Land use classification method based on multiple attention semantic segmentation

Abstract

The application discloses a land use classification method based on multiple attention semantic segmentation, and comprises the following steps: S1, a land use model based on multiple attention modules and feature enhancement layers is designed; S2, a public data set is processed, and a model data set of the application is established; S3, a data set verification and comparison model method is used: the land use model proposed in S1 is subjected to a comparison experiment, and the accuracy of the improved model is verified. The application reduces the probability of error classification caused by similar features, combines deep separable convolution and average pooling to accelerate operation efficiency, deepens features, and strengthens feature description from multiple aspects.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing information processing technology, and more specifically, relates to a land use classification method based on multiple attention semantic segmentation. Background Technology

[0002] Land use information is fundamental to understanding the dynamic changes on the Earth's surface and the interactions between society and ecology. It is indispensable in many Earth-based observational fields, such as urban and regional planning, environmental vulnerability and impact assessment, and natural disaster and hazard monitoring. Land use information classification is a crucial reference for the rational use and management of urban land resources. To this day, obtaining high-precision land use classification results from remote sensing imagery remains a hot topic.

[0003] Traditional land use classification methods include land mapping and remote sensing mapping. Land mapping, or land surveying, while highly accurate and capable of combining information at different scales, is too costly for large-area mapping due to its reliance on manpower, time, and money. It has been gradually replaced by rapidly developing remote sensing mapping technology. Remote sensing images provide a cost-effective and convenient way to obtain an overview of a study area. High-resolution remote sensing images can clearly express the spatial structure and surface texture features of ground features, distinguish more detailed internal components, and provide clearer edge information, thus providing the conditions and foundation for effective land use classification.

[0004] The improvement in computer technology and remote sensing image resolution has led to the application of machine learning methods such as artificial neural networks, support vector machines, and fuzzy adaptive resonance theory supervised prediction mapping in land classification.

[0005] Traditional semantic segmentation techniques include threshold-based image semantic segmentation, edge-based image semantic segmentation, region-based image semantic segmentation, and theory-based image semantic segmentation. Traditional semantic segmentation methods are based on low-level semantic information for image processing, utilizing only shallow feature information and lacking understanding of contextual relationships and information interaction between pixels. They also cannot cover all land types for overall classification. Therefore, land classification models need to extract high-level semantic information, utilizing both spatial and global features of the image. In recent years, deep networks with stronger learning capabilities and more comprehensive feature extraction have emerged for land classification applications. These include: SegNet, which uses a pooling index method to save memory and can segment larger images; Pyramid Scene Parsing Network, whose SPP structure increases the receptive field, enabling the network to better integrate global information; U-Net's U-shaped structure and skip connections effectively utilize semantic information at different levels of the image; and Deeplabv3+, which uses a decoder-encoder structure based on the ASPP structure to enhance feature representation capabilities, and the retention of a large amount of shallow information allows for better edge recognition of segmented objects. Besides deep networks like CNNs, transformers from natural language processing are also beginning to be applied to land classification. This type of network uses a multi-head attention mechanism as its main construction, allowing the model to focus on information from different locations, ultimately acquiring richer features of land cover types. Summary of the Invention

[0006] The technical challenge to be solved by this invention is to overcome the limitations of existing technologies and address the problem of large errors in existing land use classification methods. This invention provides a land use classification method based on multiple attention semantic segmentation, which improves upon the original model by adding different feature extraction modules to obtain more accurate land use information.

[0007] To address at least one of the aforementioned technical problems, according to one aspect of the present invention, a land use classification method based on multiple attention semantic segmentation is provided, comprising the following steps:

[0008] S1. Design a land use model based on multiple attention modules and feature enhancement layers:

[0009] The land use classification model SCU-net, which incorporates multiple computational power modules, is an improvement on the U-net network. It can fuse contextual information and extract global features from downsampled local feature maps, and perform feature enhancement on upsampled feature maps for land use classification. The improved model includes SC modules with various attention structures and feature enhancement layers.

[0010] S2. Process the public dataset to create your own model dataset:

[0011] Since the format of the publicly available dataset does not meet the training format of the model, it is necessary to crop all the images to 512*512 pixels to create the dataset.

[0012] S3. Methods for validating and comparing models using datasets:

[0013] A comparative experiment was conducted on the land use model proposed in S1 to verify the accuracy of the improved model: the new land use classification model showed outstanding performance in complex land cover classification scenarios.

[0014] Regarding the land use classification method SCU-net based on multiple attention semantic segmentation described in S1, its

[0015] Specifically, it includes:

[0016] S11. Construction of the basic framework U-net model:

[0017] This land use classification model is an improvement on U-net, which uses ResNet50 as its backbone network. The main structure of ResNet50 is as follows: first, the remote sensing image is convolved with a 7*7 kernel size and a stride of 2, then max pooling is performed, followed by Bottleneck structures with 3, 4, 6, and 3 layers respectively. Bottleneck is specifically a residual structure composed of a 1*1 convolution, a 3*3 convolution, and a 1*1 convolution in sequence.

[0018] S12. Construction of SC modules containing multiple attention structures:

[0019] This module combines the attention computation methods of CNN and Transformer to enhance the acquisition of local and global feature information from feature maps, thereby revealing the relationships between different land use types. This module comprises two key components: Swing Transformer Blocks and the Convolutional Block Attention Module (CBAM).

[0020] Swin Transformer Blocks consist of two consecutive blocks. The input feature map is first processed into 1x1 blocks using Embeddedpatches, and then flattened along the channel direction. After passing through the W-MSA module and then the LayerNorm operation, it enters the Window Multi-head Self-Attention (W-MSA) layer. W-MSA divides the feature map into multiple non-overlapping windows, and then calculates self-attention within each window. If the given feature map size is H*W*C and the segmentation window size is M, then the computational complexity of the window multi-head self-attention is:

[0021] Ω(W-MSA)=4HWC 2 +2M 2 HWC#(1)

[0022] The window size is much smaller than the entire image, reducing computational cost. After passing through the MLP layer and Dropout, the feature map is input into the second block. Both are similar, but the difference lies in that W-MSA is replaced by Shifted Window Multi-head Self-Attention (SW-MSA). SW-MSA shifts the segmentation window to generate new windows. To maintain the original window computational cost while enabling information exchange between different windows, some windows are shifted and then Masked Multi-head Self-Attention (Masked MSA) is applied to reduce information misalignment.

[0023] CBAM consists of two parts: Channel Attention Module (CAM) and Spatial Attention Module (SAM). The input feature map (F) is processed through two parallel MaxPool layers and AvgPool layers, a Share MLP module, summation, and sigmoid activation operations to finally obtain the channel attention feature F'. The channel attention calculation process is shown in equations (2) and (3).

[0024]

[0025] Unlike CAM, SAM performs max pooling and average pooling on the input feature map F', then concatenates the two results, and finally activates it with sigmoid to obtain a feature map F'' that integrates spatial attention and channel attention. The specific calculation process is shown in equations (4) and (5).

[0026]

[0027] S13. Upsampling and feature enhancement layer processing:

[0028] The feature maps obtained from Swing Transformer Blocks and CBAM are stacked and then convolutionally adjusted to reduce the number of channels from 4096 to 2048. Then, upsampling convolution is performed to double the size of the feature map each time, and the number of channels is successively changed to 512, 256, and 128. Finally, the feature enhancement layer is entered to further enhance the features.

[0029] The feature enhancement layer receives two inputs: an upsampled feature map, which has undergone feature extraction by the SC module and integrates deep semantic information with global effective features; and a 64-channel feature map that has only undergone one downsampling process. This downsampling compensates for information loss during feature extraction and supplements features for less frequently used land use types to prevent them from being overlooked. In the feature enhancement layer, the two feature maps are multiplied by learnable parameters X1 and X2, respectively, using the following formula: where a1 and a2 are self-generated parameters, and b is a constant.

[0030]

[0031]

[0032] After adding the two, ordinary convolution, depthwise separable convolution, and adaptive pooling are applied separately. Depthwise separable convolution reduces computational cost compared to ordinary convolution. Through feature enhancement operations in three branches, the effective land use types identified are improved, the perturbation caused by similar features is reduced, and the overall land classification accuracy is increased.

[0033] The dataset used in S2 is the GID dataset. To facilitate model training, all images were cropped to 512*512 pixels. There are a total of 6 land use types: background, forest, grassland, farmland, water body, and buildings, labeled with black, indigo, green, yellow, blue, and red, respectively. The cross-entropy loss function is used for training.

[0034] The comparison models used in S3 are U-net, PSPNet, Segnet, and Deeplabv3+, all with the same settings, a batch size of 4, and the use of cross-entropy loss function for training.

[0035] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements S in the land use classification method based on multiple attention semantic segmentation of the present invention.

[0036] According to another aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements S in the land use classification method based on multiple attention semantic segmentation of the present invention.

[0037] Compared with the prior art, the present invention has at least the following beneficial effects:

[0038] This invention presents a land use classification method based on multi-attention semantic segmentation, which fully leverages the capabilities of CNNs in capturing local image features and Transformers in extracting overall features. By combining CBAM and Swin Transformer Blocks to extract global contextual information, it compensates for the CNN's neglect of relationships between pixels. Deep features are processed through the SC module to obtain more overall image information, improving the model's ability to recognize multiple categories. In the feature enhancement layer, learnable parameters are introduced to adjust the influence of the input image on the segmentation results, improving the model's classification accuracy with minimal computation. Multi-branch convolution is used for feature enhancement, and the separable convolution further reduces parameters, which is beneficial for image feature classification. This invention not only extracts water bodies more accurately but also significantly reduces feature loss during the learning process, achieving relatively high-precision land use classification images even in complex multi-category remote sensing images. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0040] Figure 1 This is a flowchart of the method of the present invention;

[0041] Figure 2 This is a diagram of the SCU-net architecture for the land use classification method in Embodiment 1 of the present invention.

[0042] Figure 3 The diagram shows the overall structure and process of the Swin Transformer Blocks in the SC module of Embodiment 1 of the present invention, where (a) is the overall structure diagram, (b) is the W-MSA calculation process diagram, and (c) is the SW-MSA calculation process diagram.

[0043] Figure 4 The diagram shows the overall structure of the CBAM module and its components in the SC module of Embodiment 1 of the present invention, where (a) is the overall structure of the CBAM module, (b) is the structure of the channel attention module, and (c) is the structure of the spatial attention module.

[0044] Figure 5 This is a structural diagram of the feature enhancement layer in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0046] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0047] like Figure 1-5 As shown,

[0048] Example 1:

[0049] like Figure 1 As shown, this invention designs a land use classification method based on multiple attention semantic segmentation, which specifically includes the following steps:

[0050] Step 1. Design a land use model based on multiple attention modules and feature enhancement layers:

[0051] The land use classification model SCU-net, which incorporates multiple computational power modules, is an improvement on the U-net network. It can fuse contextual information and extract global features from downsampled local feature maps, and perform feature enhancement on upsampled feature maps for land use classification. The improved model includes SC modules with various attention structures and feature enhancement layers.

[0052] For the land use classification method SCU-net based on multiple attention semantic segmentation described in step 1, its

[0053] Specifically, it includes:

[0054] Step 1.1. Construction of the basic framework U-net model:

[0055] This land use classification model is an improvement on U-net, which uses ResNet50 as its backbone network. The main structure of ResNet50 is as follows: first, the remote sensing image is convolved with a 7*7 kernel size and a stride of 2, then max pooling is performed, followed by Bottleneck structures with 3, 4, 6, and 3 layers respectively. Bottleneck is specifically a residual structure composed of a 1*1 convolution, a 3*3 convolution, and a 1*1 convolution in sequence.

[0056] Step 1.2. Construction of SC modules containing multiple attention structures:

[0057] This module combines the attention computation methods of CNN and Transformer to enhance the acquisition of local and global feature information from feature maps, thereby revealing the relationships between different land use types. This module comprises two key components: Swing Transformer Blocks and the Convolutional Block Attention Module (CBAM).

[0058] like Figure 3 As shown, the Swin Transformer Blocks consist of two consecutive blocks. The input feature map is first processed into 1*1 blocks using embedded patches, and then flattened along the channel direction. After passing through the W-MSA module and then the LayerNorm operation, it enters the Window Multi-head Self-Attention (W-MSA) layer. W-MSA divides the feature map into multiple non-overlapping windows, and then calculates self-attention within each window. If the given feature map size is H*W*C and the segmentation window size is M, then the computational complexity of the window multi-head self-attention is:

[0059] Ω(W-MSA)=4HWC 2 +2M 2 HWC#(1)

[0060] The window size is much smaller than the entire image, reducing computational cost. After passing through the MLP layer and Dropout, the feature map is input into the second block. Both are similar, but the difference lies in that W-MSA is replaced by Shifted Window Multi-head Self-Attention (SW-MSA). SW-MSA shifts the segmentation window to generate new windows. To maintain the original window computational cost while enabling information exchange between different windows, some windows are shifted and then Masked Multi-head Self-Attention (Masked MSA) is applied to reduce information misalignment.

[0061] like Figure 4As shown, CBAM consists of two parts: Channel Attention Module (CAM) and Spatial Attention Module (SAM). The input feature map (F) is processed through two parallel MaxPool layers and AvgPool layers, a Share MLP module, summation, and sigmoid activation operations to finally obtain the channel attention feature F'. The channel attention calculation process is shown in equations (2) and (3).

[0062]

[0063] Unlike CAM, SAM performs max pooling and average pooling on the input feature map F', then concatenates the two results, and finally activates it with sigmoid to obtain a feature map F'' that integrates spatial attention and channel attention. The specific calculation process is shown in equations (4) and (5).

[0064]

[0065] Step 1.3 Upsampling and Feature Enhancement Layer Processing:

[0066] The feature maps obtained from Swing Transformer Blocks and CBAM are stacked and then convolutionally adjusted to reduce the number of channels from 4096 to 2048. Then, upsampling convolution is performed to double the size of the feature map each time, and the number of channels is successively changed to 512, 256, and 128. Finally, the feature enhancement layer is entered to further enhance the features.

[0067] The feature enhancement layer receives two inputs: an upsampled feature map, which has undergone feature extraction by the SC module and integrates deep semantic information with global effective features; and a 64-channel feature map that has only undergone one downsampling process. This downsampling compensates for information loss during feature extraction and supplements features for less frequently used land use types to prevent them from being overlooked. In the feature enhancement layer, the two feature maps are multiplied by learnable parameters X1 and X2, respectively, using the following formula: where a1 and a2 are self-generated parameters, and b is a constant.

[0068]

[0069]

[0070] After adding the two, ordinary convolution, depthwise separable convolution, and adaptive pooling are applied separately. Depthwise separable convolution reduces computational cost compared to ordinary convolution. Through feature enhancement operations in three branches, the effective land use types identified are improved, the perturbation caused by similar features is reduced, and the overall land classification accuracy is increased.

[0071] The dataset used in step 2 is the GID dataset. To facilitate model training, all images were cropped to 512*512 pixels. There are a total of 6 land use types: background, forest, grassland, farmland, water body, and buildings, labeled with black, indigo, green, yellow, blue, and red, respectively. The cross-entropy loss function is used for training.

[0072] The comparison models used in step 3 are U-net, PSPNet, Segnet, and Deeplabv3+. All settings are the same, the batch size is set to 4, and the cross-entropy loss function is used for training.

[0073] The effectiveness of this invention was validated on the GID dataset. Overall accuracy (OA), precision, recall, F1 score, and mean intersection-over-union (MIoU) were used. Higher values ​​indicate better segmentation performance. Recall assesses the severity of undersegmentation; a lower value indicates more severe undersegmentation. Precision assesses the severity of oversegmentation; a lower value indicates more severe oversegmentation. F1 score measures the balance between precision and recall. As shown in Table 1, SCU-net achieved the best OA, F1 score, and MIoU compared to other models. Furthermore, ablation experiments were conducted to verify the impact of adding modules on classification; see Table 2.

[0074] Table 1. Comparison of the proposed method with other models on the GID dataset.

[0075]

[0076] Table 2. Ablation Experiment

[0077]

[0078] It can be seen that the method proposed in this invention has better performance than other existing methods, with OA reaching 77.35, F1-score reaching 76.38, and MIoU reaching 61.80.

[0079] In summary, this invention fully leverages the capabilities of CNNs in capturing local image features and Transformers in extracting global features. By combining CBAM and Swin Transformer Blocks to extract global contextual information, it compensates for the CNN's neglect of relationships between pixels. Deep features are processed through the SC module to obtain more overall image information, improving the model's ability to recognize multiple categories. In the feature enhancement layer, learnable parameters are introduced to adjust the influence of the input image on the segmentation results, improving the model's classification accuracy with minimal computation. This invention not only extracts different land use types more accurately but also significantly reduces feature loss during the learning process, achieving relatively high-precision land use classification images even in complex multi-category remote sensing images.

[0080] Example 2:

[0081] The computer-readable storage medium of this embodiment stores a computer program that, when executed by a processor, implements the steps in the land use classification method based on multiple attention semantic segmentation of Embodiment 1.

[0082] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0083] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0084] Example 3:

[0085] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the land use classification method based on multiple attention semantic segmentation of Embodiment 1.

[0086] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0087] Those skilled in the art will understand that the content disclosed in the embodiments can be provided as a method, system, or computer program product. Therefore, this solution can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.

[0088] This solution is described with reference to flowchart illustrations and / or block diagrams of methods and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0092] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

Claims

1. A land use classification method based on multi-attention semantic segmentation, characterized in that, Includes the following steps: S1. Design a land use model based on multiple attention modules and feature enhancement layers: Introduce a land use classification model SCU-net with multiple computational power modules. The land use classification model SCU-net is an improvement on the U-net network. It can fuse contextual information and extract global features from the downsampled local feature map, and perform feature enhancement on the upsampled feature map for land use classification. The improved model includes SC modules with multiple attention structures and feature enhancement layers. S2. Process the public dataset to create your own model dataset; S3. Methods for validating and comparing models using datasets: Comparative experiments were conducted on the land use model proposed in S1 to verify the accuracy of the improved model; The SCU-net land use classification model based on multi-attention semantic segmentation in S1 specifically includes: S11. Construction of the basic framework U-net model: First, the remote sensing image is convolved with a kernel size of 7*7 and a stride of 2, then max pooling is performed, followed by Bottleneck structures with numbers of 3, 4, 6, and 3 respectively; Bottleneck is specifically a residual structure composed of a 1*1 convolution, a 3*3 convolution, and a 1*1 convolution in sequence. S12. Construction of SC modules containing multiple attention structures: The SC module, which includes multiple attention structures, combines the attention calculation methods of CNN and Transformer to enhance the collection of local and global feature information from the feature map and obtain the relationship between different land use types. S13. Upsampling and feature enhancement layer processing: After stacking the feature maps obtained from Swing Transformer Blocks and CBAM, convolution is performed to adjust the number of channels from 4096 to 2048. Then, upsampling convolution is performed to restore the feature map size to twice its size each time, and the number of channels is successively changed to 512, 256, and 128. Finally, the feature enhancement layer is entered to further enhance the features. The feature enhancement layer receives two inputs: an upsampled feature map, which has undergone feature extraction by the SC module, fusing deep semantic information with global effective features; and a 64-channel feature map that has only undergone one downsampling process. This downsampling compensates for information loss during feature extraction and supplements features for less frequently used land use types to prevent them from being overlooked. In the feature enhancement layer, the two feature maps are multiplied by learnable parameters X1 and X2, respectively, using the following formula: where a1 and a2 are self-generated parameters, and b is a constant. After adding the two, ordinary convolution, depthwise separable convolution, and adaptive pooling are performed respectively. Depthwise separable convolution reduces the amount of computation compared to ordinary convolution. Through feature deepening operations in three branches, the effective land use types classified are improved, the perturbation of similar features on classification is reduced, and the overall land classification accuracy is improved.

2. The method according to claim 1, characterized in that, The SC module, which includes multiple attention structures, comprises two important components: Swing Transformer Blocks and Convolutional Block Attention Module; Swin Transformer Blocks consist of two consecutive blocks. The input feature map is first processed into 1*1 blocks using Embeddedpatches, and then flattened along the channel direction. After passing through the W-MSA module and the LayerNorm operation, it enters the Window Multi-head Self-Attention (W-MSA) layer. W-MSA divides the feature map into multiple non-overlapping windows, and then calculates self-attention within each window. If the given feature map size is H*W*C and the segmentation window size is M, then the computational complexity of the window multi-head self-attention is: Ω(W-MSA)=4HWC 2 +2M 2 HWC (1) CBAM consists of two parts: Channel Attention Module and Spatial Attention Module; the input feature map goes through two parallel MaxPool layers and AvgPool layers, a Share MLP module, summation and sigmoid activation operations to finally obtain the channel attention feature F'; The channel attention calculation process is shown in equations (2) and (3): M C (F)<σ(MLP(AvgPool(F)))+MLP(MaxPool(F)) ( 2) Unlike CAM, SAM performs max pooling and average pooling on the input feature map F', then concatenates the two results, and finally activates it with sigmoid to obtain a feature map F'' that integrates spatial attention and channel attention; the specific calculation process is shown in equations (4) and (5): M s (F)<σ(f7*7([AvgPool(F');MaxPool(F')])) (4) 3. The method according to claim 1, characterized in that, S2 specifically involves creating a dataset by cropping images to a size of 512*512.

4. The method according to claim 3, characterized in that, The dataset used in S2 is the GID dataset. To facilitate model training, the images were cropped to 512*512 pixels. There are a total of 6 land use types: background, forest, grassland, farmland, water body, and buildings. The colors are labeled as black, indigo, green, yellow, blue, and red, respectively. The cross-entropy loss function is used for training.

5. The method according to claim 1, characterized in that, The comparison models used in S3 are U-net, PSPNet, Segnet, and Deeplabv3+, all with the same settings, a batch size of 4, and the use of cross-entropy loss function for training.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps in the land use classification method based on multiple attention semantic segmentation as described in any one of claims 1 to 5.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the land use classification method based on multiple attention semantic segmentation as described in any one of claims 1 to 5.

Images