Sentinel-2 remote sensing image super-resolution reconstruction method based on attention mechanism

Abstract

The invention discloses a Sentinel-2 remote sensing image super-resolution reconstruction method based on an attention mechanism. The method comprises the following steps: 1, obtaining a Sentinel-2 remote sensing image data set; 2, carrying out the preprocessing of the Sentinel-2 remote sensing image data; 3, randomly dividing the preprocessed Sentinel-2 remote sensing image data into a training set and a test set; 4, constructing a deep learning model structure of the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism; 5, training the built model by using the training set; and 6, testing the test set by using the trained model to obtain a Sentinel-2 remote sensing image super-division image. According to the method, the characteristics of the Sentinel-2 satellite image are fully utilized, and the remote sensing image fusion thought is combined, so that the quality of the super-resolution image is improved, and the accuracy and reliability of the earth surface information are enhanced.

Description

Technical Field

[0001] The present invention relates to a remote sensing image super-resolution reconstruction method, and in particular to a Sentinel-2 remote sensing image super-resolution reconstruction method based on an attention mechanism. Background Art

[0002] Remote sensing satellite images play an important role in forestry applications. They not only provide efficient and accurate technical means for the investigation and monitoring of forest resources, but also play a key role in forest health assessment, carbon storage research and vegetation coverage extraction. However, the spatial resolution of remote sensing images is often limited by factors such as sensor technology, observation height and cost, and it is difficult to meet the needs of many application scenarios. With the increasing demand for refined analysis of remote sensing images, improving the spatial resolution of remote sensing images has become an important research topic.

[0003] Sentinel-2 is an earth observation satellite developed by the European Space Agency (ESA). It carries a multispectral imager (MSI) and consists of two identical satellites, 2A and 2B. Sentinel-2 remote sensing images are an important type of remote sensing images. They have the characteristics of multispectral imaging capabilities, high spatial resolution, large coverage, and free access. They provide rich spectral information and fine surface details, and can be used in forest loss change monitoring, chlorophyll content estimation, etc. Sentinel-2 remote sensing images include 13 spectral bands with three different spatial resolutions: 10m, 20m, and 60m. Since the spatial resolution of some bands in Sentinel-2 remote sensing images is low, only 20m, this will affect the capture of image details and refined analysis, limiting its effectiveness in some applications. Therefore, it is of great significance to improve the spatial resolution of these bands.

[0004] Image super-resolution (SR) is a highly-regarded computer vision task that aims to convert low-resolution (LR) images into high-resolution (HR) images with richer details and higher visual quality. With the continuous development of deep learning technology, significant progress has been made in achieving image super-resolution using deep learning technology. Many super-resolution reconstruction methods based on convolutional neural networks have been widely used and achieved remarkable results.

[0005] In recent years, learning-based methods have attracted much attention. In 2010, Yang et al. first used sparse representation in the field of super-resolution reconstruction. The algorithm combined sparse coding theory to construct high-resolution and low-resolution image dictionaries respectively, and reconstructed high-resolution images by learning the mapping relationship between dictionaries. With the continuous development of deep learning technology, many super-resolution reconstruction methods based on convolutional neural networks have been widely used and achieved remarkable results. SRCNN first applied convolutional neural networks to image super-resolution tasks, using a three-layer convolution structure to map low-resolution features to high-resolution features, so that the final reconstructed high-resolution image has high quality. In order to learn the multi-level representation of remote sensing images, including local details and global environment priors, Lei et al. designed a "branch" structure network (Local-Global Combined Network, LGCnet) to learn the multi-scale representation of remote sensing images. Ledig et al. first applied GAN to the field of image SR and proposed the SRGAN network model, combining perceptual loss and adversarial loss to train the network through generative adversarial methods. Considering that the Sentinel-2 satellite has both high-resolution and low-resolution band images, the idea of ​​remote sensing image fusion can be combined to improve the resolution of Sentinel-2 images. Panchromatic sharpening, as one of the important technologies for multi-source remote sensing image fusion, achieves the goal of generating multispectral images with high spatial resolution by fusing low-spatial-resolution multispectral images with high-spatial-resolution single-band panchromatic images. In a comparative study, Vaiopoulos and Karantzalos evaluated 21 panchromatic sharpening algorithms to enhance the 20m visible and near-infrared (VNIR) and short-wave infrared (SWIR) bands of Sentinel-2. Palsson et al. proposed a method to fuse fine spatial resolution bands with coarse spatial resolution bands to obtain finer spatial resolution of the coarse bands.

[0006] Both convolutional neural networks (CNNs) and other algorithms have certain limitations. Sparse representation methods lose a lot of information in the process of degradation processing of high-resolution remote sensing images. Methods based on convolutional neural networks have fewer layers and cannot fit the relationship between high and low resolution in remote sensing images well. The SRGAN network model trains the network by generating adversarial methods, but the objective evaluation indicators such as PSNR and SSIM of this method are relatively low. Summary of the invention

[0007] The purpose of the present invention is to provide a Sentinel-2 remote sensing image super-resolution reconstruction method based on the attention mechanism. In the remote sensing image super-resolution process, the method combines the characteristics of Sentinel-2 satellite images to perform network design, thereby improving the remote sensing image resolution.

[0008] The objective of the present invention is achieved through the following technical solutions:

[0009] A Sentinel-2 remote sensing image super-resolution reconstruction method based on an attention mechanism comprises the following steps:

[0010] Step 1: Get the Sentinel-2 remote sensing image dataset. According to the different spatial resolutions, the Sentinel-2 remote sensing image bands are divided into three categories: 10m, 20m and 60m. Each band has only one spatial resolution.

[0011] Step 2: Preprocess the Sentinel-2 remote sensing image data, downsample the original spatial resolution bands of 10m and 20m by two times, and use the satellite image with the original resolution of 20m band as the ground truth data reference. The process is shown in the following formula:

[0012] X 10 ↓2→X' 10

[0013] X 20 ↓2→X' 20

[0014] T(X' 10 ,X' 20 )→Y 20

[0015] Among them, X 10 Represents the original 10m band image, X 20 represents the original 20m band image, ↓2 represents downsampling by two times, X' 10 represents the downsampled 10m band image, X' 20 represents the downsampled 20m band image, T(·) represents the super-resolution architecture, and Y 20 It represents the 20m band image after super resolution;

[0016] Step 3: Randomly divide the preprocessed Sentinel-2 remote sensing image data into training set and test set;

[0017] Step 4: Build a deep learning model structure for the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism:

[0018] The deep learning model of the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism consists of a spatial feature extraction branch, a channel feature extraction branch and a trunk, wherein:

[0019] The spatial feature extraction branch extracts y by using the spatial attention mechanism AContains spatial information, y A It is a high-resolution image obtained by downsampling the original 10m band image by two times;

[0020] The channel feature extraction branch extracts y' by using the channel attention mechanism B Contains channel information, y B The original 20m band image is downsampled twice to obtain a low-resolution image. Bilinear interpolation is used to upsample it twice to convert it into a low-resolution image with the same resolution as y. A An image y' with the same spatial dimensions B ;

[0021] After the features extracted by the spatial feature extraction branch and the channel feature extraction branch are fused through the trunk, feature extraction is performed again using convolution and residual scaling mechanisms, and finally the final result is obtained through the residual structure;

[0022] The spatial attention mechanism performs global average pooling and global maximum pooling on the high-resolution input to obtain an effective feature descriptor, which is then cascaded according to the channel dimension and sent to a standard convolutional layer to obtain a spatial attention map through a Sigmoid function. Finally, the original input is multiplied by the spatial attention map to rescale the original input.

[0023] The computation of the spatial attention map is expressed as:

[0024] G Avg =AvgPool(G)

[0025] G Max =MaxPool(G)

[0026] G s =σ(ω(G Avg +G Max ))

[0027] in, G is the input feature map, G Avg is the feature map after global average pooling, G Max is the feature map after global maximum pooling: G∈R C×H×W , G Avg ∈R 1×H×W , G Max ∈R 1×H×W , ω represents the convolution operation with a filter size of 3×3, σ represents the Sigmoid function, and C, H, and W represent the number of bands, length, and width of the remote sensing image used, respectively;

[0028] The channel attention mechanism converts the global spatial information of the channel direction into a channel feature descriptor by using global average pooling, and then obtains the weight coefficient of each channel through a convolution activation layer, and multiplies the weight coefficient with the original feature to obtain a scaled channel attention map, so as to re-weight the features of different channels, and finally uses the residual to add the original input and the extracted channel attention map;

[0029] The calculation of the channel attention map is expressed as:

[0030] F c =F c-1 +R(F)·F

[0031]

[0032] Among them, F c-1 and F are the input and output of the entire large residual block, F Avg refers to the feature map after global average pooling, R(·) represents channel attention, ω1 represents the convolution operation with a filter size of 3×3, ω2 represents the convolution operation with a filter size of 1×1, represents the ReLU activation function, σ represents the Sigmoid function;

[0033] Step 5: Use the training set to train the built model;

[0034] Step 6: Use the trained model to test the test set to obtain the super-resolution image of Sentinel-2 remote sensing image.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] 1. Full utilization of spatial information and spectral information: Traditional methods usually only focus on improving spatial resolution and may ignore the effective utilization of spectral information. The present invention extracts spatial feature information from high-resolution images through a spatial attention mechanism and extracts spectral feature information from low-resolution images through a channel attention mechanism. It can simultaneously retain and utilize spatial and spectral information in the image and improve the overall quality of the image.

[0037] 2. Multi-resolution feature fusion: The features extracted from high-resolution and low-resolution images are fused and secondary feature extraction is performed, which comprehensively utilizes the advantages of images with different resolutions and improves the effect of super-resolution reconstruction.

[0038] 3. Strong targeting: It is specially designed according to the characteristics of Sentinel-2 satellite images, which can better adapt to and give full play to the advantages of the satellite images and improve the quality of super-resolution images.

[0039] 4. Super-resolution quality: Since the present invention makes full use of the characteristics of Sentinel-2 satellite images and extracts high-resolution and low-resolution image information respectively, higher-quality super-resolution images can often be obtained in remote sensing image super-resolution tasks.

[0040] 5. Application advantages: In remote sensing image super-resolution, the present invention can effectively improve the spatial resolution of Sentinel-2 remote sensing image bands, improve the application accuracy and reliability of image data in agricultural monitoring, environmental protection, urban planning and other fields, and promote research and application development in related fields.

[0041] 6. The present invention makes full use of the characteristics of Sentinel-2 satellite images and combines the idea of ​​remote sensing image fusion to improve the quality of super-resolution images and enhance the accuracy and reliability of surface information. Compared with traditional prediction methods, it has obvious advantages in full utilization of spatial information and spectral information, multi-resolution feature fusion and improvement of remote sensing image quality, providing data support for the subsequent application of Sentinel-2 satellite images in various fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a structural diagram of the deep learning model of the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism;

[0043] Figure 2 This is the structure diagram of the spatial attention network;

[0044] Figure 3 This is the channel attention network structure diagram;

[0045] Figure 4 This is a flow chart of the Sentinel-2 remote sensing image super-resolution reconstruction method based on the attention mechanism. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0047] The present invention provides a super-resolution reconstruction method of Sentinel-2 remote sensing images based on an attention mechanism. The method realizes super-resolution reconstruction of remote sensing images through a deep learning framework Pytorch and a corresponding data processing module. The specific implementation principle is as follows: Aiming at the characteristics of Sentinel-2 satellite images themselves, combined with remote sensing image fusion and deep learning methods, a deep learning model of a Sentinel-2 remote sensing image super-resolution method based on an attention mechanism is proposed. For low spatial resolution bands, due to their rich spectral information, feature extraction is performed through a channel attention mechanism; for high spatial resolution bands, due to their rich spatial information, feature extraction is performed through a spatial attention mechanism; after feature fusion of bands with different resolutions, secondary feature extraction is performed through convolution and residual structure to obtain the final result. Finally, the spatial resolution of the band corresponding to the 20m resolution of the Sentinel-2 satellite is improved to reach a spatial resolution of 10m, providing data support for subsequent applications such as precise investigation and monitoring of forest resources. The present invention solves the following technical problems:

[0048] 1. Aiming at the spatial resolution limitation problem faced by existing satellite remote sensing images in refined analysis, a deep learning-based image super-resolution method is designed. This method uses spatial attention mechanism and channel attention mechanism to extract spatial and channel features, successfully increasing the 20m resolution band to 10m, significantly expanding its application potential in forestry resource survey and monitoring, while meeting the needs of refined analysis of forestry resources.

[0049] 2. In traditional technical applications, high-resolution images and low-resolution images are often directly cascaded, which makes it difficult to fully utilize image features. The present invention proposes to extract features separately before directly cascading high-resolution and low-resolution images, and use spatial and channel attention mechanisms to apply them to high-resolution and low-resolution band images in a dual-branch structure to more effectively extract features, thereby improving the image super-resolution quality.

[0050] 3. In order to address the problem that images generated by existing traditional methods such as bilinear interpolation and bicubic interpolation have obvious edge effects and cannot effectively restore high-frequency details, the present invention introduces super-resolution reconstruction based on deep learning and proposes a remote sensing image super-resolution method based on an attention mechanism, so that the final reconstructed high-resolution image has higher quality.

[0051] (I) Deep learning model of Sentinel-2 remote sensing image super-resolution method based on attention mechanism

[0052] The deep learning model of the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism proposed in this invention is as follows Figure 1As shown in Figure 1, the model consists of two branches and a trunk. The model receives two parts of input, y A The high-resolution image is obtained by downsampling the original 10m band image by two times. The spatial feature extraction branch uses the spatial attention mechanism to focus on extracting y A Contains spatial information; y B The original 20m band image is downsampled twice to obtain a low-resolution image. Bilinear interpolation is used to upsample it twice to convert it into a low-resolution image with the same resolution as y. A An image y' with the same spatial dimensions B , the channel feature extraction branch focuses on extracting y' by using the channel attention mechanism B The features extracted by the two branches are fused through the trunk, and then the features are extracted again using convolution and residual scaling mechanisms, and the final result is obtained through the residual structure. The convolution layer before entering the trunk performs shallow feature extraction and maintains dimensional consistency.

[0053] The present invention adopts a dual-branch structure, combined with spatial attention and channel attention mechanisms, so that the model can make full use of the characteristics of Sentinel-2 remote sensing images and further extract features from the original remote sensing images. At the same time, the backbone structure performs secondary feature extraction on the results of dual-branch fusion to improve the quality of super-resolution reconstruction.

[0054] 2. Spatial Attention Mechanism

[0055] The role of the spatial attention mechanism: For high spatial resolution bands, due to their rich spatial information, feature extraction through the spatial attention mechanism can effectively focus on and enhance important spatial details in the image. This not only improves the clarity of image details, but also enhances the expressiveness of features, thereby achieving more accurate and higher quality reconstruction effects.

[0056] Since Sentinel-2 satellite images have four high-resolution bands and rich spatial information, by extracting features from their spatial information, we can further obtain richer features and improve fusion accuracy. The present invention uses a spatial attention mechanism to perform global average pooling and global maximum pooling on high-resolution inputs to obtain effective feature descriptors, which are then cascaded according to the channel dimension and sent to a standard convolutional layer. The spatial attention map is obtained through the Sigmoid function, and finally the original input is multiplied by the spatial attention map to rescale the original input.

[0057] The overall structure of the spatial attention module is as follows Figure 2 As shown. The calculation of the spatial attention map can be expressed as:

[0058] G Avg =AvgPool(G)

[0059] G Max =MaxPool(G)

[0060] G s =σ(ω(G Avg +G Max ))

[0061] in, G is the input feature map, G Avg is the feature map after global average pooling, G Max is the feature map after global maximum pooling: G∈R C×H×W , G Avg ∈R 1×H×W , G Max ∈R 1×H×W , ω represents the convolution operation with a filter size of 3×3, σ represents the Sigmoid function, and C, H, and W represent the number of bands, length, and width of the remote sensing image used, respectively.

[0062] 3. Channel Attention Mechanism

[0063] The role of the channel attention mechanism: For low spatial resolution bands, due to their rich spectral information, feature extraction through the channel attention mechanism can effectively emphasize the correlation and importance between bands. The channel attention mechanism assigns different weights to different spectral channels, thereby improving the accuracy and effect of feature extraction. It plays an important role in enhancing spectral features and improving the overall quality of images, and helps to more accurately restore and reconstruct surface information in remote sensing images.

[0064] The goal of the present invention is to super-resolve low-resolution bands to high resolution. The original low-resolution bands contain rich spectral information, which is treated equally between channels, thus hindering the representation ability of CNN. The interdependence between feature channels can be used, combined with the residual idea, to generate a residual channel attention mechanism. The global spatial information in the channel direction is converted into a channel feature descriptor by using global average pooling, and then a convolution activation layer is used to obtain the weight coefficient of each channel. The weight coefficient is multiplied by the original feature to obtain a scaled channel attention map, so as to realize the re-weighted distribution of features of different channels. Finally, the residual is used to add the original input and the extracted channel attention map.

[0065] The overall structure of the channel attention module is as follows Figure 3 As shown. The calculation of the channel attention map can be expressed as:

[0066] F c =F c-1 +R(F)·F

[0067]

[0068] Among them, F c-1 and F are the input and output of the entire large residual block, F Avg refers to the feature map after global average pooling, R(·) represents channel attention, ω1 represents the convolution operation with a filter size of 3×3, ω2 represents the convolution operation with a filter size of 1×1, represents the ReLU activation function, and σ represents the Sigmoid function.

[0069] (IV) Sentinel-2 remote sensing image super-resolution reconstruction method based on attention mechanism

[0070] The present invention adopts a comprehensive method based on the attention mechanism and uses a dual-branch structure to complete feature extraction of high- and low-resolution images, thereby realizing super-resolution reconstruction of Sentinel-2 remote sensing satellite images. Figure 4 As shown, the specific steps are as follows:

[0071] The first step is to obtain a remote sensing image dataset for training and testing. The training and test images used in this invention come from the Level-1C products of Sentinel-2A and Sentinel-2B satellites, and each image spans 110×110km. 2 area. In order to ensure the generalization and robustness of the model, three images from different regions and collected at different times are selected for training, and the image size is 10980×10980; for the test images, data from three regions of Canada, France and Russia are selected and cropped into image blocks of size 512×512 for testing. They show different characteristics, so that the model proposed in this invention can be more comprehensively verified. According to the different spatial resolutions, the Sentinel-2 remote sensing image bands can be divided into three categories: 10m (B2, B3, B4 and B8), 20m (B5, B6, B7, B8a, B11 and B12) and 60m (B1, B9 and B10). For each band, there is only one unique spatial resolution.

[0072] The second is the preprocessing of remote sensing image data. The goal of this invention is to increase the resolution of the six bands with a spatial resolution of 20m to 10m. However, due to technical limitations and other reasons, there is a lack of ground truth data reference with a resolution of 10m. Therefore, the training data set is obtained by downsampling the bands with the original spatial resolution of 10m and 20m by two times, and the satellite image with the original resolution of 20m band is used as the ground truth data reference. The process is shown in the following formula:

[0073] X 10 ↓2→X'10

[0074] X 20 ↓2→X' 20

[0075] T(X' 10 ,X' 20 )→Y 20

[0076] Among them, X 10 Represents the original 10m band image, X 20 represents the original 20m band image, ↓2 represents downsampling by two times, X' 10 represents the downsampled 10m band image, X' 20 represents the downsampled 20m band image, T(·) represents the super-resolution architecture, and Y 20 represents the 20m band image after super resolution. In addition, since the Sentinel-2 image is too large to be directly input into the neural network, 8000 image pairs are sampled for each training image, and each image block is 32*32, with a total of 24000 image blocks. In order to ensure numerical stability, the original reflectance value is divided by 2000 before processing. 90% of the 24000 image blocks after cropping are used for training, and the remaining 10% are used for validation.

[0077] Subsequently, a deep learning model structure of the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism was built. This method uses a dual-branch structure to complete the feature extraction of high- and low-resolution images. The spatial feature extraction branch uses the spatial attention mechanism to complete the spatial information extraction of the high-resolution band; the channel feature extraction branch uses the channel attention mechanism to complete the spectral information extraction of the low-resolution band. Finally, the result of the dual-branch fusion is subjected to secondary feature extraction through convolution and residual structure to complete super-resolution reconstruction. According to the above process, the corresponding model structure is built.

[0078] Next, the built model is trained using the prepared training data. The deep learning framework used in the present invention is Pytorch, the gradient optimization algorithm uses the adaptive moment estimation Adam optimizer, and the initial learning rate is set to lr = 10 -4 , the loss function uses the mean absolute pixel error between the real and predicted high-resolution images, that is, the L1 loss function.

[0079] Finally, the trained model is compared with other super-resolution models. The data evaluation indicators in the experiment are root mean square error (RMSE), structural similarity (SSIM) and signal to reconstruction error ratio (SRE). RMSE is a commonly used image similarity measurement method. The smaller its value, the more similar the image generated by the algorithm is to the original image, and the better the performance; SSIM is used to evaluate the similarity between the reconstructed image and the original image. The closer its value is to 1, the smaller the gap between the reconstructed image and the original image; SRE is used to measure the relative signal power error. The larger its value, the higher the match between the image generated by the algorithm and the original image. The comparison results are shown in Table 1.

[0080] Table 1 Quantitative indicators of reconstruction results of different algorithms on the Sentinel-2 dataset

[0081]

[0082] This method integrates data processing and deep learning models, providing a comprehensive and effective solution for super-resolution reconstruction of Sentinel-2 satellite remote sensing images.

Claims

1. A super-resolution reconstruction method for Sentinel-2 remote sensing images based on attention mechanism, characterized in that The method comprises the following steps: Step 1: Get the Sentinel-2 remote sensing image dataset. According to the different spatial resolutions, the Sentinel-2 remote sensing image bands are divided into three categories: 10m, 20m and 60m. Each band has only one spatial resolution. Step 2: Preprocess the Sentinel-2 remote sensing image data, downsample the original spatial resolution bands of 10m and 20m by two times, and use the satellite image with the original resolution of 20m band as the ground truth data reference. The process is shown in the following formula: X 10 ↓2→X' 10 X 20 ↓2→X' 20 T(X' 10 ,X' 20 )→Y 20 Among them, X 10 Represents the original 10m band image, X 20 represents the original 20m band image, ↓2 represents downsampling by two times, X' 10 represents the downsampled 10m band image, X' 20 represents the downsampled 20m band image, T(·) represents the super-resolution architecture, and Y 20 It represents the 20m band image after super resolution; Step 3: Randomly divide the preprocessed Sentinel-2 remote sensing image data into training set and test set; Step 4: Build a deep learning model structure for the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism: The deep learning model of the Sentinel-2 remote sensing image super-resolution method based on the attention mechanism consists of a spatial feature extraction branch, a channel feature extraction branch and a trunk, wherein: The spatial feature extraction branch extracts y by using the spatial attention mechanism A Contains spatial information, y A It is a high-resolution image obtained by downsampling the original 10m band image by two times; The channel feature extraction branch extracts y' by using the channel attention mechanism B Contains channel information, y B The original 20m band image is downsampled twice to obtain a low-resolution image. Bilinear interpolation is used to upsample it twice to convert it into a low-resolution image with the same resolution as y. A An image y' with the same spatial dimensions B ; After the features extracted by the spatial feature extraction branch and the channel feature extraction branch are fused through the trunk, feature extraction is performed again using convolution and residual scaling mechanisms, and finally the final result is obtained through the residual structure; Step 5: Use the training set to train the built model; Step 6: Use the trained model to test the test set to obtain the super-resolution image of Sentinel-2 remote sensing image.

2. The Sentinel-2 remote sensing image super-resolution reconstruction method based on the attention mechanism according to claim 1 is characterized in that The spatial attention mechanism performs global average pooling and global maximum pooling on the high-resolution input to obtain an effective feature descriptor, which is then cascaded according to the channel dimension and sent to a standard convolutional layer to obtain a spatial attention map through a Sigmoid function. Finally, the original input is multiplied by the spatial attention map to rescale the original input. The computation of the spatial attention map is expressed as: G Avg =AvgPool(G) G Max =MaxPool(G) G s =σ(ω(G Avg +G Max )) Among them, G is the input feature map, G Avg is the feature map after global average pooling, G Max is the feature map after global maximum pooling: G∈R C×H×W , G Avg ∈R 1×H×W , G Max ∈R 1×H×W , ω represents the convolution operation with a filter size of 3×3, σ represents the Sigmoid function, and C, H, and W represent the number of bands, length, and width of the remote sensing image used, respectively.

3. The Sentinel-2 remote sensing image super-resolution reconstruction method based on the attention mechanism according to claim 1 is characterized in that The channel attention mechanism converts the global spatial information of the channel direction into a channel feature descriptor by using global average pooling, and then obtains the weight coefficient of each channel through a convolution activation layer, and multiplies the weight coefficient with the original feature to obtain a scaled channel attention map, so as to re-weight the features of different channels, and finally uses the residual to add the original input and the extracted channel attention map; The calculation of the channel attention map is expressed as: F c =F c-1 +R(F)·F Among them, F c-1 and F are the input and output of the entire large residual block, F Avg refers to the feature map after global average pooling, R(·) represents channel attention, ω1 represents the convolution operation with a filter size of 3×3, and ω2 represents the convolution operation with a filter size of 1×1. represents the ReLU activation function, and σ represents the Sigmoid function.

Images