarded to the deep layer. shallow Add noise gradually to construct the noise scale sequence {σ t}(t=1,2,...,T), generate a noisy feature sequence {F t}(t=1,2,...,T), noise scale σ t Incrementing in logarithmic space, the noisy feature map and time step t are input into the inverse denoising process, the noise is predicted and denoised iteratively to enhance the robustness of the feature;

[0051] S4, inputting the preprocessed gas remote sensing image into the CNN network model, performing feature extraction through the shallow network of CNN, and obtaining a shallow feature map;

[0052] The preprocessed image is input into the CNN shallow network, and the most basic local features in the image are captured through the convolution layer (3×3 convolution kernel) and the pooling layer. The output result is a multi-scale shallow feature map F shallow ∈R H×W×C, where H×W is the spatial size, C is the number of channels, and the shallow features include the edges of gas clouds, spectral textures and other underlying information;

[0053] S5. Input the shallow feature map into the pre-trained diffusion module and use the inverse process of the diffusion model to denoise and enhance the shallow feature map;

[0054] In the training phase, the shallow parameters of CNN are fixed first, the diffusion module is trained separately, and the shallow feature map F is input. shallow Noisy feature F after superposition with random noise noisy =F shallow + tadpole·σ t (~N(0,1)), output prediction noise The diffusion model parameters are optimized by the mean square error loss function, and the optimization target is L = E t,ò ||Tadpole-|| 2 , the shallow feature map F shallow Input pre-trained diffusion module, from the maximum noise scale σ t At the beginning, iterate 10 to 50 times to obtain the enhanced feature map F denoised , thereby removing artifacts and irrelevant noise in the feature map, retaining the key structure of gas distribution, enhancing weak signal features, and improving subsequent classification / regression accuracy;

[0055] S6. Input the enhanced feature map into the deep network of CNN to perform classification and regression operations to obtain the final gas remote sensing image features;

[0056] F denoised The deep CNN network is input to extract high-level semantic features, such as gas type and concentration level. Feature fusion is performed through the activation function layer and the fully connected layer. The classification head outputs the probability distribution of gas types, such as distinguishing CO2 leaks from background noise. The regression head outputs the predicted gas concentration value or generates a concentration distribution heat map, thereby enhancing the feature extraction operation in the gas remote sensing image.

[0057] Specifically, the steps for constructing the architecture of the diffusion model embedded in the CNN network are as follows:

[0058] Construct a basic CNN network. The CNN network includes several convolutional layers, pooling layers, fully connected layers and activation function layers. The diffusion model is integrated with the CNN network. The number of input and output channels of the diffusion module is consistent with the number of channels of the shallow feature map of CNN. The diffusion model is used to enhance the feature map output by the basic CNN network. The shallow network of the CNN network contains multiple convolutional layers and pooling layers, which are used to extract the basic features and local features of the gas remote sensing image. The deep network of the CNN network contains several convolutional layers, global average pooling layers, fully connected layers and activation function layers, which are used to classify and regress the feature maps processed by the diffusion module and output the gas information of the gas remote sensing image. The diffusion model is embedded in the CNN network structure as an independent diffusion module and connected to the convolutional layer and pooling layer of the CNN network. Information fusion is achieved through feature map splicing. The shallow layer of CNN is used as a shared encoder for the diffusion model and CNN, and feature extraction and input preparation of the diffusion model are completed at the same time. For example, the output feature map F of the first three layers of CNN enc , can be directly input into the deep layer of CNN for traditional feature extraction, or input into the diffusion model for feature denoising, and output enhanced feature F denoised ;

[0059] like Figure 1 As shown in the figure, the gas remote sensing image dataset includes remote sensing images of different time, different geographic information, and different gas types, and the gas remote sensing image dataset is grouped according to time, geographic information, and gas type to facilitate subsequent analysis;

[0060] like Figure 1 As shown, step S4 specifically includes:

[0061] S41. Input the preprocessed gas remote sensing image into the shallow network of the CNN network model. The shallow network includes at least 3 convolution layers and 2 pooling layers. Each convolution layer uses a 3×3 convolution kernel to extract edge and texture underlying features in the image through convolution operations.

[0062] S42, connect the ReLU activation function layer after each convolution layer;

[0063] S43. Use the pooling layer to downsample the convolution feature map, reduce the size and computational complexity of the feature map, retain the main features, and generate a shallow feature map containing multi-scale information;

[0064] Specifically, multi-scale information refers to features extracted by different convolution kernel sizes. The shallow network is close to the input layer and is responsible for extracting low-level, localized basic features such as edges, textures, colors, etc. The deep network is close to the output layer and is responsible for integrating shallow features through multi-layer nonlinear transformations to form high-level, abstract semantic features such as object shape, structure, category, etc. The shallow network includes the first 1 to 3 convolution layers, the first 1 to 2 pooling layers, and the activation function layer after the shallow convolution. The deep network includes subsequent convolution layers, global average pooling layers, fully connected layers, and deep convolution or activation function layers after full connection. The 3-layer convolution group of the shallow network uses a 3×3 convolution kernel and 64-128 channels in each layer to extract underlying features such as edges, texture directions, and spectral gradients in turn. The convolution operation expression is F conv =σ(W*F in +b), where W is the convolution kernel weight, F in is the input feature map, σ is the ReLU activation function, the pooling kernel size of the first 1 to 2 pooling layers is 2×2, the step size is 2, the convolution feature map is downsampled, the spatial dimension is reduced to 1 / 4 of the input size, the spatial invariance of the local features is retained, the feature maps output by each convolution layer are fused through channel splicing, and the shallow feature map is F shallow ∈R H / 4×W / 4×C , where C is the total number of channels. The diffusion model performs denoising optimization on gas remote sensing images to avoid detail loss caused by traditional pooling operations;

[0065] like Figure 1 As shown, step S5 specifically includes:

[0066] S51. During the training phase, Gaussian noise is gradually added to the shallow feature map according to a preset noise scale sequence to construct a noisy feature map sequence;

[0067] S52, using the U-Net structure to build a diffusion model, input the noisy feature map and the corresponding time step, and add noise through the diffusion model prediction;

[0068] S53, after adding Gaussian noise of the maximum noise scale to the shallow feature map, input it into the pre-trained diffusion module, and update the feature map through the reverse iterative denoising process to obtain the denoised and enhanced feature map;

[0069] Specifically, the preset noise scale sequence adopts the logarithmic space cosine annealing strategy, that is, where σ min =10 -4 ,σ max =10 2, T = 100 is the total number of time steps. The diffusion module adopts the U-Net structure, which consists of an encoder, a decoder and an intermediate layer. The encoder includes 3 layers of downsampling blocks, and the number of channels increases from 64 to 256. The decoder includes 3 layers of upsampling blocks, and the number of channels decreases from 256 to 64. The intermediate layer includes a bottleneck convolution block. The denoising stage adopts the DDIM fast sampling algorithm, starting from σ t Start reverse iteration S = 20 times, the iteration formula is in It is Gaussian noise or zero vector, which can effectively suppress complex interferences such as atmospheric scattering noise and sensor inherent noise commonly found in gas remote sensing images;

[0070] like Figure 1-Figure 2 As shown in the figure, the pre-training process of the diffusion model includes: after building a CNN network model based on the diffusion model enhancement, the diffusion model is trained using the pre-processed gas remote sensing image dataset to learn the image denoising and enhancement mode. The loss function used in the training process is the mean square error loss function;

[0071] Specifically, the diffusion module uses the Adam optimizer with a learning rate of 0.001 and a training set accounting for 70%;

[0072] like Figure 1 As shown, step S6 specifically includes:

[0073] S61, inputting the enhanced feature map into the deep network of the CNN network model, where the deep network includes at least 3 convolutional layers, 1 global average pooling layer, and 2 fully connected layers;

[0074] S62, further extract high-level semantic features of gas through the convolutional layer of the deep network;

[0075] S63, using a global average pooling layer to reduce the dimension of the convolution feature map to obtain a global feature vector;

[0076] S64. Input the global feature vector into the fully connected layer, output the probability distribution of gas types through the classification head, and output the gas concentration value and concentration distribution heat map through the regression head to obtain the final gas remote sensing image features;

[0077] Specifically, the deep network contains three layers of residual convolution blocks, and the number of channels increases from 128 to 512. It fuses cross-layer features through skip connections to extract high-level semantics such as gas distribution patterns and concentration gradients. It uses global average pooling to compress the spatial dimension to 1×1 and obtain the global feature vector v∈R 512 , the calculation formula is The multi-task output head includes a classification head and a regression head, which is suitable for complex scenarios and multi-source data;

[0078] like Figure 2As shown in FIG, the gas remote sensing image feature extraction system based on the diffusion model includes:

[0079] Database 10 is used to store multi-source gas remote sensing image data and supports data addition, deletion, modification, and version management operations. It can be used to store annotated data, manage metadata, and provide data support for the system.

[0080] The data acquisition module 20 is used to acquire gas remote sensing images in real time and batches in accordance with the image format and simultaneously collect auxiliary data, including but not limited to meteorological parameters and geographic information. The auxiliary data can parse the image format and extract spectral and spatial dimension data for subsequent feature fusion;

[0081] The pre-processing module 30 is used to perform noise reduction and normalization on the collected data, and can perform atmospheric correction, cropping, data enhancement and other operations to expand the diversity of training samples;

[0082] The feature extraction and enhancement module 40 is used to extract low-level features through the shallow CNN network, perform inverse denoising on the shallow feature map using the pre-trained diffusion module, merge feature maps at different levels, and extract low-level features such as edges and spectral textures through the shallow CNN network, preserving details and structural information;

[0083] The deep semantic analysis module 50 is used to integrate the enhanced feature maps through the CNN deep network, perform classification and regression tasks, and extract semantic features such as gas distribution patterns and concentration gradients;

[0084] The result output module 60 is used to visually display gas remote sensing image data, related feature data and analysis conclusions. It can derive gas concentration values ​​and classification confidence levels, and visually display pseudo-color heat maps, concentration curve comparison maps, spatial distribution vector maps, time series dynamic maps, etc. The interactive interface supports Web / GUI interface query, result download and docking with GIS system;

[0085] Specifically, the multi-source gas remote sensing image data in the database 10 is transmitted to the data acquisition module 20, which loads the original image. The pre-processing module 30 cleans the data and stores the processed data in a temporary storage area for subsequent modules to call. The diffusion model parameters (from the diffusion module, not listed in the figure) are called by the feature extraction and enhancement module 40, and the shallow features are denoised to generate an optimized shallow feature map. The enhanced feature map is integrated by the deep semantic analysis module 50, and the analysis results are output. The result output module 60 exports the results in the form of visualization. The classification or regression results can be fed back to the database 10 for historical data archiving and model iterative training.

[0086] like Figure 2As shown, the system also includes a memory and a processor, wherein the memory is used to store a program for implementing the gas remote sensing image feature extraction method based on the diffusion model, and the processor is used to execute the program for implementing the gas remote sensing image feature extraction method based on the diffusion model to implement the steps of the gas remote sensing image feature extraction method based on the diffusion model, and push the image data to the visual display interface through the processor to support historical data query and realize subsequent analysis operations.

[0087] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A gas remote sensing image feature extraction method based on a diffusion model, characterized in that: Applied to the gas remote sensing image feature extraction system based on the diffusion model, the specific steps include: S1. Obtain gas remote sensing image dataset; S2. Preprocessing the gas remote sensing image dataset; S3, build the basic CNN network model of the fusion diffusion model; S4, inputting the preprocessed gas remote sensing image into the CNN network model, performing feature extraction through the shallow network of CNN, and obtaining a shallow feature map; S5. Input the shallow feature map into the pre-trained diffusion module, and perform denoising and enhancement processing on the shallow feature map using the inverse process of the diffusion model; S6. Input the enhanced feature map into the deep network of CNN for classification and regression operations to obtain the final gas remote sensing image features.

2. The method for extracting gas remote sensing image features based on a diffusion model according to claim 1, wherein: The gas remote sensing image dataset includes remote sensing images of different times, different geographical information, and different gas types, and the gas remote sensing image dataset is grouped according to time, geographical information, and gas type.

3. The method for extracting gas remote sensing image features based on a diffusion model according to claim 2, wherein: The step S2 specifically includes: performing noise reduction processing on the gas remote sensing image, and performing normalization processing on the noise reduced gas remote sensing image.

4. The method for extracting gas remote sensing image features based on a diffusion model according to claim 3, wherein: The step S3 specifically includes: constructing a basic CNN network, the CNN network includes several convolutional layers, pooling layers, fully connected layers and activation function layers, fusing the diffusion model with the CNN network, the diffusion model is used to enhance the feature map output by the basic CNN network, the shallow network of the CNN network includes multiple convolutional layers and pooling layers, which are used to extract the basic features and local features of the gas remote sensing image, the deep network of the CNN network includes several convolutional layers, global average pooling layers, fully connected layers and activation function layers, which are used to classify and regress the feature map processed by the diffusion module, and output the gas information of the gas remote sensing image, the diffusion model is embedded in the CNN network structure as an independent diffusion module, and is connected to the convolutional layer and pooling layer of the CNN network, and information fusion is achieved through feature map splicing.

5. The method for extracting gas remote sensing image features based on a diffusion model according to claim 4, characterized in that: The step S4 specifically includes: S41, inputting the preprocessed gas remote sensing image into the shallow network of the CNN network model, wherein the shallow network includes at least 3 convolution layers and 2 pooling layers, each convolution layer uses a 3×3 convolution kernel, and extracts edge and texture underlying features in the image through convolution operation; S42, connect the ReLU activation function layer after each convolution layer; S43. Use the pooling layer to downsample the convolutional feature map, reduce the size and computational complexity of the feature map, retain the main features, and generate a shallow feature map containing multi-scale information.

6. The method for extracting gas remote sensing image features based on a diffusion model according to claim 5, characterized in that: The step S5 specifically includes: S51. During the training phase, Gaussian noise is gradually added to the shallow feature map according to a preset noise scale sequence to construct a noisy feature map sequence; S52, using the U-Net structure to build a diffusion model, input the noisy feature map and the corresponding time step, and add noise through the diffusion model prediction; S53. After adding Gaussian noise of the maximum noise scale to the shallow feature map, the map is input into the pre-trained diffusion module, and the feature map is updated through the reverse iterative denoising process to obtain the denoised and enhanced feature map.

7. The method for extracting gas remote sensing image features based on a diffusion model according to claim 6, wherein: The pre-training process of the diffusion model includes: after constructing a CNN network model based on the diffusion model enhancement, using the pre-processed gas remote sensing image dataset to train the diffusion model, learning image denoising and enhancement modes, and the loss function used in the training process is the mean square error loss function.

8. The method for extracting gas remote sensing image features based on a diffusion model according to claim 7, wherein: The step S6 specifically includes: S61, inputting the enhanced feature map into the deep network of the CNN network model, wherein the deep network comprises at least 3 convolutional layers, 1 global average pooling layer and 2 fully connected layers; S62, further extract high-level semantic features of gas through the convolutional layer of the deep network; S63, using a global average pooling layer to reduce the dimension of the convolution feature map to obtain a global feature vector; S64. Input the global feature vector into the fully connected layer, output the probability distribution of gas types through the classification head, and output the gas concentration value and concentration distribution heat map through the regression head to obtain the final gas remote sensing image features.

9. The method for extracting gas remote sensing image features based on a diffusion model according to claim 1, wherein: The gas remote sensing image feature extraction system based on the diffusion model includes: A database (10) is used to store multi-source gas remote sensing image data and support data addition, deletion, modification, query and version management operations; A data acquisition module (20) is used to acquire gas remote sensing images in batches in real time that conform to an image format and simultaneously acquire auxiliary data, the auxiliary data including but not limited to meteorological parameters and geographic information; A pre-processing module (30) is used to perform noise reduction and normalization processing on the collected data; A feature extraction and enhancement module (40) is used to extract low-level features through a shallow CNN network, perform an inverse denoising operation on the shallow feature map using a pre-trained diffusion module, and merge feature maps at different levels; A deep semantic analysis module (50) for integrating the enhanced feature maps through a CNN deep network to perform classification and regression tasks; The result output module (60) is used to display the gas remote sensing image data, related feature data and analysis conclusions in a visual manner.

10. The method for extracting gas remote sensing image features based on a diffusion model according to claim 9, characterized in that: The system also includes a memory and a processor, the memory is used to store a program for implementing the gas remote sensing image feature extraction method based on the diffusion model, and the processor is used to execute the program for implementing the gas remote sensing image feature extraction method based on the diffusion model to implement the steps of the gas remote sensing image feature extraction method based on the diffusion model described in any one of claims 1-8.

Images