A cloud detection method for remote sensing images based on multi-attention and multi-level feature hybrid

Abstract

The present invention relates to a remote sensing image cloud detection method based on multi-attention and multi-level feature mixing, which belongs to the field of satellite remote sensing image processing technology, including: constructing a cloud and fog detection data set through an existing data set or collecting real remote sensing satellite image data, and marking the cloud and fog areas therein; constructing a cloud and fog detection network model; training and verifying the cloud and fog detection network model; inputting the remote sensing image to be detected into the cloud and fog detection network model to obtain the cloud and fog mask area, and realizing the cloud detection task. The method of the present invention is simple and reasonably designed. The encoder module adopts an efficient self-attention mechanism and a channel attention mechanism to enhance the ability of the model network to extract image features, and mixes image features from different directions and angles respectively. The decoder module introduces multi-level MLP and upsampling, which fully improves the utilization ability of features at all levels, improves the detection ability of clouds and fogs of different scales, and improves the cloud detection accuracy.

Description

Technical Field

[0001] The invention relates to a remote sensing image cloud detection method based on multi-attention and multi-level feature mixing, and belongs to the technical field of satellite remote sensing image processing. Background Art

[0002] Driven by the progress of remote sensing technology in the past few decades, optical remote sensing images have been widely used. As a natural phenomenon in the atmosphere, clouds and fog have a significant impact on remote sensing satellite images. They reduce the contrast of images by scattering and absorbing light, resulting in blurred details, which in turn affects the clarity and usability of images. In addition, dense clouds and fog can obscure the true features of the surface, making it difficult for satellite sensors to fully capture ground information, which poses a major obstacle to application areas that require accurate surface information, such as geological exploration, urban planning, and agricultural monitoring. Clouds and fog also interfere with spectral reflectance in different bands, which makes it difficult to interpret remote sensing data and increases the complexity of data interpretation. In terms of meteorological monitoring, the presence of clouds and fog limits the ability of satellites to obtain key information such as atmospheric temperature and humidity, affecting the accuracy of meteorological monitoring.

[0003] To address these challenges, researchers have developed a variety of technologies to improve the quality of remote sensing images under foggy conditions. The application of multispectral technology reduces the impact of foggy conditions by combining data at different wavelengths, thereby improving the clarity of images. Synthetic aperture radar (SAR) technology provides an effective means for obtaining surface information due to its ability to penetrate foggy conditions. In addition, researchers have developed advanced fog correction algorithms to reduce the interference of foggy conditions on images. By optimizing satellite orbit planning, the interference of foggy conditions on satellite observations can be reduced. Data fusion technology improves the accuracy and reliability of data by combining satellite images with ground observation data, meteorological data, etc.

[0004] Cloud detection in remote sensing images is a key step in remote sensing image processing and is crucial for subsequent image analysis and applications. Cloud detection methods can be divided into three categories: spectral threshold-based methods, classical machine learning-based methods, and deep learning-based methods. Spectral threshold-based methods mainly use spectral information to distinguish cloud and non-cloud pixels by setting thresholds, but this method relies on various information and may have incorrect pixel classification affected by similar information.

[0005] Machine learning-based methods avoid manually setting thresholds by learning the characteristics of sample data, thereby improving accuracy and flexibility. Deep learning methods, especially those based on convolutional neural networks (CNNs), can automatically learn the characteristics of images and achieve high-precision cloud detection, but these methods may require a large amount of training data and parameter adjustment. At the same time, existing deep learning technologies are applied to cloud detection in remote sensing images. Compared with other methods, the accuracy of cloud detection is significantly improved, but the effect on the thin fog part at the edge of the cloud is poor, resulting in reduced accuracy.

[0006] Therefore, a cloud detection method with reasonable design and high cloud detection accuracy is needed. Summary of the invention

[0007] In order to address the shortcomings of the prior art, the present invention provides a remote sensing image cloud detection method based on multi-attention and multi-level feature mixing. The method is simple and reasonably designed. An efficient self-attention mechanism and a channel attention mechanism are adopted in the encoder module to enhance the ability of the model network to extract image features, and image features are mixed from different directions and angles. Multi-level MLP and upsampling are introduced in the decoder module to fully improve the utilization ability of features at all levels, improve the detection ability of clouds and fog of different scales, and improve the cloud detection accuracy.

[0008] The invention adopts the following technical solution:

[0009] A remote sensing image cloud detection method based on multi-attention and multi-level feature mixing includes the following steps:

[0010] (1) Construct a cloud and fog detection dataset by using existing datasets or collecting real remote sensing satellite image data and marking the cloud and fog areas therein;

[0011] (2) Constructing a cloud and fog detection network model;

[0012] (3) Train and verify the cloud and fog detection network model, and supervise the model training through the cross entropy loss between the generated results and the training labels;

[0013] (4) The model accuracy can be tested using the data from the training set and the validation set. Once the accuracy requirements are met, the remote sensing image to be detected is input into the cloud and fog detection network model to obtain the cloud and fog mask area and realize the cloud detection task.

[0014] Preferably, in step (1), the remote sensing image is subjected to sliding cropping, and the image containing clouds and fog is visually selected, and the clouds and fog are annotated point by point at the pixel level using an image annotation tool, such as labelme, V7, Labelbox or any other annotation tool, to outline the pixel edges and pixel features.

[0015] Preferably, in step (2), the cloud and fog detection network model includes an encoder and a decoder, the encoder is constructed based on the Transformer network model and is responsible for extracting deep features of the image to generate a high-dimensional feature map, and the decoder is responsible for feature fusion of the high-dimensional feature map and outputting the final result;

[0016] The encoder consists of an Overlap Patch Embeddings module and four Multi Attention Block modules. First, the image input Overlap Patch Embeddings module is divided into small blocks and embedded with features of specific dimensions. Then, the encoder extracts features step by step through four Multi Attention Block modules using a hierarchical structure. Each level reduces the spatial resolution through a specific mechanism while increasing the channel dimension to capture semantic information from local to global.

[0017] The decoder gradually recovers from small-scale feature maps to large-scale feature maps through step-by-step upsampling and feature fusion to complete high-resolution output. It includes 3 MLPUP2 Block modules and 1 MLP module. The decoder receives the multi-scale feature map output by the encoder, and gradually recovers the resolution and improves the feature expression capability through step-by-step feature fusion and upsampling.

[0018] Preferably, each Multi Attention Block module includes an SE Block module, an EfficientSelf-Attn module, a Mix-FFN and an Overlap Patch Merging module, wherein the SE Block module is a channel attention module, which has a dual-branch structure, wherein one branch first performs global average pooling on the image, then performs channel fusion through a convolution operation, and finally obtains the attention weight of the branch through a GeLU activation function; the other branch first performs global maximum pooling on the image, then performs channel fusion through a convolution operation, and finally obtains the attention weight of the branch through a GeLU activation function; the results of the two branches are summed and activated using a sigmoid function, and the attention weight obtained by activation is multiplied by the module input to obtain the result of the channel attention; the EfficientSelf-Attn module is a self-attention module, which is an efficient self-attention mechanism for capturing the dependency between different positions in the image; Mix-FFN is a hybrid feedforward neural network, and the Overlap Patch Merging module is used to aggregate overlapping patches.

[0019] Preferably, the mixed feedforward neural network Mix-FFN consists of a convolution operation and an activation function. First, a convolution operation with a convolution kernel of 1 is used to expand the feature map channel, and then a convolution operation with a convolution kernel of 3 is used to extract features. Then, a GELU activation function is entered for nonlinear mapping, and finally, a convolution layer is entered to restore the number of channels to the initial number.

[0020] Preferably, each MLPUP2 Block module is composed of an MLP and a 2x upsampling module connected in series, so as to achieve feature extraction of the feature map and a 2x amplification effect at the same time.

[0021] Preferably, the encoder forms feature maps of four stages through four Multi Attention Block modules, which are named layer1, layer2, layer3, and layer4 respectively. The shape of layer1 is (H / 4, W / 4, C1), the shape of layer2 is (H / 8, W / 8, C2), the shape of layer3 is (H / 16, W / 16, C3), and the shape of layer4 is (H / 32, W / 32, C4), where H represents the image height, W represents the image width, Ci represents the i-th Multi Attention Block, and i takes the value of 1, 2, 3, or 4;

[0022] In the decoder, the feature map of layer4 is first upsampled twice through the first MLPUP2 Block module to obtain (H / 16, W / 16, C), and the result is merged with the feature map of layer3 to obtain (H / 16, W / 16, 2C); the result (H / 16, W / 16, 2C) is upsampled twice through the second MLPUP2 Block module to obtain (H / 8, W / 8, 2C), and the result is merged with the feature map of layer2 to obtain (H / 8, W / 8, 3C); the result (H / 8, W / 8, 3C) is upsampled twice through the third MLPUP2 Block module to obtain (H / 4, W / 4, 3C), and the result is merged with the feature map of layer1 to obtain (H / 4, W / 4, 4C); then the result is subjected to an MLP for feature enhancement to obtain the final result.

[0023] In the present invention, the decoder fuses and upsamples the multi-level cloud features of different depths in the encoder with the feature maps of different scales in the decoder in a tandem manner. As the upsampling process in the decoder proceeds step by step, the level of feature fusion gradually increases, generating a higher level feature map, thereby delivering more accurate location information.

[0024] Preferably, during the training process of step (3), the loss value of the mask of the cloud area in the image output by the cloud detection network model and the mask of the actual cloud area is calculated through a loss function, and the network is repeatedly iteratively trained using the gradient descent method according to the loss value, so that the recognition effect is gradually corrected to obtain the final output result.

[0025] Preferably, during training, the initial learning rate is set to 0.001, and a total of 200 cycles are trained. The learning rate is reduced every 30 cycles, and the learning rate becomes one tenth of the original value. When the loss value remains below the threshold after 200 training cycles, the model is considered to have converged and no further training is performed.

[0026] Preferably, data enhancement is performed on real remote sensing satellite image data using random cropping, random magnification, random horizontal flipping, random vertical flipping, random noise, and random image distortion.

[0027] For any details not provided in the present invention, please refer to the prior art.

[0028] The beneficial effects of the present invention are:

[0029] 1. The present invention connects the self-attention module and the channel attention module in the cloud and fog detection network model in parallel, which not only retains the advantages of the self-attention mechanism in capturing long-distance dependencies and dynamic weight allocation, but also increases the advantage of channel attention in allocating different weights to each channel, thereby improving the performance of the model. This improvement enables the model to increase its perception of the importance of different channels without losing its original capabilities, and specifically allocates higher weights to the channels that contribute the most.

[0030] 2. The decoder of the present invention fuses and upsamples the multi-level fog features of different depths in the encoder with the feature maps of different scales in the decoder in a tandem manner. As the upsampling process in the decoder proceeds step by step, the level of feature fusion gradually increases, generating a higher level feature map, thereby delivering more accurate location information. The advantage of this method is that the feature map obtained in the decoder stage is more sufficient, and the deep-level features are subjected to multiple 2x upsampling and multiple feature fusions interspersed in the middle, so that the network can better learn the intrinsic connection between images, thereby improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0032] Figure 1 This is an architecture diagram of the remote sensing image cloud detection method based on multi-attention and multi-level feature mixing of the present invention;

[0033] Figure 2 It is the module structure diagram of Multi Attention Block of the present invention;

[0034] Figure 3 It is the module structure diagram of MLPUP2 Block of the present invention;

[0035] Figure 4 Image input for the cloud and fog detection network model;

[0036] Figure 5 The cloud mask image output by the cloud detection network model;

[0037] Figure 6 This is the cloud mask overlay output by the cloud detection network model. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the drawings in the implementation of this specification, but are not limited to this. Anything not fully described in the present invention shall be based on the conventional technology in the art.

[0039] Example 1

[0040] A remote sensing image cloud detection method based on multi-attention and multi-level feature mixing includes the following steps:

[0041] (1) Use existing data sets or collect real remote sensing satellite image data, and annotate the cloud and fog areas in them to build a cloud and fog detection data set, including training set, validation set, and test set;

[0042] (2) Constructing a cloud and fog detection network model:

[0043] (3) Train and verify the cloud and fog detection network model, and supervise the model training through the cross entropy loss between the generated results and the training labels;

[0044] (4) The model accuracy can be tested using the data from the training set and the validation set. Once the accuracy requirements are met, the remote sensing image to be detected is input into the cloud and fog detection network model to obtain the cloud and fog mask area and realize the cloud detection task.

[0045] The present invention can not only effectively improve the accuracy of traditional cloud detection, but also significantly enhance the detection efficiency.

[0046] Example 2

[0047] A remote sensing image cloud detection method based on multi-attention and multi-level feature mixing is as described in Example 1, except that, during specific implementation, the technology of the present invention can be experimented using the Python programming language, and can also be engineered using the C / C++ programming language.

[0048] In step (1), the open source dataset LandSet8 cloud detection dataset is selected as the training set and verification set of the cloud detection network model proposed in the present invention, which contains original remote sensing images and cloud annotation information. There are 64 large-scale remote sensing images with an image size of about 8000×8000. 44 images are used as the training set and 30 images are used as the verification set. During training, due to the limitation of GPU memory, the embodiment of the present invention divides the 8000×8000 images in the training set and the verification set into 352×352 images for model training and verification. For the test set, the present invention uses images obtained by the Sentinel-2 satellite for testing, which includes 14 large-scale remote sensing images with an image size of 10980×10980. During testing, the test image is divided into 1024×1024, and the divided test results are re-joined in the order of blocks to obtain the corresponding large-scale remote sensing image cloud mask.

[0049] Example 3

[0050] A remote sensing image cloud detection method based on multi-attention and multi-level feature mixing is as described in Example 2, except that in step (2), the cloud and fog detection network model includes an encoder and a decoder, the encoder is constructed based on the Transformer network model, and is responsible for extracting deep features of the image to generate a high-dimensional feature map, and the decoder is responsible for feature fusion of the high-dimensional feature map and outputting the final result; the Transformer structure can effectively extract feature information in the image, which is conducive to more accurate extraction of cloud areas.

[0051] The encoder consists of an Overlap Patch Embeddings module and four Multi Attention Block modules. First, the image input Overlap Patch Embeddings module is divided into small blocks and embedded with features of specific dimensions. Then, the encoder extracts features step by step through four Multi Attention Block modules using a hierarchical structure. Each level reduces the spatial resolution through a specific mechanism while increasing the channel dimension to capture semantic information from local to global.

[0052] The decoder gradually recovers from small-scale feature maps to large-scale feature maps through step-by-step upsampling and feature fusion to complete high-resolution output. It includes 3 MLPUP2 Block modules and 1 MLP module. The decoder receives the multi-scale feature map output by the encoder, and gradually recovers the resolution and improves the feature expression capability through step-by-step feature fusion and upsampling.

[0053] like Figure 2 As shown in the figure, each Multi Attention Block module includes an SE Block module, an EfficientSelf-Attn module, a Mix-FFN and an Overlap Patch Merging module, wherein the SE Block module is a channel attention module, which has a dual-branch structure. One branch first performs global average pooling on the image, then performs channel fusion through a convolution operation, and finally obtains the attention weight of the branch through a GeLU activation function; the other branch first performs global maximum pooling on the image, then performs channel fusion through a convolution operation, and finally obtains the attention weight of the branch through a GeLU activation function; the results of the two branches are summed and activated using a sigmoid function, and the attention weight obtained by activation is multiplied by the module input to obtain the result of the channel attention. The channel attention module can enhance attention to key areas and optimize feature expression. Finally, the encoder outputs multi-scale feature maps, which contain rich semantic information and spatial details, and are finally mapped to a high-dimensional feature space; the EfficientSelf-Attn module is a self-attention module, which is an efficient self-attention mechanism used to capture the dependencies between different positions in the image; Mix-FFN is a hybrid feedforward neural network, and the Overlap Patch Merging module reduces the spatial resolution by aggregating overlapping patches, while integrating local and global features to optimize the consistency and expressiveness of features. It is an important operation that takes into account both efficiency and effect.

[0054] The mixed feedforward neural network Mix-FFN consists of convolution operations and activation functions. First, a convolution operation with a convolution kernel of 1 is used to expand the feature map channel, and then a convolution operation with a convolution kernel of 3 is used for feature extraction. After that, the GELU activation function is entered for nonlinear mapping, and finally the convolution layer is entered to restore the number of channels to the initial number.

[0055] The Multi Attention Block module is used for feature extraction in the encoder. Each pass will reduce the scale of the feature map by two times. The module input enters the self-attention module and the channel attention module respectively. The results of the two modules are added and entered into the hybrid feedforward network for feature mixing. This operation can enhance the model's attention to the feature map channels that have made outstanding contributions on the basis of the original self-attention mechanism without affecting the effect of the self-attention mechanism. The number of modules can be adjusted according to the size of the required model.

[0056] like Figure 3 As shown, each MLPUP2 Block module is composed of an MLP and a 2x upsampling module connected in series, so as to extract features from the feature map and achieve a 2x amplification effect at the same time.

[0057] Furthermore, the encoder forms feature maps of four stages through four Multi Attention Block modules, named layer1, layer2, layer3, and layer4 respectively. The shape of layer1 is (H / 4, W / 4, C1), the shape of layer2 is (H / 8, W / 8, C2), the shape of layer3 is (H / 16, W / 16, C3), and the shape of layer4 is (H / 32, W / 32, C4), where H represents the image height, W represents the image width, Ci represents the i-th Multi Attention Block, and i takes the value of 1, 2, 3, or 4.

[0058] In the decoder, the feature map of layer4 is first upsampled twice through the first MLPUP2 Block module to obtain (H / 16, W / 16, C), and the result is merged with the feature map of layer3 to obtain (H / 16, W / 16, 2C); the result (H / 16, W / 16, 2C) is upsampled twice through the second MLPUP2 Block module to obtain (H / 8, W / 8, 2C), and the result is merged with the feature map of layer2 to obtain (H / 8, W / 8, 3C); the result (H / 8, W / 8, 3C) is upsampled twice through the third MLPUP2 Block module to obtain (H / 4, W / 4, 3C), and the result is merged with the feature map of layer1 to obtain (H / 4, W / 4, 4C); then the result is subjected to an MLP for feature enhancement to obtain the final result. Compared with the method of uniformly amplifying different stages to the same multiple and then merging them, this method obtains a more complete feature map in the decoder stage. The deep features are upsampled by 2 times multiple times and multiple feature fusions are interspersed in the middle, so that the network can better learn the intrinsic connection between images, thereby improving the recognition accuracy.

[0059] The reason for constructing the MLPUP2 Block module is to fuse the features from different stages of the encoder according to their feature map sizes; in the decoder, the feature maps of each stage are gradually processed and channel-adjusted to form a consistent feature representation.

[0060] The role of MLP is to linearly transform the channel dimension of the feature map while enhancing its semantic expression ability. First, the spatial dimension of the feature map is flattened to form a set of channel feature vectors, which are linearly transformed through the fully connected layer of MLP to extract higher-dimensional semantic information. Subsequently, the activation function (such as ReLU or GELU) introduces nonlinear capabilities to further enhance feature expression. Through multiple layers of linear transformation and nonlinear operations, MLP not only optimizes the feature distribution, but also adjusts the channel dimension, ultimately generating a more recognizable feature map to provide support for subsequent upsampling and feature fusion of the decoder.

[0061] In the present invention, the decoder fuses and upsamples the multi-level cloud features of different depths in the encoder with the feature maps of different scales in the decoder in a tandem manner. As the upsampling process in the decoder proceeds step by step, the level of feature fusion gradually increases, generating a higher level feature map, thereby delivering more accurate location information.

[0062] Example 4

[0063] A remote sensing image cloud detection method based on multi-attention and multi-level feature mixing is as described in Example 3, except that, during the training process of step (3), the mask of the cloud area in the image output by the cloud detection network model and the mask of the real cloud area are used to calculate the loss value through the loss function, and the network is repeatedly iteratively trained using the gradient descent method according to the loss value, so that the recognition effect is gradually corrected to obtain the final output result.

[0064] During training, the initial learning rate is set to 0.001, and a total of 200 cycles are trained. The learning rate is reduced every 30 cycles to one tenth of the original value. The training is performed on the training set and the changes in the loss function are recorded. When the loss value remains below the threshold (such as 0.005) after 200 training cycles, the model is considered to have converged and no further training is performed.

[0065] After each training of the cloud and fog detection network model, the accuracy of the model is verified on the validation set, and the error between the estimated value and the true value is calculated. The smaller the error, the more accurate the model prediction.

[0066] In specific implementation, it is recommended to use the open source deep learning framework Pytorch for model training when training the cloud and fog detection network model, and use GPU to accelerate the training of the model.

[0067] Since the image scale of the test set is 10980×10980, due to the limitation of GPU memory, the image needs to be loaded in blocks, each block is 1024×1024. Each time a block of image is loaded, the cloud mask of the image is automatically generated until the last block of image is loaded into the model, and the cloud mask of the entire image is obtained by splicing the cloud mask of each block of image.

[0068] In order to facilitate understanding of the technical effects of the present invention, the following are provided: an input image of the cloud and fog detection network model of the present invention, an output cloud and fog area mask image, and an overlay image of the cloud and fog area and the original remote sensing image. Figure 4 , is the original image. Figure 5 This is the mask image of the cloud area. Figure 6 This is the overlay of the cloud area and the original image. It can be seen that the cloud area mask output by the cloud detection model is basically consistent with the cloud area of ​​the original image.

[0069] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A remote sensing image cloud detection method based on multi-attention and multi-level feature mixing, characterized in that: The steps include: (1) Construct a cloud and fog detection dataset by using existing datasets or collecting real remote sensing satellite image data and marking the cloud and fog areas therein; (2) Constructing a cloud and fog detection network model; (3) Train and verify the cloud and fog detection network model, and supervise the model training through the cross entropy loss between the generated results and the training labels; (4) The remote sensing image to be detected is input into the cloud detection network model to obtain the cloud mask area and realize the cloud detection task.

2. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 1 is characterized in that: In step (1), the remote sensing image is subjected to sliding cropping, and the image containing clouds and fog is visually selected. The cloud and fog are annotated point by point at the pixel level using an image annotation tool to outline the pixel edges and pixel features.

3. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 2 is characterized in that: In step (2), the cloud and fog detection network model includes an encoder and a decoder. The encoder is constructed based on the Transformer network model and is responsible for extracting deep features of the image to generate a high-dimensional feature map. The decoder is responsible for feature fusion of the high-dimensional feature map and outputting the final result. The encoder consists of an Overlap Patch Embeddings module and four Multi Attention Block modules. First, the image input Overlap Patch Embeddings module is divided into small blocks and embedded with features of specific dimensions. Then, the encoder uses a hierarchical structure through four Multi Attention Block modules to extract features step by step. Each level reduces the spatial resolution through a specific mechanism, while increasing the channel dimension, capturing semantic information from local to global. The decoder gradually recovers from small-scale feature maps to large-scale feature maps through step-by-step upsampling and feature fusion to complete high-resolution output. It includes 3 MLPUP2 Block modules and 1 MLP module. The decoder receives the multi-scale feature map output by the encoder, and gradually recovers the resolution and improves the feature expression capability through step-by-step feature fusion and upsampling.

4. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 3 is characterized in that: Each Multi Attention Block module includes SE Block module, EfficientSelf-Attn module, Mix-FFN and Overlap Patch Merging module, among which SE Block module is a channel attention module, EfficientSelf-Attn module is a self-attention module, Mix-FFN is a hybrid feedforward neural network, and Overlap PatchMerging module is used to aggregate overlapping patches.

5. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 4 is characterized in that: The mixed feedforward neural network Mix-FFN consists of convolution operations and activation functions. First, a convolution operation with a convolution kernel of 1 is used to expand the feature map channel, and then a convolution operation with a convolution kernel of 3 is used for feature extraction. After that, the GELU activation function is entered for nonlinear mapping, and finally the convolution layer is entered to restore the number of channels to the initial number.

6. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 5 is characterized in that: Each MLPUP2 Block module consists of an MLP and a 2x upsampling module in series, which can extract features from the feature map and achieve a 2x amplification effect at the same time.

7. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 6 is characterized in that: The encoder forms feature maps of four stages through four Multi Attention Block modules, named layer1, layer2, layer3, and layer4 respectively. The shape of layer1 is (H / 4, W / 4, C1), the shape of layer2 is (H / 8, W / 8, C2), the shape of layer3 is (H / 16, W / 16, C3), and the shape of layer4 is (H / 32, W / 32, C4), where H represents the image height, W represents the image width, Ci represents the i-th MultiAttention Block, and i takes the value of 1, 2, 3, or 4. In the decoder, the feature map of layer4 is first upsampled twice through the first MLPUP2 Block module to obtain (H / 16, W / 16, C), and the result is merged with the feature map of layer3 to obtain (H / 16, W / 16, 2C); the result (H / 16, W / 16, 2C) is upsampled twice through the second MLPUP2 Block module to obtain (H / 8, W / 8, 2C), and the result is merged with the feature map of layer2 to obtain (H / 8, W / 8, 3C); the result (H / 8, W / 8, 3C) is upsampled twice through the third MLPUP2 Block module to obtain (H / 4, W / 4, 3C), and the result is merged with the feature map of layer1 to obtain (H / 4, W / 4, 4C); then the result is subjected to an MLP for feature enhancement to obtain the final result.

8. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 7 is characterized in that: During the training process of step (3), the loss value of the cloud area mask in the image output by the cloud detection network model and the real cloud area mask is calculated through the loss function, and the network is repeatedly iterated and trained using the gradient descent method according to the loss value, so that the recognition effect is gradually corrected to obtain the final output result.

9. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 8 is characterized in that: During training, the initial learning rate was set to 0.001, and a total of 200 cycles were trained. The learning rate was reduced every 30 cycles, and the learning rate became one tenth of the original value. When the loss value remained below the threshold after 200 training cycles, the model was considered to have converged and no further training was carried out.

10. The remote sensing image cloud detection method based on multi-attention and multi-level feature mixing according to claim 9 is characterized in that: The real remote sensing satellite image data is enhanced by random cropping, random magnification, random horizontal flipping, random vertical flipping, random noise and random image distortion.

Images