Cross-layer collaborative interaction remote sensing image defogging method

Abstract

A cross-layer collaborative interaction remote sensing image defogging method comprises the following steps: step S1, collecting a remote sensing image data set containing different concentrations of haze, and carrying out image screening and preprocessing; s2, constructing a cross-layer collaborative interaction remote sensing image defogging network framework, wherein the framework comprises a U-shaped defogging network and a cross-layer collaborative interaction module CLCIM; s3, based on the preprocessed remote sensing image data set, using a proper loss function and an optimizer to train a defogging network, and optimizing the performance of the model in the aspects of defogging effect and image quality; and S4, performing optimization by adopting the trained defogging model, and performing online reasoning and defogging functions of the real-time remote sensing image. Through the steps, defogging of the remote sensing image can be realized. The invention aims to solve the technical problem that the existing remote sensing image defogging network based on a U-shaped frame is insufficient in the aspect of cross-layer information interaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a technology for removing haze from remote sensing images using deep learning technology, and specifically relates to a cross-layer collaborative interaction remote sensing image dehazing method. BACKGROUND

[0002] Haze has a serious impact on the image quality captured by optical sensors, resulting in problems such as insufficient contrast, texture deviation, and color distortion in the image. These problems not only weaken the visual effect of the image, but also severely limit subsequent advanced vision tasks such as target detection, image segmentation, and scene understanding, leading to performance degradation. In the field of remote sensing, haze-induced spatial resolution and spectral information distortion greatly restrict the effectiveness of remote sensing images in key applications such as environmental monitoring, disaster warning, agricultural monitoring, and urban planning. Therefore, dehazing technology has become a research hotspot in the field of remote sensing image processing.

[0003] Early research on remote sensing image dehazing technology mainly relied on atmospheric scattering models, introducing various prior knowledge to estimate the transmission map and atmospheric light, thereby restoring the haze-free image. For example, dark channel prior, color attenuation prior, and non-local prior. These methods can improve image clarity to some extent, and can achieve good dehazing results in simple scenes. However, as the complexity of the scene increases, the estimation error of the model also increases, limiting the accuracy of the dehazing effect. In recent years, deep learning technology has driven innovation in the field of image restoration, and researchers have begun to build end-to-end models, such as network models based on U-shaped framework and straight tube structure, to learn the non-linear mapping between hazy images and clear images. Among them, the dehazing algorithm based on the U-shaped framework effectively preserves texture details and significantly improves visual quality through multi-layer feature fusion and layer-by-layer refinement of image reconstruction, and is widely used in various complex scenes. Although the dehazing network based on the U-shaped framework has made some progress in remote sensing image dehazing, there are still some challenges in the feature extraction and transmission process of existing models: Existing frameworks lack effective hierarchical feature interaction, for example, Song et al. in Learning an effective transformer for remote sensing satellite image dehazing designed RSDformer, which introduced a detail compensation attention mechanism, dynamic filtering to enhance feature learning, and multi-scale feature fusion to improve the ability to capture global and local dependencies in remote sensing image dehazing. However, this method ignores the interaction of different levels of features in the encoding and decoding stages, and fails to fully capture the correlation between global and local information, limiting the network's ability to understand the global context, resulting in poor dehazing results. SUMMARY

[0004] The purpose of this invention is to address the shortcomings of current U-shaped frame-based remote sensing image dehazing networks in cross-layer information interaction, in line with the technical limitations of such networks. This is based on the work of Song et al. in their paper "Learning an effective transformer for remote sensing satellite image..." Taking the RSDformer proposed in "Dehazing" as an example, it adopts a U-shaped framework composed of multi-level encoders and decoders, and achieves the alignment and transfer of shallow and deep features through skip connections. However, RSDformer does not design an explicit cross-layer collaborative interaction mechanism. In the encoding stage, there is a lack of lateral information exchange between encoders at each level, making it difficult to collaboratively model multi-scale semantic features. In the decoding stage, there is also no cross-layer feedback and interaction channel between different decoders, which limits the ability to complete details and restore structures during the decoding process. In addition, the encoder and decoder only perform feature splicing through simple skip connections, lacking a deep semantic fusion strategy, making it difficult to fully release the complementary potential of features at different levels. These structural limitations cause the model to face problems such as insufficient feature transfer and unsatisfactory detail restoration when processing high-resolution remote sensing image dehazing tasks. To address this, this invention proposes a cross-layer collaborative interaction mechanism, aiming to strengthen cross-layer information interaction in the encoding and decoding stages of the U-shaped network, improve feature fusion efficiency and expressive power, and thus significantly enhance the performance of remote sensing image dehazing.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A cross-layer collaborative interactive method for dehazing remote sensing images includes the following steps: Step S1: Collect a dataset of remote sensing images containing haze of different concentrations, and perform image filtering and preprocessing; Step S2: Construct a cross-layer collaborative interactive remote sensing image dehazing network framework, which includes a U-shaped dehazing network and a cross-layer collaborative interactive module CLCIM; Step S3: Based on the preprocessed remote sensing image dataset, train the dehazing network using an appropriate loss function and optimizer to improve the model's performance in terms of dehazing effect and image quality; Step S4: Optimize the trained dehazing model to perform online inference and dehazing functions for real-time remote sensing images; The above steps can be used to remove fog from remote sensing images.

[0006] In step S1, remote sensing haze images from different concentration conditions are collected to ensure data diversity; image cleaning and standardization are performed to ensure data quality and consistency, and to enhance the model's generalization ability in remote sensing images so as to effectively support subsequent model training. In step S3, supervised learning training is conducted using the collected remote sensing haze dataset. Based on the task characteristics and model requirements, multiple loss functions are selected, including smooth L1 loss for pixel-level error constraints. Adversarial loss used to improve the realism of generated results Perceptual loss used to maintain visual consistency of images And multi-scale structural similarity loss that can measure structural similarity at multiple levels. During training, the Adam optimizer is used, and peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used as core evaluation metrics. The model performance is monitored and analyzed in real time to ensure that the best defogging effect and excellent generalization ability can be achieved under different smog concentrations and diverse scenarios. In step S4, after training is completed, further fine-tuning can be carried out to optimize the network structure in a targeted manner, improving the overall visual quality while maintaining the defogging effect. The fully optimized and validated model will be deployed to the UAV platform to achieve efficient and low-latency remote sensing defogging processing using its onboard computing power, ensuring that clear images can still be output stably under complex weather and variable environmental conditions, providing reliable and real-time technical support for practical application scenarios such as environmental monitoring, air quality assessment and emergency response.

[0007] In step S3, the loss function used includes: Smoothing L1 loss Introducing a smooth L1 loss function to effectively constrain network behavior; the formula is as follows: ; ; in, This represents a cross-layer collaborative interactive remote sensing image dehazing network, where the first layer in a clear image and the second layer in a blurry image... The values ​​of each pixel are respectively determined by and This indicates that the total number of pixels in the image is ; Combat losses : Adopting the adversarial loss function in generative adversarial networks, where the generator The goal is to convert blurry images into clear, fog-free images, while the discriminator... The task is to distinguish between real, clear images and fog-free images generated by the generator; under this adversarial learning mechanism, the generator continuously adjusts its generation process to make the generated images closer to real images; the formula is defined as follows: ; Perceived loss : To realize the multi-level constraint of image from low-level texture to high-level semantics, the quality of the defogged image is optimized; The specific formula can be expressed as: ; Wherein, represents the input blurred image, represents the target clear image; represents the first layer feature extracted from the VGG16 pre-training model; in the calculation of perceptual loss, represents the number of feature layers in the VGG16 pre-training model used; Multi-scale structural similarity loss : The multi-scale structural similarity loss is introduced to further constrain the generation process of the network, which enhances the similarity between the generated image and the clear image in the structure level, and ensures that the image quality is effectively improved; The calculation formula is as follows: ; Wherein, represents the generated image, represents the clear image; the mean of the generated image and the clear image is and , the standard deviation is and , and represents the covariance between them; the two important terms in the equation are and , and finally, the constant term is and ; Total loss: The total loss function can be expressed as a weighted combination in the following form; The specific formula can be expressed as: ; Wherein, the hyperparameters , , , represent the weight coefficients of each loss term, to balance the contribution of each loss term to the defogging effect.

[0008] In step S2, the cross-layer collaborative interaction remote sensing image defogging network framework includes a U-shaped defogging network and a cross-layer collaborative interaction module CLCIM; Specifically: The output of the first encoder is connected to the input of the second encoder; Then, the outputs of the first encoder and the second encoder are respectively connected with the input of the first cross-layer collaborative interaction module; The output of the second encoder is continuously transmitted to the input of the third encoder, and the output of the third encoder is transmitted to the second cross-layer collaborative interaction CLCIM module together with the output of the first cross-layer collaborative interaction CLCIM module. The output of the second cross-layer collaborative interaction CLCIM module is connected to the input of the residual block module, and the output of the residual block module is transmitted to the input of the third decoder together with the output of the third encoder after residual fusion; The output of the third decoder is transmitted to the input of the second decoder together with the output of the second encoder after residual fusion for further decoding operation; subsequently, the outputs of the third decoder and the second encoder are respectively connected to the input of the third cross-layer collaborative interaction CLCIM module; The output of the second decoder is transmitted to the input of the first decoder together with the result after residual fusion of the output of the first encoder; at this time, the output of the first decoder and the output of the third cross-layer collaborative interaction CLCIM module are respectively connected to the input of the fourth cross-layer collaborative interaction CLCIM module; The input end of the first encoder is used for inputting the foggy image, and the output end of the fourth cross-layer collaborative interaction module is used for outputting the clear image.

[0009] In the encoding stage, the input foggy image is extracted layer by layer by the first encoder, the second encoder and the third encoder to extract multi-scale features; after entering the decoding stage, the network restores the spatial resolution of the image by step-by-step decoding; the third decoder, the second decoder and the first decoder are respectively connected to the third encoder, the second encoder and the first encoder through residual connection for feature fusion, thereby significantly improving the detail restoration capability of the image; finally, the network generates a clear image after dehazing; at the same time, the cross-layer collaborative interaction module, i.e. the CLCIM module, interacts with the features between layers in the encoding and decoding stages, so as to ensure that the encoder and the decoder can not only capture local details within the layer, but also extract global context information across layers.

[0010] Specifically, when constructing the U-shaped dehazing network, the input foggy image is extracted layer by layer through four times of down-sampling; the first encoder is responsible for basic feature extraction, the second encoder further enhances these features, and the third encoder focuses on capturing deep semantic features of the image; the down-sampled feature map then extracts more semantic information through the residual block; in the decoding stage, the features extracted by the encoder are step-by-step restored to the spatial resolution of the image through up-sampling; the third decoder is responsible for restoring rough features and step-by-step refining the low-resolution information of the image, the second decoder further improves the integrity of the information through feature fusion, and the first decoder focuses on restoring the details of the image, and finally restores the image to the original size; through the residual connection, the output of the encoder is connected to the decoder layer, so that the features can be directly transmitted, thereby avoiding information loss and speeding up the training process of the network.

[0011] The constructed cross-layer collaborative interaction CLCIM module includes a shallow branch network, a deep branch network and an inter-layer interaction network; For the shallow branch network: first, the initial shallow feature map output connects the input of the shallow normalization LN block to enhance the training stability; the shallow normalization LN block feature map channel is divided into a shallow local feature map and a shallow global feature map; the shallow local feature map is input to the shallow local first 3*3 depth convolution to obtain a shallow local first 3*3 depth convolution feature map, the shallow local first 3*3 depth convolution feature map is input to the shallow local first 5*5 depth convolution to obtain a shallow local first 5*5 depth convolution feature map, and the shallow local first 5*5 depth convolution feature map is input to the shallow local first 7*7 depth convolution to obtain a shallow local first 7*7 depth convolution feature map; at the same time, the shallow global feature map is input to the shallow global first 3*3 depth convolution to obtain a shallow global first 3*3 depth convolution feature map, the shallow global first 3*3 depth convolution feature map is input to the shallow global second 3*3 depth convolution to obtain a shallow global second 3*3 depth convolution feature map, and the shallow global second 3*3 depth convolution feature map is input to the shallow global Mamba to obtain a shallow global Mamba feature map; this process can be represented as: ; wherein, is the shallow local feature map, represents a depth convolution operation on the shallow local feature map, represents a cascaded depth convolution operation on the shallow local feature map, refers to the shallow local first 7*7 depth convolution feature map; is the shallow global feature map, represents a depth convolution operation on the shallow global feature map, represents a continuous depth convolution operation on the shallow global feature map, represents using the Mamba block to process the shallow global second 3*3 depth convolution feature map, is the shallow global Mamba feature map; Subsequently, the shallow local first 7*7 depth convolution feature map and the shallow global Mamba feature map are channel spliced to obtain a shallow spliced feature map; the shallow spliced feature map is input to the shallow first 1*1 convolution to obtain a shallow first 1*1 convolution feature map; this process can be represented as: ; wherein, is a splicing operation, and the spliced feature is fused using ; represents the fused shallow first 1*1 convolution feature map, the size of which is restored to ; this strategy effectively balances detail preservation and global information modeling, while maintaining efficient computation and enhancing adaptability to complex visual information; Then, the shallow first 1*1 convolution feature map is input to the shallow second 1*1 convolution to obtain a shallow second 1*1 feature map; the shallow first 1*1 convolution feature map is input to the shallow third 1*1 convolution to obtain a shallow third 1*1 feature map; this process can be represented as: ; wherein, represents a two-dimensional convolution with a kernel size and a step size of 1, respectively represent the feature extraction branch and the gating branch after convolution processing, , so ; Then, the shallow second 1*1 feature map is input to the shallow first 5*5 deep convolution to obtain a shallow first 5*5 deep convolution feature map; the shallow second 1*1 feature map and the shallow first 5*5 deep convolution feature map are connected in residual connection and multiplied by the shallow third 1*1 feature map element to obtain a first weighted feature map, and the first weighted feature map output is connected in residual connection with the shallow first 1*1 convolution feature map to obtain a first gating feature map; this process can be represented as: ; wherein, represents a deep convolution with a kernel size of 5x5 and a step size of 1, represents a feature map processed by a gating mechanism, represents the final output feature map; the gating feature enhancement block enhances key information and suppresses redundant information, maintains spatial structure integrity, optimizes pixel-level feature interaction, and provides richer feature representation.

[0012] For the deep branch network: first, the initial deep feature map output is connected to the input of the deep normalization LN block to enhance the training stability; the feature map channel of the deep normalization LN block is divided into a deep local feature map and a deep global feature map; the deep local feature map is input to the deep local first 3*3 deep convolution to obtain a deep local first 3*3 deep convolution feature map, the deep local first 3*3 deep convolution feature map is input to the deep local first 5*5 deep convolution to obtain a deep local first 5*5 deep convolution feature map, and the deep local first 5*5 deep convolution feature map is input to the deep local first 7*7 deep convolution to obtain a deep local first 7*7 deep convolution feature map; at the same time, the deep global feature map is input to the deep global first 3*3 deep convolution to obtain a deep global first 3*3 deep convolution feature map, the deep global first 3*3 deep convolution feature map is input to the deep global second 3*3 deep convolution to obtain a deep global second 3*3 deep convolution feature map, and the deep global second 3*3 deep convolution feature map is input to the deep global Mamba to obtain a deep global Mamba feature map; this process can be represented as: ​ ; wherein, is a deep local feature map, represents a deep convolution operation on the deep local feature map, represents a cascaded deep convolution operation on the deep local feature map, refers to a deep local first 7*7 deep convolution feature map; is a deep global feature map, represents a deep convolution operation on the deep global feature map, represents a successive deep convolution operation on the deep global feature map, represents processing a deep global second 3*3 deep convolution feature map using a Mamba block, is a deep global Mamba feature map; Subsequently, the deep local first 7*7 deep convolution feature map and the deep global Mamba feature map are channel spliced to obtain a deep spliced feature map; the deep spliced feature map is input into a deep first 1*1 convolution to obtain a deep first 1*1 convolution feature map; this process can be represented as: ; wherein, is a splicing operation, and the spliced feature is fused using , represents the fused deep first 1*1 convolution feature map, and the size is restored to ; this strategy effectively balances detail preservation and global information modeling, while maintaining efficient computation and enhancing adaptability to complex visual information; Next, the deep first 1*1 convolution feature map is input into a deep second 1*1 convolution to obtain a deep second 1*1 feature map; the deep first 1*1 convolution feature map is input into a deep third 1*1 convolution to obtain a deep third 1*1 feature map; this process can be represented as: ; wherein, represents a two-dimensional convolution with a convolution kernel size and a step size of 1, respectively represent a feature extraction branch and a gating branch after convolution processing, so ; Then, the deep second 1*1 feature map is input to the deep first 5*5 depth convolution to obtain a deep first 5*5 depth convolution feature map; the deep second 1*1 feature map and the deep first 5*5 depth convolution feature map are connected in residual and multiplied by the deep third 1*1 feature map element to obtain a second weighted feature map, and the second weighted feature map output is connected in residual with the deep first 1*1 convolution feature map to obtain a gated second feature map; this process can be represented as: ; wherein, represents a depth convolution with a convolution kernel of 5x5 and a step of 1, represents a feature map processed by a gating mechanism, represents a final output feature map; the gated feature enhancement block enhances key information and suppresses redundant information, maintains spatial structure integrity, optimizes pixel-level feature interaction, and provides richer feature representation; Inter-layer interaction network: to perform inter-layer interaction and capture global context information, a reshaping first operation is used on the gated first feature map to obtain a shallow sequence, and a reshaping second operation is used on the gated second feature map to obtain a deep sequence; this process can be represented as: ; wherein, represents a reshaping operation on the obtained gated feature map, represents mapping from a spatial domain to a channel-related representation, is composed of and , and the size changes from HxWxC to ; Further, the shallow sequence is subjected to a transposition operation to obtain a transposed sequence, the transposition operation changes the scale of the transposed sequence from to , and matrix multiplication is performed with the deep sequence with a shape of , and the result is input to Softmax to obtain inter-layer weights with a shape of CxC; subsequently, matrix operation is performed on the shallow sequence and the inter-layer weights; this process can be represented as: ; wherein, represents the calculated inter-layer weights, Softmax enhances relevant features and suppresses irrelevant features, and generates more effective attention scores; fuses different levels of information to generate context-related features; Finally, a reshaping third operation is used on the matrix operation result of the shallow sequence and the inter-layer weights, and an initial shallow feature map and an initial deep feature map are connected in residual and input to a 1*1 convolution to obtain a final feature map; which can be represented as: ; wherein, represents an initial shallow feature map, represents an initial deep feature map, represents a final feature map. This operation fully integrates shallow and deep features and can fully learn global context information.

[0013] Compared with the prior art, the present application has the following technical effects: 1) In order to overcome the problem of insufficient utilization of global context information in the traditional U-shaped dehazing network, the present application proposes a cross-layer collaborative interaction module CLCIM, which strengthens the inter-layer feature interaction and cross-layer collaborative optimization in the encoding and decoding stages, enhances the network's modeling ability for global context, effectively alleviates the semantic dilution problem, and improves the dense haze image restoration effect. 2) The present application proposes a cross-layer collaborative interaction remote sensing image dehazing network framework for unmanned aerial vehicle recovery remote sensing haze image. By seamlessly integrating the cross-layer collaborative interaction module into the U-shaped network, the present application significantly enhances the feature transmission and information fusion efficiency in the image restoration process while improving the dehazing effect. BRIEF DESCRIPTION OF DRAWINGS

[0014] The present application will be further described below in conjunction with the drawings and examples: Figure 1 is a method flowchart of the present application; Figure 2 is a whole network framework structure diagram of the present application; Figure 3 is Figure 2 a structure diagram of the cross-layer collaborative interaction module CLCIM. DETAILED DESCRIPTION

[0015] As Figure 1 shown, the present embodiment provides a cross-layer collaborative interaction remote sensing image dehazing method, comprising the following steps: Step S1: Collect remote sensing image data sets containing different concentrations of haze, perform image screening and preprocessing, ensure the quality of the data set, and enhance the generalization ability of the model in the remote sensing image; Step S2: Construct a cross-layer collaborative interaction remote sensing image dehazing network framework, which includes a U-shaped dehazing network and a cross-layer collaborative interaction module CLCIM; the cross-layer collaborative interaction module CLCIM enhances the learning ability of context information in the image dehazing process through inter-layer feature interaction, and improves the dehazing effect.

[0016] Step S3: Based on the pre-processed remote sensing image dataset, train the defogging network using appropriate loss functions and optimizers to optimize the performance of the model in terms of defogging effect and image quality. Step S4: Optimize the trained defogging model and deploy it to the UAV platform for online inference and defogging function of real-time remote sensing images.

[0017] Through the above steps, the remote sensing defogging function on the UAV can be realized.

[0018] Step S1 is as follows: To ensure that the model has good adaptability and generalization ability under various fog conditions, remote sensing fog images from different concentrations and geographical regions should be systematically collected, covering a variety of scenes such as light, medium and heavy fog. Then, the collected images are strictly cleaned and processed in terms of format and numerical range while keeping the original features, so as to ensure data quality and consistency, laying a solid foundation for efficient training and accurate inference of subsequent models.

[0019] Step S2 is as follows: As shown in Figure 2 The method proposes a cross-layer collaborative interaction remote sensing image defogging network framework, which includes a U-shaped defogging network and a cross-layer collaborative interaction module CLCIM. In the encoding stage, the input foggy image is extracted layer by layer through the first encoder 200, the second encoder 201 and the third encoder 202 to obtain multi-scale features. After entering the decoding stage, the network restores the spatial resolution of the image by step-by-step decoding. The third decoder 203, the second decoder 204 and the first decoder 205 are connected through residual connection and respectively perform feature fusion with the third encoder 202, the second encoder 201 and the first encoder 200, thereby significantly improving the detail restoration ability of the image. Finally, the network generates a clear image after defogging. At the same time, the cross-layer collaborative interaction module CLCIM interacts between the layers in the encoding and decoding stages to ensure that the encoder and decoder can not only capture local details within the layer, but also extract global context information across layers.

[0020] Specifically, in the construction of the U-shaped dehazing network, the input foggy image extracts features layer by layer through four times of down-sampling. The first encoder 200 is responsible for basic feature extraction, the second encoder 201 further enhances these features, and the third encoder 202 focuses on capturing deep semantic features of the image. The down-sampled feature map is then extracted by the residual module 3 to extract more semantic information. In the decoding stage, the features extracted by the encoder are gradually restored to the spatial resolution of the image through up-sampling. The third decoder 203 is responsible for restoring the relatively rough features and gradually refining the low-resolution information of the image. The second decoder 204 further improves the integrity of the information through feature fusion, and the first decoder 205 focuses on restoring the details of the image, finally restoring the image to the original size. Through the residual connection, the output of the encoder is connected with the decoder layer, so that the features can be directly transmitted, thereby avoiding information loss and speeding up the training process of the network.

[0021] As Figure 3 (a) shows, in the construction of the cross-layer collaborative interaction module CLCIM, the remote sensing image dehazing method based on the U-shaped network improves the image reconstruction effect through a unique multi-scale feature fusion mechanism. However, the existing method still has shortcomings in the design of the cross-layer interaction mechanism, which makes it difficult for deep semantic features and shallow details to be fully fused, thereby affecting the effective transmission of features in the network. Therefore, the present application designs a cross-layer collaborative interaction module based on the U-shaped network, which effectively aggregates global context information, enhances feature expression, and significantly improves the recovery quality of remote sensing image dehazing. In the encoding and decoding stages, the inter-layer features are aligned through basic convolution to ensure that they have the same resolution, and the feature information and are obtained. and .

[0022] As Figure 3 (b) shows, in order to efficiently integrate local and global information and reduce computational complexity, a Mamba-CNN fusion block MCFB is designed, which uniformly divides the input feature in the channel dimension through a channel allocation strategy, and inputs it into the local path and the global path for independent processing. The former focuses on extracting local structural features and strengthening texture. The latter first extracts features to capture edge information, and then combines the Mamba block to capture long-range dependencies and optimize overall semantic expression. The features of the double paths are mutually enhanced in the fusion stage, achieving a better balance between feature expression ability and computational efficiency. In particular, corresponding to the feature information transmitted between different layers in the encoding and decoding stages. It can be represented as.

[0023] ; wherein, is a local feature map, represents a depth convolution operation on the local feature map, represents a cascaded depth convolution operation on the local feature map, refers to an extracted local feature map in a layer; is a global feature map, represents a depth convolution operation on the global feature map, represents a successive depth convolution operation on the global feature map, represents processing of a depth convolution feature map subjected to two 3*3 depth convolution processes using a Mamba block, is an extracted global feature map in a layer.

[0024] Then, the local and global features are fused to further improve the representation ability of image features and the reconstruction effect.

[0025] ; wherein, is a concatenation operation, and the feature after concatenation is subjected to feature fusion, and the fused feature is restored to . This strategy effectively balances detail preservation and global information modeling, while maintaining efficient computation and enhancing adaptability to complex visual information.

[0026] As shown in Figure 3 (c), in order to reduce the computational complexity while preserving key feature information, a gated feature enhancement block GFEB is designed. Specifically, for the obtained fused feature , the present application first expands the input feature channel, which is divided into two parts, wherein the feature extraction branch is responsible for capturing spatial information, and the gating branch is responsible for dynamically adjusting the feature flow to enhance the expression of key features. The specific operation is as follows.

[0027] ; wherein, represents a two-dimensional convolution with a convolution kernel size and a step size of 1, respectively represent the feature extraction branch and the gating branch after convolution processing, so .

[0028] The local spatial information of is effectively extracted by depth convolution, and feature transmission is realized by combining residual connection. The gating mechanism is used to adjust the feature flow of The extracted wide-area features are refined to enhance the feature expression capability. Finally, the fused features are mapped to the output channel, combined with the residual connection and the original input features to enhance the stability of information flow.

[0029] wherein, represents a deep convolution with a 5x5 kernel and a step of 1, represents a feature map processed by a gating mechanism, represents the final output feature map. The gating feature enhancement block enhances key information and suppresses redundant information, maintains spatial structure integrity, optimizes pixel-level feature interaction, and provides richer feature representation.

[0030] As shown in Figure 3 (d), in order to enhance the cross-layer feature learning ability and information transmission, an inter-layer feature interaction block ILFIB is designed to further improve the performance of the U-shaped network in remote sensing image defogging. Specifically, the feature information and reshaped by the intra-layer enhancement is reshaped. It can be represented as: wherein, represents a reshaping operation on the obtained gating feature map, represents mapping from the spatial domain to the channel-related representation, is composed of and , and the size changes from HxWxC to In order to improve the global context understanding ability, matrix multiplication is performed to establish the relevance mapping of inter-layer features, and by adaptively adjusting the importance of different features, an attention map is generated. Finally, the attention map is used for feature enhancement to dynamically adjust the information flow, so that the network can more effectively fuse low-level details and high-level semantics.

[0031] wherein, Softmax enhances relevant features and suppresses irrelevant features to generate more effective attention scores. The context-related features are generated by fusing information of different levels.

[0032] The size of the context feature information is reshaped to , and the feature information , and are spliced. Then,​​​​ Enhance important features and remove redundant information, and finally generate a more expressive inter-layer fusion feature .

[0033] ; wherein, represents an initial shallow feature map, represents an initial deep feature map, represents a final feature map.

[0034] In summary, CLCIM effectively fuses local details and global information, strengthens feature expression, and promotes the deep fusion of shallow details and deep semantics in the encoding and decoding stages. It greatly improves the adaptability to complex scenes and enhances the image restoration ability in strong haze conditions.

[0035] Step S3 is specifically as follows: Supervised learning training is carried out by collecting remote sensing haze data sets, and a variety of loss functions are comprehensively selected according to the task characteristics and model requirements, including smooth L1 loss for pixel-level error constraint , adversarial loss for improving the authenticity of the generated results , perceptual loss for maintaining image visual consistency , and multi-scale structural similarity loss for measuring structural similarity at multiple levels . In the training process, the Adam optimizer is adopted, and the peak signal-to-noise ratio PSNR and structural similarity SSIM are used as the core evaluation indexes to monitor and analyze the model performance in real time, so as to ensure that the best haze removal effect and excellent generalization ability can be achieved under different haze concentrations and diversified scenes. The loss function of the haze removal network is specifically as follows.

[0036] Smooth L1 loss : The present application introduces a smooth L1 loss function to effectively constrain the network behavior. Compared with L1 loss, smooth L1 loss provides a smoother change near zero, and compared with L2 loss, it reduces the excessive sensitivity to extreme values. These characteristics make it perform better than traditional methods in the image haze removal task. The calculation formula is as follows.

[0037] ; ; wherein, represents a cross-layer collaborative interaction remote sensing image haze removal network, the values of the first pixel in the clear image and the blurred image are represented by and , and the total number of pixels of the image is .

[0038] Combat losses This invention employs the adversarial loss function in generative adversarial networks, where the generator... The goal is to convert blurry images into clear, fog-free images, while the discriminator... The task is to distinguish between a real, clear image and a fog-free image generated by the generator. Under this adversarial learning mechanism, the generator continuously adjusts its generation process to make the generated image closer to the real image. The formula is defined as follows.

[0039] ; Perceived loss Perceptual loss is widely used in image restoration tasks. L2 loss is calculated for different feature layers on a pre-trained VGG16 model to achieve multi-level constraints on the image from low-level texture to high-level semantics, thus optimizing the quality of the dehazed image. The specific formula can be expressed as follows.

[0040] ; in, This represents the blurred input image. This indicates that the target image is clear. This indicates the first [item] extracted from the VGG16 pre-trained model. Layer features. In the calculation of perceptual loss, This indicates the number of feature layers in the VGG16 pre-trained model used.

[0041] Multiscale structural similarity loss Multi-scale structural similarity loss is introduced to further constrain the network's generation process, enhancing the structural similarity between the generated image and the clear image, thus ensuring an effective improvement in image quality. The calculation formula is as follows.

[0042] ; in, This represents the generated image. This represents a sharp image. The mean values ​​of the generated image and the sharp image are respectively... and The standard deviations are respectively and ,and This represents the covariance between them. Two important terms in the equation are... and Finally, the constant term is and .

[0043] Total Loss: The total loss function can be expressed as a weighted combination in the following form. The specific formula is as follows.

[0044] ; wherein the hyperparameters , , , represent the weight coefficients of each loss term to balance the contribution of each loss term to the defogging effect.

[0045] Step S4 is specifically as follows: After the training is completed, fine-tuning can be further carried out, and the network structure is optimized in a targeted manner, so that the overall visual quality is improved while the defogging effect is maintained. The model that is fully optimized and verified will be deployed to the unmanned aerial vehicle platform, and the airborne computing power is used to realize efficient and low-delay remote sensing defogging processing, so that clear images can still be stably output under complex weather and variable environmental conditions, and reliable and real-time technical support is provided for practical application scenarios such as environmental monitoring, atmospheric quality evaluation and emergency response.

[0046] Regarding the framework constructed in the application, the following is specific: In step S2, a remote sensing image defogging network with cross-layer cooperative interaction is constructed, and the structure of the framework is specifically as follows: As shown in Figure 2 , a foggy image with a size of HxWx3 is processed by reflection padding. Then, the image is input into the first encoder 200, and a 7*7 convolution kernel is used to generate a preliminary feature map with a size of HxWx64.

[0047] The output of the first encoder 200 is connected to the input of the second encoder 201. The second encoder 201 uses a 3*3 convolution kernel for encoding operation, and the feature map is reduced from HxWx64 to H / 2xW / 2x128. Subsequently, the outputs of the first encoder 200 and the second encoder 201 are respectively connected with the inputs of the first cross-layer cooperative interaction module 1, which is used to capture the preliminary context information of the encoder.

[0048] The output of the second encoder 201 is further transmitted to the input of the third encoder 202, and the third encoder 202 uses a 3*3 convolution kernel for encoding operation again, and further reduces the feature map size from H / 2xW / 2x128 to H / 4xW / 4x256. At this time, the output of the third encoder 202 is transmitted to the second cross-layer cooperative interaction module 2 together with the output of the first cross-layer cooperative interaction module 1, to further capture the global context information of the encoder.

[0049] The output of the second cross-layer cooperative interaction module 2 is connected to the input of the residual block module 3, which extracts high-level features and ensures the stability of the information flow. The output of the residual block module 3 is transmitted to the input of the third decoder 203 together with the output of the third encoder 202 after residual fusion.

[0050] The third decoder 203 realizes the decoding operation through the deconvolution operation, and recovers the feature map from H / 4*W / 4*256 to H / 2*W / 2*128. At the same time, the output of the third decoder 203 is transmitted to the input of the second decoder 204 together with the output of the second encoder 201 after residual fusion, and further decoding operation is performed. Subsequently, the outputs of the third decoder 203 and the second encoder 204 are respectively connected with the input of the third cross-layer collaborative interaction module 4, which is used to capture the preliminary context information of the decoder.

[0051] The output of the second decoder 204 is transmitted to the input of the first decoder 205 together with the result after residual fusion of the output of the first encoder 200, and the feature map size is recovered from H / 2*W / 2*128 to H*W*64. At this time, the output of the first decoder 205 is respectively connected with the output of the third cross-layer collaborative interaction module 4 to the input of the fourth cross-layer collaborative interaction module 5, which further captures the global context information of the decoder.

[0052] The output of the fourth cross-layer collaborative interaction module 5 is output through the final mapping operation, combined with the 7*7 convolution kernel for convolution, to generate the clear image after defogging, and the output resolution is H*W*3. This process effectively preserves the details and ensures the integrity of the reconstructed image.

[0053] Among them, the constructed cross-layer collaborative interaction module CLCIM includes a shallow branch network, a deep branch network and an inter-layer interaction network: As shown in the shallow branch network: Figure 2 First, the initial shallow feature map 300 is output to connect the input of the shallow normalization LN block to enhance the training stability. The channel of the shallow normalization LN block feature map 301 is divided into a shallow local feature map and a shallow global feature map. The shallow local feature map is input to the shallow local first 3*3 depth convolution to obtain the shallow local first 3*3 depth convolution feature map 310, the shallow local first 3*3 depth convolution feature map 310 is input to the shallow local first 5*5 depth convolution to obtain the shallow local first 5*5 depth convolution feature map 311, and the shallow local first 5*5 depth convolution feature map 311 is input to the shallow local first 7*7 depth convolution to obtain the shallow local first 7*7 depth convolution feature map 312. At the same time, the shallow global feature map is input to the shallow global first 3*3 depth convolution to obtain the shallow global first 3*3 depth convolution feature map 313, the shallow global first 3*3 depth convolution feature map 313 is input to the shallow global second 3*3 depth convolution to obtain the shallow global second 3*3 depth convolution feature map 314, and the shallow global second 3*3 depth convolution feature map 314 is input to the shallow global Mamba to obtain the shallow global Mamba feature map 315. This process can be represented as: ; wherein, is a shallow local feature map, represents a deep convolution operation on the shallow local feature map, represents a cascaded deep convolution operation on the shallow local feature map, refers to a shallow local first 7*7 deep convolution feature map. is a shallow global feature map, represents a deep convolution operation on the shallow global feature map, represents a successive deep convolution operation on the shallow global feature map, represents processing the shallow global second 3*3 deep convolution feature map using the Mamba block, is a shallow global Mamba feature map.

[0054] Subsequently, the shallow local first 7*7 deep convolution feature map 312 and the shallow global Mamba feature map 315 are channel spliced to obtain a shallow spliced feature map 320. The shallow spliced feature map 320 is input to a shallow first 1*1 convolution to obtain a shallow first 1*1 convolution feature map 321. This process can be represented as: ; wherein, is a splicing operation, and the spliced feature is fused using , represents the fused shallow first 1*1 convolution feature map, and the size is restored to . This strategy effectively balances detail preservation and global information modeling, while maintaining efficient computation and enhancing adaptability to complex visual information.

[0055] Next, the shallow first 1*1 convolution feature map 321 is input to a shallow second 1*1 convolution to obtain a shallow second 1*1 feature map 322. The shallow first 1*1 convolution feature map 321 is input to a shallow third 1*1 convolution to obtain a shallow third 1*1 feature map 324. This process can be represented as: ; wherein, represents a two-dimensional convolution with a convolution kernel size and a step size of 1, respectively represent a feature extraction branch and a gating branch after convolution processing, so .

[0056] Then, the shallow second 1*1 feature map 322 is input to the shallow first 5*5 depth convolution to obtain a shallow first 5*5 depth convolution feature map 323. The shallow second 1*1 feature map 322 and the shallow first 5*5 depth convolution feature map 323 are connected in residual and multiplied by the shallow third 1*1 feature map 324 to obtain a first weighted feature map 325. The first weighted feature map 325 is output and connected in residual with the shallow first 1*1 convolution feature map 321 to obtain a first gated feature map 326. This process can be represented as: ; wherein, represents a depth convolution with a 5x5 kernel and a step of 1, represents a feature map processed by a gating mechanism, represents the final output feature map. The gated feature enhancement block enhances key information and suppresses redundant information, maintains spatial structure integrity, and optimizes pixel-level feature interaction to provide richer feature representation.

[0057] For the deep branch network: first, the initial deep feature map 330 output is connected to the input of the deep normalization LN block to enhance training stability. The deep normalization LN block feature map 331 is divided into a deep local feature map and a deep global feature map. The deep local feature map is input to the deep local first 3*3 depth convolution to obtain a deep local first 3*3 depth convolution feature map 340. The deep local first 3*3 depth convolution feature map 340 is input to the deep local first 5*5 depth convolution to obtain a deep local first 5*5 depth convolution feature map 341. The deep local first 5*5 depth convolution feature map 341 is input to the deep local first 7*7 depth convolution to obtain a deep local first 7*7 depth convolution feature map 342. At the same time, the deep global feature map is input to the deep global first 3*3 depth convolution to obtain a deep global first 3*3 depth convolution feature map 343. The deep global first 3*3 depth convolution feature map 343 is input to the deep global second 3*3 depth convolution to obtain a deep global second 3*3 depth convolution feature map 344. The deep global second 3*3 depth convolution feature map 344 is input to the deep global Mamba to obtain a deep global Mamba feature map 345. This process can be represented as: ; wherein, is a deep local feature map, represents a depth convolution operation on the deep local feature map, represents a cascaded depth convolution operation on the deep local feature map, is a deep local first 7*7 depth convolution feature map. is a deep global feature map, denotes a deep convolution operation on the deep global feature map, denotes a successive deep convolution operation on the deep global feature map, denotes a deep convolution operation on the deep global second 3*3 feature map using the Mamba block, for the deep global Mamba feature map.

[0058] Subsequently, the deep local first 7*7 deep convolution feature map 342 and the deep global Mamba feature map 345 are concatenated to obtain a deep concatenation feature map 350. The deep concatenation feature map 350 is input to a deep first 1*1 convolution to obtain a deep first 1*1 convolution feature map 351. This process can be represented as: ; wherein, is a concatenation operation, and the concatenated feature is fused using , denotes the fused deep first 1*1 convolution feature map, and the size is restored to . This strategy effectively balances detail preservation and global information modeling, while maintaining efficient computation and enhancing adaptability to complex visual information.

[0059] Next, the deep first 1*1 convolution feature map 351 is input to a deep second 1*1 convolution to obtain a deep second 1*1 feature map 352. The deep first 1*1 convolution feature map 351 is input to a deep third 1*1 convolution to obtain a deep third 1*1 feature map 354. This process can be represented as: ; wherein, represents a two-dimensional convolution with a kernel size and a stride of 1, represent the feature extraction branch and the gating branch after convolution processing, respectively, .

[0060] Then, the deep second 1*1 feature map 352 is input to a deep first 5*5 deep convolution to obtain a deep first 5*5 deep convolution feature map 353. The deep second 1*1 feature map 352 and the deep first 5*5 deep convolution feature map 353 are residual connected and multiplied by the deep third 1*1 feature map 354 to obtain a second weighted feature map 355. The second weighted feature map 355 is output and residual connected with the deep first 1*1 convolution feature map 351 to obtain a gated second feature map 356. This process can be represented as: ; wherein, denotes a deep convolution with a kernel of 5x5 and a stride of 1, denotes the feature map processed by the gating mechanism, denotes the final output feature map; the gating feature enhancement block enhances key information and suppresses redundant information, maintains spatial structure integrity, optimizes pixel-level feature interaction, and provides richer feature representation.

[0061] Inter-layer interaction network: To perform inter-layer interaction and capture global context information, a reshaping first operation is used on the gated first feature map 326 to obtain a shallow sequence 327, and a reshaping second operation is used on the gated second feature map 356 to obtain a deep sequence 357. This process can be represented as: ; wherein, denotes a reshaping operation on the obtained gated feature map, represents mapping from the spatial domain to the channel-related representation, consists of and , and the size changes from HxWxC to .

[0062] Further, the shallow sequence 327 is subjected to a transposition operation to obtain a transposed sequence 328. The transposition operation changes the scale of the transposed sequence 328 from to , and matrix multiplication is performed with the deep sequence 357 having a shape of , and the result is input to Softmax to obtain an inter-layer weight 360 having a shape of CxC. Subsequently, matrix operations are calculated for the shallow sequence 327 and the inter-layer weight 360. This process can be represented as: ; wherein, denotes the calculated inter-layer weight, Softmax enhances relevant features and suppresses irrelevant features, generating more effective attention scores. Fusion of different levels of information generates context-related features.

[0063] Finally, a reshaping third operation is used on the matrix operation result of the shallow sequence 327 and the inter-layer weight 360, and a residual connection is used between the initial shallow feature map 300 and the initial deep feature map 330 to input a 1*1 convolution to obtain a final feature map 361. This can be represented as: ; wherein, denotes the initial shallow feature map, denotes the initial deep feature map, denotes the final feature map.

[0064] This operation fully integrates shallow and deep features and can fully learn global context information.

[0065] In order to better understand the present application for those skilled in the art, further description is as follows: 1) Parameter setting The experiment is carried out on an NVIDIA RTX 3090 GPU, and the code is developed using the PyTorch framework. The optimizer is Adam, the learning rate is 0.0001, and the momentum decay index is 0.9 and 0.999, respectively. The present application evaluates a variety of defogging methods on the SateHaze1k remote sensing dataset. The SateHaze1k remote sensing dataset is composed of remote sensing haze scenes of different concentrations captured by GF-2, including light haze Thin, moderate haze Moderate, and thick haze Thick. In order to evaluate the performance of the algorithm proposed in the present application in the remote sensing image defogging task, the present application selects the peak signal-to-noise ratio PSNR and the structural similarity SSIM as quantitative evaluation indicators. These two indicators are widely accepted as standard methods for measuring image quality in the field of image defogging, and can well reflect the visual effect and structural integrity of the defogged image.

[0066] 2) Experimental results To verify the effectiveness of the algorithm of the present application, the present application is compared with the current excellent defogging algorithms, including DCP, DCRD, DCI-Net, Trinity-Net and DCMP-Net, for quantitative comparison, and the dataset includes three sub-datasets of SateHaze1k. Table 1 shows the comparison results of various defogging algorithms on the above datasets. From the experimental results in Table 1, it can be seen that the method proposed in the present application has achieved the optimal PSNR and SSIM indicators on the three haze data subsets of the SateHazel1k dataset, showing excellent defogging performance. Specifically, under the condition of light haze Thin, the PSNR of the method of the present application reaches 26.23, and the SSIM is 0.908, indicating that the method can effectively restore image details and structural information in light haze scenes. Under the condition of moderate haze Moderate, the method of the present application continues to lead, with a PSNR of 27.25 and an SSIM of 0.933, which is significantly better than other comparison methods, proving that the method has stronger robustness and generalization ability under moderate haze. Under the condition of thick haze Thick, the method of the present application also performs best, with a PSNR of 24.15 and an SSIM of 0.851. Compared with the classic method DCP, its PSNR and SSIM are increased by 13.29 and 0.280, respectively, indicating that the method of the present application still has a significant advantage in extremely severe haze environment. In summary, the above results show that the method proposed in the present study has good image restoration effect and structure preservation ability under various haze conditions, and has strong practical application potential and promotional value.

[0067]

[0068] Table 1 Comparison of various algorithms on SateHaze1k dataset 3) Ablation experiment In order to evaluate the effectiveness of each sub-module in the CLCIM module, the present application designs an ablation experiment according to the framework and the innovations, including two experiments: (1) Base represents a U-shaped basic frame; (2) Base + ILFIB block; (3) Base + ILFIB + MCFB block; (4) Base + ILFIB + MCFB + GFEB block.

[0069]

[0070] Table 2 PSNR and SSIM results on Thin dataset According to Table 2, by gradually introducing different modules, the ablation experiment effectively analyzes the role of each module in remote sensing image defogging. First, the PSNR of the Base model is 23.18, and the SSIM is 0.811, indicating that its image restoration effect is relatively limited. After adding the ILFIB block, the PSNR increases to 25.72, and the SSIM increases to 0.895, which shows that ILFIB enhances the image detail restoration ability by optimizing the inter-layer feature fusion. Next, the MCFB block is added, and the PSNR is further improved to 25.98, and the SSIM increases to 0.902, proving the effectiveness of MCFB in improving the efficiency of local and global information fusion and enhancing feature transmission. Finally, the GFEB block is added, and the PSNR and SSIM reach 26.23 and 0.908, respectively, reflecting the contribution of GFEB in strengthening key features and optimizing information flow. In summary, the ablation experiment results verify the effectiveness of the cross-layer collaborative interaction module CLCIM, especially in detail preservation, global information modeling, and optimization of feature flow, indicating that the collaborative action of each module is crucial to the improvement of remote sensing image defogging quality.

Claims

1. A method for remote sensing image defogging based on cross-layer collaborative interaction, characterized in that, The method comprises the following steps: Step S1: collecting a remote sensing image dataset containing different concentrations of haze, performing image screening and preprocessing; Step S2: constructing a cross-layer collaborative interaction remote sensing image defogging network framework, which comprises a U-shaped defogging network and a cross-layer collaborative interaction module CLCIM; Step S3: based on the preprocessed remote sensing image dataset, training the defogging network using appropriate loss functions and optimizers to optimize the performance of the model in terms of defogging effect and image quality; Step S4: using the trained defogging model for optimization, performing online inference and defogging functions on real-time remote sensing images; Through the above steps, the defogging of remote sensing images can be realized.

2. The method of claim 1, wherein, In step S1, remote sensing haze images from different concentration conditions are collected to ensure data diversity; image cleaning and standardization and other preprocessing are performed to ensure data quality and consistency and to enhance the model's generalization ability in remote sensing images, thereby effectively supporting subsequent model training; In step S3, supervised learning training is carried out by using the collected remote sensing haze data set, and according to the task characteristics and model requirements, a plurality of loss functions are comprehensively selected, including a smooth L1 loss for pixel-level error constraint , an adversarial loss for improving the authenticity of the generated results , a perception loss for maintaining the visual consistency of the image , and a multi-scale structural similarity loss capable of measuring the structural similarity at multiple levels . In the training process, the Adam optimizer is used, and the peak signal-to-noise ratio PSNR and the structural similarity SSIM are used as the core evaluation indicators to monitor and analyze the model performance in real time, thereby ensuring the best defogging effect and excellent generalization ability under different haze concentrations and diverse scenarios; In step S4, after training, further fine-tuning is carried out to optimize the network structure, thereby improving the overall visual quality while maintaining the defogging effect; the model that has been fully optimized and verified is deployed to the unmanned aerial vehicle platform to realize efficient and low-latency remote sensing defogging processing using the on-board computing power, thereby ensuring stable output of clear images under complex weather and variable environmental conditions and providing reliable and real-time technical support for practical application scenarios such as environmental monitoring, atmospheric quality assessment, and emergency response.

3. The method of claim 1, wherein, In step S3, the loss function used includes: Smooth L1 loss : Introduce smooth L1 loss function to effectively constrain network behavior; formula as follows: ； ； wherein, represents a remote sensing image defogging network of cross-layer cooperative interaction, the values of the first pixels in the clear image and the blurred image are represented by and respectively, and the total number of pixels of the image is ; adversarial loss Adversarial loss function in generative adversarial networks is adopted, in which the generator aims to convert the foggy images to clear defogged images, while the discriminator aims to distinguish the real clear images from the defogged images generated by the generator; under this adversarial learning mechanism, the generator constantly adjusts its generation process to make the generated images closer to the real images; the formula is defined as follows: ； perceptual loss : for realizing multi-level constraints on image from low-level texture to high-level semantics, optimizing the quality of defogged image; the specific formula is expressed as: ； in, This represents the blurred input image. This indicates a clear image of the target; This indicates the first [item] extracted from the VGG16 pre-trained model. Layer features; in the calculation of perceptual loss, This indicates the number of feature layers in the VGG16 pre-trained model used; Multi-scale structural similarity loss : The multi-scale structural similarity loss is introduced to further constrain the generation process of the network, which enhances the structural similarity between the generated image and the clear image, and ensures that the image quality is effectively improved. The calculation formula is as follows: ； where denotes the generated image, denotes the sharp image; the mean of the generated image and the sharp image are and the standard deviations are and and denotes the covariance between them; the two important terms in the equation are and and finally, the constant terms are and ; Total loss: the total loss function is represented as a weighted combination in the following form; the specific formula is: ； wherein the hyperparameters , , , represent weight coefficients of each loss term to balance the contribution of each loss term to the defogging effect.

4. The method of claim 1, wherein, In step S1, remote sensing haze images from different concentration conditions are collected to ensure data diversity; image cleaning and standardization and other preprocessing are performed to ensure data quality and consistency and to enhance the model's generalization ability in remote sensing images, thereby effectively supporting subsequent model training.

5. The method of claim 1, wherein, In step S2, the cross-layer collaborative interaction remote sensing image defogging network framework comprises a U-shaped defogging network and a cross-layer collaborative interaction module CLCIM; specifically, The output of the first encoder (200) is connected to the input of the second encoder (201); then, the outputs of the first encoder (200) and the second encoder (201) are respectively connected to the input of the first cross-layer collaborative interaction module (1); The output of the second encoder (201) is further transmitted to the input of the third encoder (202), and the output of the third encoder (202) is transmitted to the output of the first cross-layer collaborative interaction CLCIM module (1) together with the output of the first cross-layer collaborative interaction CLCIM module (1); The output of the second cross-layer collaborative interaction CLCIM module (2) is connected to the input of the residual block module (3), and the output of the residual block module (3) is transmitted to the input of the third decoder (203) together with the output of the third encoder (202) after residual fusion; The output of the third decoder (203) is transmitted to the input of the second decoder (204) together with the output of the second encoder (201) after residual fusion, and further decoding operation is performed; subsequently, the outputs of the third decoder (203) and the second encoder (204) are respectively connected to the input of the third cross-layer collaborative interaction CLCIM module (4); The output of the second decoder (204) is transmitted to the input of the first decoder (205) together with the result of the output of the first encoder (200) after residual fusion; at this time, the output of the first decoder (205) is respectively connected to the input of the fourth cross-layer collaborative interaction CLCIM module (5) together with the output of the third cross-layer collaborative interaction CLCIM module (4). The input end of the first encoder (200) is used for inputting a foggy image, and the output end of the fourth cross-layer collaborative interaction module (5) is used for outputting a clear image.

6. The method of claim 5, wherein, In the encoding stage, the input foggy image is extracted layer by layer through the first encoder (200), the second encoder (201) and the third encoder (202) to obtain multi-scale features. After entering the decoding stage, the network restores the spatial resolution of the image through step-by-step decoding; the third decoder (203), the second decoder (204) and the first decoder (205) perform feature fusion through residual connection with the third encoder (202), the second encoder (201) and the first encoder (200) respectively, thereby significantly improving the detail restoration capability of the image; finally, the network generates a clear image after dehazing; at the same time, the cross-layer collaborative interaction module, i.e. the CLCIM module, interacts with the features between layers in the encoding and decoding stages, ensuring that the encoder and the decoder can not only capture local details within the layer, but also extract global context information across layers.

7. The method of claim 6, wherein, Specifically, when constructing the U-shaped dehazing network, the input foggy image is extracted layer by layer through four times of down-sampling; the first encoder (200) is responsible for basic feature extraction, the second encoder (201) further enhances these features, and the third encoder (202) focuses on capturing deep semantic features of the image; the down-sampled feature map then extracts more semantic information through the residual block; in the decoding stage, the features extracted by the encoder are gradually restored to the spatial resolution of the image through up-sampling; the third decoder (203) is responsible for restoring relatively rough features and gradually refining the low-resolution information of the image, the second decoder (204) further improves the integrity of the information through feature fusion, and the first decoder (205) focuses on restoring the details of the image, and finally restores the image to the original size; through the residual connection, the output of the encoder is connected with the decoder layer, so that the features can be directly transmitted, thereby avoiding information loss and speeding up the training process of the network.

8. The method according to one of claims 5 to 7, characterized in that, The constructed cross-layer cooperative interaction CLCIM module includes a shallow branch network, a deep branch network and an inter-layer interaction network; For the shallow branch network: first, the initial shallow feature map (300) outputs the input connected to the shallow layer normalization LN block to enhance the training stability; the shallow layer normalization LN block feature map (301) is divided into a shallow local feature map and a shallow global feature map in a channel pair; the shallow local feature map is input to a shallow local first 3*3 deep convolution to obtain a shallow local first 3*3 deep convolution feature map (310), the shallow local first 3*3 deep convolution feature map (310) is input to a shallow local first 5*5 deep convolution to obtain a shallow local first 5*5 deep convolution feature map (311), the shallow local first 5*5 deep convolution feature map (311) is input to a shallow local first 7*7 deep convolution to obtain a shallow local first 7*7 deep convolution feature map (312); at the same time, the shallow global feature map is input to a shallow global first 3*3 deep convolution to obtain a shallow global first 3*3 deep convolution feature map (313), the shallow global first 3*3 deep convolution feature map (313) is input to a shallow global second 3*3 deep convolution to obtain a shallow global second 3*3 deep convolution feature map (314), the shallow global second 3*3 deep convolution feature map (314) is input to a shallow global Mamba to obtain a shallow global Mamba feature map (315); the process can be represented as: ； wherein, is a shallow local feature map, represents a deep convolution operation on the shallow local feature map, represents a cascaded deep convolution operation on the shallow local feature map, refers to a shallow local first 7*7 deep convolution feature map; is a shallow global feature map, represents a deep convolution operation on the shallow global feature map, represents a successive deep convolution operation on the shallow global feature map, represents processing a shallow global second 3*3 deep convolution feature map using a Mamba block, is a shallow global Mamba feature map; Subsequently, the shallow local first 7*7 deep convolution feature map (312) and the shallow global Mamba feature map (315) are channel spliced to obtain a shallow spliced feature map (320); the shallow spliced feature map (320) is input to a shallow first 1*1 convolution to obtain a shallow first 1*1 convolution feature map (321); this process can be represented as: ； wherein, is a concatenation operation, the concatenated feature uses to perform feature fusion, denotes the fused shallow first 1*1 convolution feature map, the size is restored to This strategy effectively balances detail preservation and global information modeling, while maintaining efficient computation and enhancing adaptability to complex visual information. Then, the shallow first 1*1 convolution feature map (321) is input to a shallow second 1*1 convolution to obtain a shallow second 1*1 feature map (322); the shallow first 1*1 convolution feature map (321) is input to a shallow third 1*1 convolution to obtain a shallow third 1*1 feature map (324); this process can be represented as: ； wherein, represents a two-dimensional convolution with both the kernel size and the stride being 1, respectively represent the feature extraction branch and the gating branch after convolution processing, , so ;​ Then, the shallow second 1*1 feature map (322) is input to a shallow first 5*5 deep convolution to obtain a shallow first 5*5 deep convolution feature map (323); the shallow second 1*1 feature map (322) and the shallow first 5*5 deep convolution feature map (323) are residual connected and multiplied by the shallow third 1*1 feature map (324) to obtain a first weighted feature map (325), the first weighted feature map (325) is output and residual connected with the shallow first 1*1 convolution feature map (321) to obtain a first gating feature map (326); this process can be represented as: ； wherein, represents a deep convolution with a 5x5 convolution kernel and a step size of 1, represents a feature map processed by a gating mechanism, represents the final output feature map; the gating feature enhancement block enhances key information and suppresses redundant information, maintains spatial structure integrity, optimizes pixel-level feature interaction, and provides richer feature representation.

9. The method of claim 8, wherein, For the deep branch network: first, the initial deep feature map (330) outputs the input connected to the deep layer normalization LN block to enhance the training stability; the deep layer normalization LN block feature map (331) is split into a deep local feature map and a deep global feature map in a channel pair; the deep local feature map is input to a deep local first 3*3 depth convolution to obtain a deep local first 3*3 depth convolution feature map (340), the deep local first 3*3 depth convolution feature map (340) is input to a deep local first 5*5 depth convolution to obtain a deep local first 5*5 depth convolution feature map (341), the deep local first 5*5 depth convolution feature map (341) is input to a deep local first 7*7 depth convolution to obtain a deep local first 7*7 depth convolution feature map (342); at the same time, the deep global feature map is input to a deep global first 3*3 depth convolution to obtain a deep global first 3*3 depth convolution feature map (343), the deep global first 3*3 depth convolution feature map (343) is input to a deep global second 3*3 depth convolution to obtain a deep global second 3*3 depth convolution feature map (344), the deep global second 3*3 depth convolution feature map (344) is input to a deep global Mamba to obtain a deep global Mamba feature map (345); this process can be represented as: ； wherein, is a deep local feature map, represents a deep convolution operation on the deep local feature map, represents a cascaded deep convolution operation on the deep local feature map, refers to a deep local first 7*7 deep convolution feature map; is a deep global feature map, represents a deep convolution operation on the deep global feature map, represents a successive deep convolution operation on the deep global feature map, represents processing a deep global second 3*3 deep convolution feature map using a Mamba block, is a deep global Mamba feature map; Subsequently, the deep local first 7*7 depth convolution feature map (342) and the deep global Mamba feature map (345) are channel spliced to obtain a deep splicing feature map (350); the deep splicing feature map (350) is input to a deep first 1*1 convolution to obtain a deep first 1*1 convolution feature map (351); this process can be represented as: ； wherein, is a concatenation operation, the concatenated feature uses to perform feature fusion, denotes the fused deep first 1*1 convolution feature map, the size is restored to This strategy effectively balances detail preservation and global information modeling, while maintaining efficient computation and enhancing adaptability to complex visual information. Then, the deep first 1*1 convolution feature map (351) is input to a deep second 1*1 convolution to obtain a deep second 1*1 feature map (352); the deep first 1*1 convolution feature map (351) is input to a deep third 1*1 convolution to obtain a deep third 1*1 feature map (354); this process can be represented as: ； wherein, represents a two-dimensional convolution with both the kernel size and the stride being 1, respectively represent the feature extraction branch and the gating branch after convolution processing, , so ;​ Then, the deep second 1*1 feature map (352) is input to a deep first 5*5 depth convolution to obtain a deep first 5*5 depth convolution feature map (353); the deep second 1*1 feature map (352) and the deep first 5*5 depth convolution feature map (353) are residual connected and multiplied by the deep third 1*1 feature map (354) to obtain a second weighted feature map (355), the second weighted feature map (355) is output and residual connected with the deep first 1*1 convolution feature map (351) to obtain a gated second feature map (356); this process can be represented as: ； wherein, represents a deep convolution with a 5x5 convolution kernel and a step of 1, represents a feature map processed by a gating mechanism, represents the final output feature map; the gating feature enhancement block enhances key information and suppresses redundant information, maintains spatial structure integrity, optimizes pixel-level feature interaction, and provides richer feature representation; Inter-layer interaction network: in order to perform inter-layer interaction and capture global context information, a reshaping first operation is used on the gated first feature map (326) to obtain a shallow sequence (327), and a reshaping second operation is used on the gated second feature map (356) to obtain a deep sequence (357); this process can be represented as: ； wherein, denotes a reshaping operation on the obtained gated feature map, represents a reshaping operation, from spatial domain to channel-wise representation, by and consists of, the size changes from H x W x C to ; Further, the shallow sequence (327) is subjected to a transposition operation to obtain a transposed sequence (328), the transposition operation changing the dimension of the transposed sequence (328) from to , and then subjected to a matrix multiplication operation with the deep sequence (357) having a shape of , and the result is input to Softmax to obtain the inter-layer weight (360) having a shape of CxC; subsequently, a matrix operation is performed on the shallow sequence (327) and the inter-layer weight (360); this process can be represented as: ； wherein, represents the calculated inter-layer weight, the Softmax enhances the relevant features and suppresses the irrelevant features, and generates more effective attention scores; fuses different levels of information to generate context-related features.

10. The method of claim 9, wherein, Finally, the matrix operation result of the shallow sequence (327) and the inter-layer weight (360) uses a reshaping third operation, and a residual connection is used to input a 1*1 convolution of the initial shallow feature map (300) and the initial deep feature map (330) to obtain a final feature map (361); which can be represented as: ； wherein, denotes an initial shallow feature map, denotes an initial deep feature map, denotes a final feature map; This operation fully integrates shallow and deep features and can fully learn global context information.

Images