Remote sensing image defogging method

Abstract

The invention discloses a remote sensing image defogging method, which relates to the field of computer vision and image processing, adopts a convolutional network comprising a multi-scale and multi-dimensional module to fit a mapping relation between a foggy image and a clear image, and comprises the following steps: S1, establishing a defogging model; s2, a defogging network of a multi-scale and multi-dimensional module is designed, multi-scale and multi-dimensional features are extracted from each layer of the network, and then the extracted features enter an atmospheric scattering model unit to be processed; s3, collecting a training data set; s4, training a defogging network of a multi-scale and multi-dimensional module; and S5, defogging the remote sensing image. The method has the advantages that the method is provided with a multi-scale and multi-dimensional module, extraction of fine details can be enhanced, fine details between foggy and fogless images can be extracted, haze can be effectively removed, the model is allowed to adaptively and preferentially consider global context or significant details according to needs, and the method is high in robustness and high in robustness. The combination of various pooling methods can ensure better feature fusion and improve the image defogging performance.

Description

Technical Field

[0001] The invention relates to the fields of computer vision and image processing, and in particular to a remote sensing image defogging method. Background Art

[0002] As one of the important tasks in remote sensing image processing, remote sensing image dehazing aims to solve the impact of scattering effects such as haze and smoke in the atmosphere on image quality. In remote sensing images, suspended particles in the atmosphere will cause image contrast to decrease and details to blur, affecting the accurate extraction of surface information. Especially in areas with high humidity and severe pollution, haze interference will greatly reduce the usability of remote sensing images.

[0003] Traditional remote sensing image dehazing methods are usually based on physical models, such as atmospheric scattering models, which estimate the impact of haze effects to restore clear images. However, due to the complexity of atmospheric conditions and the diversity of haze distribution in different scenes, physical model-based methods have limited effectiveness in practical applications. With the development of deep learning technology, data-driven dehazing methods have gradually become a research hotspot. By constructing a large-scale remote sensing image dehazing dataset, the deep learning network can automatically learn the impact of haze on images, thereby achieving efficient dehazing in complex scenes.

[0004] Recent methods have turned to end-to-end learning frameworks, aiming to completely bypass the limitations of physical models. Examples of such methods include multi-scale convolutional neural networks (CNNs), which directly map blurry images to dehazed images. The success of these methods is largely due to their ability to automatically extract powerful discriminative features through multi-layer convolution operations. However, one of the inherent limitations of convolution is that it is a local operation. This means that while convolutional layers are excellent at capturing fine local details, they have difficulty modeling long-range dependencies and global structural information, which are often crucial for accurately restoring blurry images. For the above reasons, the present invention designs a dehazing method for remote sensing images. Summary of the invention

[0005] The purpose of the present invention is to solve the problems in the prior art and to propose a remote sensing image defogging method.

[0006] A remote sensing image defogging method uses a convolutional network including multi-scale and multi-dimensional modules to fit the mapping relationship between foggy images and clear images, including the following steps:

[0007] S1: Building a dehazing model

[0008] S2: Design a dehazing network with multi-scale and multi-dimensional modules. Each layer of the network extracts multi-scale and multi-dimensional features and then enters the atmospheric scattering model unit for processing;

[0009] S3: Collect training dataset;

[0010] S4: training a dehazing network with multi-scale and multi-dimensional modules;

[0011] S5: Remote sensing image dehazing.

[0012] In the above remote sensing image defogging method, in step S1, the defogging model:

[0013] r=F(h)

[0014] Among them, h represents the foggy image, r represents the restored clear image, and the function F represents the mapping relationship between the foggy image and the corresponding clear image. The mapping relationship F(h) between the foggy image and the clear image is fitted through the scale multi-dimensional physical convolutional network. In this structure, the overall framework of the network is based on U-Net, and each layer of upsampling or downsampling is composed of a feature extraction block. In the extraction block structure, the multi-scale block and the multi-dimensional extraction block extract multi-scale and multi-dimensional information in parallel, and then enter the atmospheric scattering model unit for effective defogging. The atmospheric scattering model is described as:

[0015] I(x)=J(x)t(x)+A[1-t(x)]

[0016] Where I(x) is the image affected by haze; J(x) is the haze-free image to be restored; t(x) is the transmittance of light through the atmospheric medium; x represents the image coordinate; A is the global atmospheric light constant.

[0017] In the above remote sensing image defogging method, the multi-scale and multi-dimensional module includes a multi-scale extraction block and a multi-dimensional extraction block, and the multi-scale extraction block includes the following steps:

[0018] S211: Apply multiple parallel dilated convolutions with different dilation rates

[0019] S212: Fusion of multi-scale features using selective kernel module;

[0020] S213: After processing the feature map, the learned features will go through the pixel attention mechanism;

[0021] S214: Introduced a channel attention mechanism that runs in parallel and adaptively highlights fog-related features in the entire image;

[0022] S215: The features extracted by the two attention mechanisms are fused together by summing to create a feature set. In order to refine the fused features, a 3×3 convolution is finally applied

[0023] The multidimensional extraction block includes the following steps:

[0024] S221: Apply deep convolution to preserve feature representation while ensuring computational efficiency;

[0025] S222: normalize to stabilize feature distribution, thereby improving convergence during training;

[0026] S223: Use a combination of 1×1 convolution and 3×3 convolution layers to expand the number of channels and enhance the richness of feature information. 1×1 convolution increases the depth of the feature map, while 3×3 convolution captures local patterns.

[0027] S224: The normalized features extracted from the convolution are fused together to more fully represent the input data. After fusion, another 3×3 convolution is applied to further refine the extracted features;

[0028] S225: After fusion, another 3×3 convolution is applied to further refine the extracted features, and then the refined features are multiplied with another branch processed by the Sigmoid function, effectively rescaling the feature values ​​and enhancing the dimensional representation;

[0029] S226: Use 1×1 convolution to reduce dimensionality to compress useful information and prevent overfitting;

[0030] S227: The reduced feature map is passed to another deep convolutional layer and the residual connection is merged to retain the original feature information, producing the final output, which effectively combines the enhanced and retained features, ready for subsequent processing tasks.

[0031] In the above remote sensing image dehazing method, in step S2, for the number of blocks in each layer, the configuration of [N1, N2, N3, N4, N5] is set to [2, 2, 4, 2, 2], and the corresponding embedding channels are [24, 48, 96, 48, 24], and multiple sub-maps in the dehazing model are learned respectively. The last layer of the network maps the feature map to the original image dimension.

[0032] In the above remote sensing image dehazing method, in step S3, two synthetic remote sensing haze datasets SateHaze1k and HRSD are used;

[0033] The SateHaze1k dataset includes three subsets: light fog, moderate fog, and thick fog. Each subset contains a training set and a test set. Light fog images use haze masks from real clouds, while moderate haze samples combine the features of fog and moderate haze. Thick fog images are generated using the transmission map of dense fog.

[0034] The HRSD dataset consists of two subsets: LHID and DHID. LHID contains training images and test images generated by the atmospheric scattering model to train the network to handle different degrees of haze in remote sensing images. At the same time, DHID includes images created from multiple real haze maps to present the haze more realistically. The combination of these two subsets ensures a comprehensive evaluation of dehazing methods in both synthetic haze scenes and more realistic haze scenes.

[0035] In the above remote sensing image defogging method, in step S4, the learning-based defogging method uses labeled fog samples for training, and the data sets used are SateHaze1k and HRSD. In this step, the network is trained with the collected foggy image samples to learn the mapping relationship between foggy images and clear images. The L1 loss function is used to facilitate defogging, where:

[0036] L1=||J-GT||1

[0037] Among them, J is the actual output result of the network, GT is the true value image, multi-angle random rotation and horizontal flipping are used to enhance the training data set, and the stochastic gradient descent method is selected to optimize the L1 loss function.

[0038] In the above remote sensing image defogging method, in step S5, the original foggy image is re-entered into the atmospheric scattering model unit at the final output of the defogging network to guide the network to restore according to the original image structure.

[0039] In the above remote sensing image dehazing method, in step S212, the selective kernel module adopts hybrid SK fusion, and the hybrid SK fusion introduces global maximum pooling in parallel with the global average pooling to better preserve important details such as edges and textures. In order to balance the contributions of the global average pooling and the global maximum pooling, a hybrid module with a learnable parameter m is introduced, and this parameter is converted into a weight λ using the Sigmoid function.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] 1. The network proposed in the present invention is built on the U-net architecture and has a multi-scale and multi-dimensional module composed of a multi-scale extraction block and a multi-dimensional extraction block, which can enhance the extraction of fine details, help to extract fine details between foggy and fog-free images, and effectively remove haze.

[0042] 3. The present invention allows the model to adaptively prioritize global context or significant details as needed. The combination of multiple pooling methods adopted can ensure better feature fusion and improve the performance of tasks such as image dehazing by effectively capturing global and local features.

[0043] 3. The remote sensing image defogging method designed by the present invention is end-to-end. Once the network model is trained, it is only necessary to input the image to be defogged into the network. Through the forward propagation of the network, a restored clear image can be obtained at the output end, and the operation is convenient. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A diagram showing the steps in a remote sensing image defogging method proposed in the present invention.

[0045] Figure 2 This is a model diagram of the multi-scale and multi-dimensional module in the remote sensing image defogging method proposed in the present invention.

[0046] Figure 3 This is a model diagram of the atmospheric scattering model in the remote sensing image defogging method proposed in the present invention.

[0047] Figure 4 This is a structural diagram of the selective kernel module fusion in the remote sensing image defogging method proposed in the present invention.

[0048] Figure 5 This is a visual comparison diagram of two images in the Haze1k-thin dataset in the remote sensing image dehazing method proposed in the present invention.

[0049] Figure 6 This is a visual comparison chart of two images in the Haze1k-moderate dataset in the remote sensing image dehazing method proposed in the present invention.

[0050] Figure 7 This is a visual comparison diagram of two images in the Haze1k-thick dataset in the remote sensing image dehazing method proposed in the present invention.

[0051] Figure 8 This is a visual comparison diagram of three images in the LHID data in the remote sensing image defogging method proposed in the present invention.

[0052] Fig. 9 This is a visual comparison diagram of three images in the DHID data in the remote sensing image defogging method proposed in this invention. DETAILED DESCRIPTION

[0053] Reference Figure 1-9 A remote sensing image defogging method uses a convolutional network including multi-scale and multi-dimensional modules to fit the mapping relationship between foggy images and clear images. The flowchart of the present invention is as follows: Figure 1As shown, the computer configuration adopts: Intel(R) Xeon(R) CPU E5-2620 v4 processor, Nvidia GeForce GTX 1080Ti graphics processor, main frequency 2.10GHz, memory 12GB, operating system Ubuntu 18.04. The defogging method is implemented based on the Pytorch framework. The present invention is a defogging method based on a convolutional neural network, which specifically includes the following steps:

[0054] Step 1: Build a dehazing model

[0055] Let h represent the foggy image, r represent the restored clear image, and function F represent the mapping relationship between the foggy image and the corresponding clear image. Then the defogging problem is modeled as follows (i.e., the defogging model):

[0056] r=F(h)

[0057] According to the above formula, once the mapping relationship F is obtained, given a foggy image h, a clear image can be obtained through functional relationship mapping, thereby achieving image defogging.

[0058] The present invention uses a convolutional network with a multi-scale and multi-dimensional module to fit the mapping relationship F(h) between the foggy image and the clear image. In this structure, the overall framework of the network is based on U-Net, and each layer of upsampling or downsampling is composed of a feature extraction block. In the extraction block structure, the multi-scale block and the multi-dimensional extraction block extract multi-scale and multi-dimensional information in parallel, and then enter the atmospheric scattering model unit for effective defogging. The atmospheric scattering model is described as:

[0059] I(x)=J(x)t(x)+A[1-t(x)]

[0060] Where I(x) is the image affected by haze; J(x) is the haze-free image to be restored; t(x) is the transmittance of light through the atmospheric medium; x represents the image coordinate; A is the global atmospheric light constant.

[0061] In order to better restore the foggy image, the original foggy image is re-entered into the atmospheric scattering model unit at the final output of the network to guide the network to restore according to the original image structure.

[0062] Step 2: Design a dehazing network with multi-scale and multi-dimensional modules

[0063] According to the defogging model established in step 1, the design is as follows Figure 2The defogging network of the multi-scale multi-dimensional module shown in the figure includes a multi-scale extraction block and a multi-dimensional extraction block. The multi-scale extraction block is inspired by the pixel attention (PA) and channel attention (CA) proposed in FFA-Net. Before the pixel attention block, the present invention uses dilated convolutions with dilation rates of 1, 2, 4, and 8 to extract multi-scale detail features. The multi-dimensional extraction block uses a depth-separable convolution layer to extract features with more dimensions. In addition, the present invention designs a defogging unit (physical model) based on an atmospheric scattering model, extracts atmospheric light (A) through a fully connected layer, extracts transmittance (t) through a convolution layer, and finally obtains a restored clear image (J). This architecture helps to extract fine details between foggy and non-fog images and can effectively remove haze.

[0064] When the method is used, four parallel dilated convolutions with different dilation rates are first applied, which allows the extraction of features at multiple scales and effectively captures fine details and broader contextual information. In order to further enhance the feature representation, the present invention uses an optimized selective kernel module to fuse these multi-scale features, which helps to retain important information at each scale. The selective kernel module adopts hybrid SK fusion, which is an optimization based on SK fusion (SK-Fusion) and enhances image feature selection by adjusting the receptive field. In order to improve the original SK fusion, which only uses global average pooling and may miss key features, the hybrid SK fusion layer introduces global maximum pooling in parallel with global average pooling to better retain important details such as edges and textures. In order to balance the contributions of global average pooling and global maximum pooling, a hybrid module with a learnable parameter m is introduced, and this parameter is converted into a weight λ using a Sigmoid function. This weight is dynamically adjusted during the training process, allowing the model to adaptively prioritize global context or significant details as needed. The combination of these pooling methods ensures better feature fusion and improves the performance of tasks such as image defogging by effectively capturing global and local features. In the specific setting, in the setting of the initial parameter m, when the second layer downsampling and the first layer upsampling are fused, m is set to 0.4, and when the first layer downsampling and the second layer upsampling are fused, m is set to 0.8;

[0065] The combination of the above operations ensures that the local features and multi-scale contextual features of the feature map are well preserved. After processing the feature map in this way, the learned features are passed through the pixel attention mechanism. This step is crucial for extracting key scale-specific details by emphasizing the most relevant pixel-level information. The pixel attention mechanism focuses on identifying and enhancing important areas of the image, which are crucial for tasks such as image defogging, where local clarity is important. In addition to pixel attention, the present invention also introduces a channel attention mechanism that runs in parallel. While pixel attention focuses on specific spatial regions, channel attention focuses on global information and adaptively highlights fog-related features throughout the image. This dual attention method - combining pixel and channel attention - ensures that the network can capture both local fine details and broader scene features. The features extracted by the two attention mechanisms are then fused together by summing, creating a more comprehensive and richer feature set. In order to further refine the fused features, a 3×3 convolution is finally applied. This step helps smooth and refine the features, making them more suitable for downstream tasks such as image restoration or defogging.

[0066] The present invention uses different dilation rates in dilated convolutions. Small dilated convolutions are particularly effective in recovering detailed texture information, while larger dilations provide a wider receptive field, allowing the model to capture long-range dependencies and more extensive structural information in the image. In the multidimensional extraction block (MDB) architecture, the present invention first applies deep convolutions to maintain feature representation while ensuring computational efficiency. Normalization is then performed to stabilize the feature distribution, thereby improving convergence during training. Next, the present invention uses a combination of 1×1 convolution and 3×3 convolution layers to expand the number of channels and enhance the richness of feature information. 1×1 convolution increases the depth of the feature map, while 3×3 convolution captures local patterns. The normalized features extracted from these convolutions are then fused together to more comprehensively represent the input data. After fusion, the present invention applies another 3×3 convolution to further refine the extracted features. The present invention then multiplies the refined features with another branch processed by the Sigmoid function, effectively rescaling the feature values ​​and enhancing the dimensional representation. Then, the present invention uses 1×1 convolution to reduce the dimension to compress useful information and prevent overfitting. Finally, the present invention passes the reduced feature map to another deep convolution (DW Conv) layer and merges the residual connection to retain the original feature information. This produces the final output, which effectively combines the enhanced and retained features, ready for subsequent processing tasks.

[0067] The foggy image is input into the network, and the output is the corresponding clear image. The network in the present invention is an end-to-end structure, wherein each layer of the network extracts multi-scale and multi-dimensional features, and then enters the atmospheric scattering model unit for processing. In the atmospheric scattering model unit, the present invention uses Figure 3The model diagram re-represents the clear image (J): In order to better extract the feature information of atmospheric light (A) and transmittance (t), the present invention introduces two dilated convolution blocks before extracting features. Each dilated convolution block consists of a dilated convolution layer, a batch normalization layer, and an activation function layer. The dilated convolution can avoid information loss while expanding the receptive field, especially when processing details such as image edges and textures, it can better retain these key features. In the dehazing process, it can capture contextual information in a larger range, which helps to better understand the haze distribution and global structure in the image. Considering that atmospheric light is usually assumed to be uniform, the present invention uses global average pooling to remove redundant information in the feature space. However, due to information loss, the present invention cannot apply global average pooling to the transmittance (t) because the transmission map is non-uniform. Therefore, the present invention chooses to use a series of 3x3 convolutional layers to extract features. For the number of blocks in each layer, the present invention sets the configuration of [N1, N2, N3, N4, N5] to [2, 2, 4, 2, 2], and the corresponding embedding channels are [24, 48, 96, 48, 24]. Multiple sub-maps in the dehazing model are learned separately. The last layer of the network maps the feature map to the original image dimension.

[0068] Step 3: Collect training dataset

[0069] The present invention uses two synthetic remote sensing (RS) haze datasets, SateHaze1k and HRSD. SateHaze1k includes three subsets: thin haze, moderate haze, and thick haze, each of which contains a training set of 320 images and a test set of 45 images. The thin haze images utilize haze masks from real clouds, while the moderate haze samples combine the features of fog and moderate haze. The thick haze images are generated using a transmission map of dense fog. The HRSD dataset is divided into two subsets: LHID and DHID. LHID contains a total of 30,517 training images and 500 test images, all of which are generated using an atmospheric scattering model. This subset simulates various haze conditions to train the network to handle different degrees of haze in remote sensing images. On the other hand, DHID consists of a total of 14,990 images, which are created using real haze maps and can present haze more realistically. Among them, 14,490 images are used for training and 500 are used for testing. The combination of these two subsets ensures a comprehensive evaluation of dehazing methods in both synthetic haze scenes and more realistic haze scenes.

[0070] Step 4: Train the dehazing network with multi-scale and multi-dimensional modules

[0071] Learning-based dehazing methods require labeled fog samples for training. Currently, the most commonly used datasets are SateHaze1k and HRSD.

[0072] In this step, the network is trained with the collected foggy image samples to achieve the goal of transforming foggy images into clear images.

[0073] Learning the mapping relationship between images. In the field of dehazing, studies have shown that L1 loss is more conducive to dehazing. The L1 loss function is:

[0074] L1=||J-GT||1

[0075] Among them, J is the actual output result of the network, and GT is the true value image. The present invention uses the PyTorch framework for training on a system equipped with four NVIDIAGeForce GTX 1080Ti GPUs. In order to enhance the training data set, the present invention performs random rotations of 90, 180 and 270 degrees and horizontal flipping. The input of the network includes RGB remote sensing images cropped to a size of 256×256 pixels. The present invention uses the Adam optimizer (β1=0.9, β2=0.999) and uses a batch size of 4 to train each sub-dataset. The initial learning rate is set to 0.0002, and the present invention uses a cosine annealing strategy to gradually reduce it to 0. The present invention selects the stochastic gradient descent method to optimize the loss function, iteratively learns the network with foggy images, updates the network parameters, and ends the training when the loss value of the network tends to be stable. At this time, the saved network parameters are the trained defogging network model.

[0076] Step 5: Remote sensing image dehazing

[0077] The remote sensing image defogging method designed by the present invention is end-to-end. Once the network model is trained, the image to be defogged only needs to be input into the network, and the restored image can be obtained at the output end through the forward propagation of the network.

[0078] Clear image. The present invention uses the test set in the data set to test the effect. In order to show the competitiveness of the dehazing method of the present invention, the present invention compares the network with seven advanced dehazing methods: DCP, AOD-Net, FCTF-Net, GridDehaze-Net, FFA-Net, MixDehaze-Net and OK-Net. In order to ensure a fair comparison, the present invention uses the official implementations of these deep learning models during the training process. These models are provided by their respective authors. For intuitive effects, please refer to the attached figure.

[0079] In the figure below, PSNR stands for Peak Signal-to-Noise Ratio, SSIM stands for Structural Similarity, Hazy Image stands for Hazy Image, and GT stands for True Image.

[0080] Table 1 shows the SateHaze1k results. For the compared methods, the present invention uses bold and underline to mark the best and second best methods.

[0081] Table 1

[0082]

[0083] Table 2 shows the HRSD results. For the compared methods, the present invention uses bold and underline to mark the best and second best methods;

[0084] Table 2

[0085]

[0086] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A remote sensing image defogging method, characterized in that: A convolutional network including multi-scale and multi-dimensional modules is used to fit the mapping relationship between the foggy image and the clear image, including the following steps: S1: Establish a dehazing model; S2: Design a dehazing network with multi-scale and multi-dimensional modules. Each layer of the network extracts multi-scale and multi-dimensional features and then enters the atmospheric scattering model unit for processing; S3: Collect training dataset; S4: training a dehazing network with multi-scale and multi-dimensional modules; S5: Remote sensing image dehazing.

2. A remote sensing image defogging method according to claim 1, characterized in that: In step S1, the dehazing model: r=F(h) Among them, h represents the foggy image, r represents the restored clear image, and the function F represents the mapping relationship between the foggy image and the corresponding clear image. The mapping relationship F(h) between the foggy image and the clear image is fitted through the scale multi-dimensional physical convolutional network. In this structure, the overall framework of the network is based on U-Net, and each layer of upsampling or downsampling has a feature extraction block. In the extraction block structure, the multi-scale block and the multi-dimensional extraction block extract multi-scale and multi-dimensional information in parallel, and then enter the atmospheric scattering model unit for effective defogging. The atmospheric scattering model is described as: I(x)=J(x)t(x)+A[1-t(x)] Where I(x) is the image affected by haze; J(x) is the haze-free image to be restored; t(x) is the transmittance of light through the atmospheric medium; x represents the image coordinate; A is the global atmospheric light constant.

3. The remote sensing image defogging method according to claim 1, characterized in that: In step S2, the multi-scale and multi-dimensional module includes a multi-scale extraction block and a multi-dimensional extraction block, and the multi-scale extraction block includes the following steps: S211: Apply multiple parallel dilated convolutions with different dilation rates S212: Fusion of multi-scale features using selective kernel module; S213: After processing the feature map, the learned features will go through the pixel attention mechanism; S214: Introduced a channel attention mechanism that runs in parallel and adaptively highlights fog-related features in the entire image; S215: The features extracted by the two attention mechanisms are fused together by summing to create a feature set. In order to refine the fused features, a 3×3 convolution is finally applied The multidimensional extraction block includes the following steps: S221: Apply deep convolution to preserve feature representation while ensuring computational efficiency; S222: normalize to stabilize feature distribution, thereby improving convergence during training; S223: Use a combination of 1×1 convolution and 3×3 convolution layers to expand the number of channels and enhance the richness of feature information. 1×1 convolution increases the depth of the feature map, while 3×3 convolution captures local patterns. S224: The normalized features extracted from the convolution are fused together to more fully represent the input data. After fusion, another 3×3 convolution is applied to further refine the extracted features. S225: After fusion, another 3×3 convolution is applied to further refine the extracted features, and then the refined features are multiplied with another branch processed by the Sigmoid function, effectively rescaling the feature values ​​and enhancing the dimensional representation; S226: Use 1×1 convolution to reduce dimensionality to compress useful information and prevent overfitting; S227: The reduced feature map is passed to another deep convolutional layer and the residual connection is merged to retain the original feature information, producing the final output, which effectively combines the enhanced and retained features, ready for subsequent processing tasks.

4. The remote sensing image defogging method according to claim 1, characterized in that: In step S2, for the number of blocks in each layer, the configuration of [N1, N2, N3, N4, N5] is set to [2, 2, 4, 2, 2], and the corresponding embedding channels are [24, 48, 96, 48, 24], respectively learning multiple sub-maps in the dehazing model, and the last layer of the network maps the feature map to the original image dimension.

5. The remote sensing image defogging method according to claim 1, characterized in that: In step S3, two synthetic remote sensing haze datasets SateHaze1k and HRSD are used; The SateHaze1k dataset includes three subsets: light fog, moderate fog, and thick fog. Each subset contains a training set and a test set. Light fog images use haze masks from real clouds, while moderate haze samples combine the features of fog and moderate haze. Thick fog images are generated using the transmission map of dense fog. The HRSD dataset consists of two subsets: LHID and DHID. LHID contains training images and test images generated by the atmospheric scattering model to train the network to handle different degrees of haze in remote sensing images. At the same time, DHID includes images created from multiple real haze maps to present the haze more realistically. The combination of these two subsets ensures a comprehensive evaluation of dehazing methods in both synthetic haze scenes and more realistic haze scenes.

6. The remote sensing image defogging method according to claim 1, characterized in that: In step S4, the defogging method based on learning uses labeled fog samples for training. The datasets used are SateHazelk and HRSD. In this step, the network is trained with the collected foggy image samples to learn the mapping relationship between foggy images and clear images. The L1 loss function is used to facilitate defogging, where: L1=||J-GT||1 Among them, J is the actual output result of the network, GT is the true value image, multi-angle random rotation and horizontal flipping are used to enhance the training data set, and the stochastic gradient descent method is selected to optimize the L1 loss function.

7. The remote sensing image defogging method according to claim 1, characterized in that: In step S5, the original foggy image is re-entered into the atmospheric scattering model unit at the final output of the defogging network to guide the network to restore according to the original image structure.

8. The remote sensing image defogging method according to claim 1, characterized in that: In step S212, the selective kernel module adopts hybrid SK fusion, which introduces global maximum pooling in parallel with global average pooling to better preserve important details such as edges and textures. In order to balance the contributions of global average pooling and global maximum pooling, a hybrid module with a learnable parameter m is introduced. The Sigmoid function is used to convert this parameter into a weight λ, which is dynamically adjusted during training, allowing the model to adaptively prioritize global context or significant details as needed.

Images