Remote sensing image scene classification method based on MCA attention mechanism efficient network
By introducing the MCA attention mechanism in EfficientNet B0, combining average pooling and maximum pooling, the problem of insufficient feature capture in lightweight networks in high-resolution remote sensing image processing is solved, and efficient remote sensing image scene classification is achieved, which significantly Improved classification accuracy and model performance.
Patent Information
- Application Number
- CN202410971507.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-08
AI Technical Summary
In the classification of remote sensing image scenes, lightweight convolutional neural networks such as EfficientNet B0 are difficult to fully capture subtle features and semantic information when processing high-resolution complex images, and there are dimensional disasters and complexity problems in parameter adjustment.
The MCA attention mechanism was introduced in EfficientNet B0. The coordinate attention mechanism was combined with average pooling and maximum pooling to build the MCA-EfficientNet B0 network, using ImageNet pre-training weights for transfer learning, and inserting MCA attention after the convolution layer Modules are represented with enhanced features.
It significantly accelerates the speed of model training, improves classification accuracy, maintains a low amount of parameters and calculation, and is better than a variety of deep learning models, especially when processing complex remote sensing images.
Smart Images

Figure CN118918372A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image classification, and in particular to a remote sensing image scene classification method based on an MCA attention mechanism efficient network. Background Art
[0002] In recent years, with the continuous advancement of satellite and aerial remote sensing technology, the resolution of remote sensing images has become higher and higher, and the information contained in them has become richer and richer. These high-resolution remote sensing images have been widely used in the fields of land use, resource survey, natural disaster observation, etc. However, with the increase in resolution, the semantic information of remote sensing images has become more complex, and their spatial texture has become more diverse. At the same time, the differences within the class are significant, while the differences between classes are relatively small. Therefore, how to quickly and accurately classify remote sensing images has become a research hotspot in the field of computer vision.
[0003] Early remote sensing image scene classification methods mainly include methods based on manual features and methods based on mid-level features. Methods based on manual features refer to the use of manually designed features to describe the content of remote sensing images, such as color, texture, structure and other traditional image processing features. The process of extracting these features is time-consuming, and it is usually difficult to cover all the information in the image, and it cannot adapt to complex and changeable remote sensing image scenes. Methods based on mid-level features refer to encoding and aggregating low-level visual features to obtain more discriminative and expressive features. This method usually requires complex feature engineering and a lot of parameter adjustments, and may encounter dimensionality disasters when processing data with high intrinsic dimensions.
[0004] In the past decade, artificial intelligence and deep learning have developed rapidly. Deep learning models represented by convolutional neural networks (CNN) and Transformer have made important progress in image classification, object detection, natural language processing and other fields. Through end-to-end training, deep neural networks can automatically learn and extract high-level abstract features, avoiding tedious feature engineering, and perform well in understanding image content and improving the accuracy of scene classification. Therefore, many experts and scholars have applied them to remote sensing image scene classification tasks. Among them, lightweight convolutional neural networks perform particularly well. Compared with large networks, they have fewer parameters and computational complexity, while maintaining certain classification performance, and can be easily deployed in resource-constrained mobile devices and embedded systems. As a representative lightweight CNN, the EfficientNet B0 network is designed through Neural Architecture Search (NAS) to optimize accuracy and efficiency. Its core advantage lies in the use of a compound scaling method that improves performance by balancing the depth, width and resolution of the network.
[0005] Although EfficientNet B0 is built based on high-efficiency design principles and performs well in general image classification tasks, it may not be able to fully capture all subtle features and semantic information when processing particularly complex or detail-rich image data, such as high-resolution remote sensing images. Summary of the invention
[0006] In order to solve the above technical problems, the present invention provides a remote sensing image scene classification method based on an efficient network of an MCA attention mechanism, comprising the following steps:
[0007] S1. Obtain a remote sensing image dataset and divide it into training samples and test samples in proportion;
[0008] S2, preprocessing the remote sensing image dataset, including cropping, data enhancement and standardization;
[0009] S3. Add maximum pooling to coordinate attention CA to build MCA attention module;
[0010] S4, insert the MCA attention module after the last convolutional layer of the EfficientNet B0 convolutional neural network and before the average pooling and fully connected layers to build the MCA-EfficientNet B0 network model;
[0011] S5. Load the pre-trained weights on the ImageNetlK dataset into the MCA-EfficientNet B0 network model, then input the training samples into the MCA-EfficientNet B0 network model, continuously optimize the weighted cross entropy loss function through the back propagation algorithm, adjust the network parameters, and finally obtain the trained network model;
[0012] S6. Input the test sample into the trained network model to obtain the remote sensing image scene classification result.
[0013] The technical solution further defined in the present invention is:
[0014] Furthermore, in step S1, the three remote sensing image scene classification datasets, UC Merced Land-Use, AID and NWPU-RESISC45, are divided into training samples and test samples according to the ratios of 8:2, 5:5 and 2:8 respectively.
[0015] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, in step S2, the images in the remote sensing image dataset are cropped to a resolution of 224×224, and then the images are horizontally flipped with a probability p, and finally the image data is standardized.
[0016] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, in step S2, the probability p is set to 0.5.
[0017] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, in step S3, in the coordinate information embedding stage, for the input X, firstly, average pooling and maximum pooling encoding operations are performed on each channel along the horizontal coordinate and the vertical coordinate, and then 1×1 convolution and ReLu activation function are used to process each channel independently; finally, the results of the two pooling operations are added element by element; therefore, the output of the cth channel with a height of h is:
[0018]
[0019] Similarly, the output of the cth channel with width w is:
[0020]
[0021] Among them, x c (h, i) represents the feature of the cth channel with height h, x c (j, w) represents the feature of the cth channel with a width of w; W represents the width of the feature map, and H represents the height of the feature map; F 1 is the convolution operation; δ is the activation function.
[0022] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, in step S3, in the coordinate attention generation stage, the outputs of the two transformations are concatenated, and then a 1×1 convolution transformation function F is used. 1 Perform transformation operations on the concatenated features:
[0023] f=δ(F 1 ([z h , z w ]))
[0024] Among them, [·,·] represents the concatenation operation along the spatial dimension; z h is the feature in the vertical direction; Z w is the feature in the horizontal direction; δ represents the nonlinear activation function; f represents the intermediate feature map that encodes the spatial information in the horizontal and vertical directions; r is the reduction rate in the SE block, and the number of channels of f can be increased or decreased by controlling r;
[0025] Then, decompose f into two tensors along the spatial dimension and Where C / r represents the number of channels after the reduction using SE blocks, is a real vector space, f h is the tensor in the vertical direction, f w is the tensor in the horizontal direction;
[0026] Then use another two 1×1 convolution transformations F h and F w Separately put f h and f w Transformed into a tensor with the same number of channels as the input X, we get:
[0027] g h =σ(F h (f h ))
[0028] g w =σ(F w (f w ))
[0029] Among them, σ is the Sigmoid activation function, g h and g w They are C×H×1 and C×1×W feature maps respectively; they are then expanded and used as attention weights; the final output Y is expressed as:
[0030]
[0031] Among them, x c (i, j) is the input of the cth channel, is the attention weight in the vertical direction, is the attention weight in the horizontal direction.
[0032] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, step S5 specifically includes the following sub-steps:
[0033] S5.1, extract shallow information through 3×3 convolution layer;
[0034] S5.2, extract deep features through 7 MBConv modules in sequence;
[0035] S5.3, through a 3×3 convolutional layer, further extract features;
[0036] S5.4, aggregate key information through the MCA attention module;
[0037] S5.5. Finally, the classification results of the training samples are obtained through average pooling and fully connected layers.
[0038] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, in step S1, the UCMerced Land-Use dataset contains remote sensing images of 21 scene categories, including airplanes, beaches, baseball fields, and buildings. Each scene category includes 100 images of 256×256 pixels, totaling 2100 images, and the spatial resolution of the images is 0.3m.
[0039] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, in step S1, the AID dataset contains remote sensing images of 30 scene categories, including airports, bridges, churches, and densely populated residential areas. Each scene category includes 220 to 420 images of 600×600 pixels, totaling 10,000 images, and the spatial resolution of the images varies between 0.5 and 8 meters.
[0040] As described above, in the remote sensing image scene classification method based on the MCA attention mechanism efficient network, in step S1, the NWPU-RESISC45 dataset contains remote sensing images of 45 scene categories, each scene category includes 700 images of 256×256 pixels, totaling 31,500 images, and the spatial resolution of the images varies between 0.2 and 30 meters.
[0041] The beneficial effects of the present invention are:
[0042] (1) In the present invention, the pre-trained weights on the ImageNet1K dataset are used for transfer learning, which significantly speeds up the training speed of the model and saves training time;
[0043] (2) In this paper, maximum pooling is added to coordinate attention CA, and an MCA attention mechanism is proposed. By combining average pooling and maximum pooling, MCA can capture the overall information of remote sensing images while highlighting their local features;
[0044] (3) In the present invention, the MCA attention module is inserted after the last convolutional layer of the EfficientNet B0 model and before the average pooling and fully connected layers, which can further aggregate key information, enhance feature representation, and improve the classification accuracy of the model;
[0045] (4) In the present invention, a large number of experiments were carried out on three remote sensing image scene classification benchmark datasets, and the results were better than a variety of deep learning models while still maintaining a low number of parameters and computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a schematic diagram of the overall process of the present invention;
[0047] Figure 2 Schematic diagram of the structure of the MCA attention module in an embodiment of the present invention;
[0048] Figure 3 : is a training flow chart of the MCA-EfficientNet B0 convolutional neural network in an embodiment of the present invention;
[0049] Figure 4 is a representative image of the UC Merced Land-Use dataset in an embodiment of the present invention;
[0050] Figure 5 is the confusion matrix of the classification results of the embodiment of the present invention on the UC Merced Land-Use dataset;
[0051] Figure 6 is a representative diagram of the AID data set in an embodiment of the present invention;
[0052] Figure 7 is a confusion matrix of the classification results of the embodiment of the present invention on the AID data set;
[0053] Figure 8 It is a representative diagram of the NWPU-RESISC45 data set in an embodiment of the present invention;
[0054] Fig. 9 This is the confusion matrix of the classification results of the embodiment of the present invention on the NWPU-RESISC45 dataset. DETAILED DESCRIPTION
[0055] This embodiment provides a remote sensing image scene classification method based on an efficient network of MCA attention mechanism, such as Figure 1 As shown, the following steps are included:
[0056] S1. Obtain a remote sensing image dataset and divide it into training samples and test samples according to the ratio. Divide the training samples and test samples into the ratios of 8:2, 5:5 and 2:8 for the three remote sensing image scene classification datasets, UC MercedLand-Use, AID and NWPU-RESISC45, respectively.
[0057] S2. Preprocess the remote sensing image dataset, including cropping, data enhancement and standardization; crop the images in the remote sensing image dataset to 224×224 resolution, and then flip the images horizontally with a probability of p=0.5, which can increase sample diversity, and finally standardize the image data.
[0058] S3. Improve the coordinate attention (CA) by adding maximum pooling to the original average pooling and constructing the MCA attention module, such as Figure 2 shown.
[0059] For the MCA attention module, in the coordinate information embedding stage, for the input X, firstly, average pooling and maximum pooling encoding operations are performed on each channel along the horizontal and vertical coordinates, and then 1×1 convolution and ReLu activation functions are used to process each channel independently to increase the nonlinear ability and expression ability of the network; finally, the results of the two pooling operations are added element by element; therefore, the output of the cth channel with a height of h is:
[0060]
[0061] Similarly, the output of the cth channel with width w is:
[0062]
[0063] Among them, x c (h, i) represents the feature of the cth channel with height h, x c (j, w) represents the feature of the cth channel with a width of w; W represents the width of the feature map, and H represents the height of the feature map; F 1 is the convolution operation; δ is the activation function.
[0064] The above two transformations aggregate features along two spatial directions respectively, producing a pair of direction-aware feature maps, which can obtain the global receptive field and accurately encode the position information.
[0065] In the coordinate attention generation stage, the outputs of the above two transformations are concatenated and then a 1×1 convolution transformation function F is used. 1 Perform transformation operations on the concatenated features:
[0066] f=δ(F 1 ([z h , z w ]))
[0067] Among them, [·,·] represents the concatenation operation along the spatial dimension; z h is the feature map obtained after performing the above series of operations along the horizontal coordinate (i.e., the feature in the vertical direction); Z WW is the feature map obtained after performing the above series of operations along the vertical coordinate (i.e., the feature in the horizontal direction); δ represents the nonlinear activation function; f represents the intermediate feature map that encodes the spatial information in the horizontal and vertical directions; r is the reduction rate in the SE (Squeeze-and-Excitation) block, and the number of channels of f can be increased or decreased by controlling r.
[0068] Then, decompose f into two tensors along the spatial dimension and Where C / r represents the number of channels after the reduction using SE blocks, is a real vector space, f h is the tensor in the vertical direction, f W is the horizontal tensor.
[0069] Next, use another two 1×1 convolutions to transform F h and F w Separately put f h and f w Transformed into a tensor with the same number of channels as the input X, we get:
[0070] g h =σ(F h (f h ))
[0071] g W =σ(F w (f w ))
[0072] Among them, σ is the Sigmoid activation function, g h and g w They are C×H×1 and C×1×W feature maps respectively; they are then expanded and used as attention weights; the final output Y is expressed as:
[0073]
[0074] Among them, x c (i, j) is the input of the cth channel, is the attention weight in the vertical direction, is the attention weight in the horizontal direction.
[0075] S4. Insert the MCA attention module after the last convolutional layer of the EfficientNet B0 convolutional neural network and before the average pooling and fully connected layers to build the MCA-EfficientNet B0 network model.
[0076] S5. Load the pre-trained weights on the ImageNet1K dataset into the MCA-EfficientNet B0 network model, then input the training samples into the MCA-EfficientNet B0 network model, continuously optimize the weighted cross entropy loss function through the back propagation algorithm, adjust the network parameters, and finally obtain the trained network model.
[0077] Step S5 specifically includes the following sub-steps:
[0078] S5.1, extract shallow information through 3×3 convolution layer;
[0079] S5.2, extract deep features through 7 MBConv modules in sequence;
[0080] S5.3, through a 3×3 convolutional layer, further extract features;
[0081] S5.4, aggregate key information through the MCA attention module;
[0082] S5.5. Finally, the classification results of the training samples are obtained through average pooling and fully connected layers.
[0083] In this process, the model continuously optimizes the weighted cross entropy loss function through the back propagation algorithm, adjusts the network parameters, and obtains a trained network model; Figure 3 As shown in the figure, the complete training process of the remote sensing image scene classification method based on MCA-EfficientNet B0 is demonstrated.
[0084] S6. Input the test sample into the trained network model to obtain the remote sensing image scene classification result.
[0085] In one example, the UC Merced Land-Use dataset contains remote sensing images of 21 scene categories, including airplanes, beaches, baseball fields, and buildings. Each scene category includes 100 images of 256×256 pixels, totaling 2,100 images, with a spatial resolution of 0.3m. Figure 4 shown.
[0086] Table 1 below gives a comparison of the classification performance of this embodiment with other deep learning methods on the UC Merced Land-Use dataset. As can be seen from Table 1, except for VGG16 and GoogLeNet, the accuracy of other methods is above 98%, indicating that the classification accuracy of the dataset has reached saturation and is difficult to improve. However, this embodiment still achieved the highest accuracy of 99.76%. Compared with EfficientNet B0, it increased by 0.38%, which is the same as ACNet, and significantly better than other methods. At the same time, the number of parameters and floating-point operation size of this embodiment are only slightly increased compared to EfficientNet B0, and compared to ResNet 50 and ViT-B 16, it is only a few tenths of them.
[0087] Table 1
[0088]
[0089] like Figure 5 As shown in FIG. , the confusion matrix of this embodiment on the UC Merced Land-Use dataset is shown. Figure 5 As can be seen from the figure, this example shows excellent performance, and almost all scenes on the UC Merced Land-Use dataset are accurately identified. Only in the sparse residential area, the accuracy is 95%, and the remaining 5% of the samples are classified as oil tanks. This is because some oil tank scenes contain building information, and sparse residential areas also have circular structures similar to oil tanks, which makes it difficult for the model to effectively distinguish them.
[0090] In another example, the AID dataset contains remote sensing images of 30 scene categories, including airports, bridges, churches, and densely populated residential areas. Each scene category includes 220 to 420 images of 600×600 pixels, totaling 10,000 images. The spatial resolution of the images varies between 0.5 and 8 meters. A representative image of the AID dataset is shown below. Figure 6 shown.
[0091] Table 2 below shows the comparison of the classification performance of this embodiment with other deep learning methods on the AID dataset. As can be seen from Table 2, this embodiment has achieved the best classification performance, reaching 96.22%. Compared with EfficientNetBO, it has increased by 0.68%, and compared with ShuffleNet V2 and MobileNet V3, two lightweight CNNs, it has increased by 1.08% and 0.88% respectively. Even compared with ViT-B 16, this embodiment has a slight improvement, but it has a smaller number of parameters and calculations.
[0092] Table 2
[0093]
[0094] like Figure 7 As shown in FIG. , the confusion matrix of this embodiment on the AID data set is shown. Figure 7 It can be seen that among the 30 scene classifications, 21 categories have an accuracy rate of more than 95%, 27 categories have an accuracy rate of more than 90%, and the only 3 categories with an accuracy rate below 90% are commercial areas, parks, and resorts. The reason is that these categories of scenes have many common features, making it difficult for the model to distinguish them. However, in scenes such as bare land, farmland, and forests, this embodiment has achieved an accuracy rate of 100%, indicating that this embodiment still has a strong scene recognition capability.
[0095] In another example, the NWPU-RESISC45 dataset contains remote sensing images of 45 scene categories. Each scene category includes 700 images of 256×256 pixels, totaling 31,500 images. The spatial resolution of the images varies between 0.2 and 30 meters. Compared with other datasets, the intra-class diversity and inter-class similarity of this dataset are more obvious, and the resolution of some special terrains (such as islands, lakes, mountains, etc.) is low, so it is quite challenging to classify. Representative images of the NWPU-RESISC45 dataset are shown below: Figure 8 shown.
[0096] Table 3 below gives a comparison of the classification performance of this embodiment with other deep learning methods on the NWPU-RESISC45 dataset. As can be seen from Table 3, this embodiment still achieved the highest accuracy of 94.25%. It is 0.72% higher than EfficientNet B0, and is significantly higher than other methods except ResNet50. It is worth mentioning that ViT-B 16 has an accuracy of only 93.36%, mainly because the dataset uses a training test split ratio of 2:8, and Transformer models are more prone to overfitting when there are few training samples.
[0097] Table 3
[0098]
[0099] like Fig. 9 As shown in FIG. , the confusion matrix of this embodiment on the NWPU-RESISC45 data set is shown. Fig. 9 It can be seen that among the 45 scene classifications, 41 categories have an accuracy rate of more than 90%, and only four categories of scenes, namely churches, medium-sized residential areas, palaces, and railway stations, have not reached this standard. Among them, the accuracy rates of medium-sized residential areas and railway stations are both above 89%, close to 90%. The accuracy rates of churches and palaces are only 84% and 78%, which greatly lowers the overall classification accuracy. The reason is that the structure of the palace is similar to that of the church, and there is not much difference between the two internally.
[0100] In another example, a relevant ablation experiment was conducted to verify the effectiveness of the MCA attention mechanism in this embodiment. Specifically, using Efficient B0 as the backbone network, CBAM, CA, and MCA attention mechanisms were added to it for comparison. The experimental results are shown in Table 4 below.
[0101] Table 4
[0102]
[0103] As can be seen from Table 4, for the UC Merced Land-Use dataset, since its classification accuracy has reached saturation, the accuracy rate has almost no change after adding CBAM or CA, but it has been significantly improved after adding MCA. For the AID dataset, the accuracy rate has been slightly improved after adding CBAM or CA, but there is still a significant gap compared to adding MCA. For the NWPU-RESISC45 dataset, the accuracy rate has almost no change after adding CBAM, and there is a slight improvement of 0.22% after adding CA, but it has been improved by 0.72% after adding MCA. The results of the above ablation experiments also show that the MCA attention module of this embodiment does have certain advantages over traditional CBAM and CA.
[0104] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.
Claims
1. A remote sensing image scene classification method based on an efficient network with MCA attention mechanism, characterized by: The following steps are involved: S1. Obtain a remote sensing image dataset and divide it into training samples and test samples in proportion; S2, preprocessing the remote sensing image dataset, including cropping, data enhancement and standardization; S3. Add maximum pooling to coordinate attention CA to build MCA attention module; S4, insert the MCA attention module after the last convolutional layer of the EfficientNet B0 convolutional neural network and before the average pooling and fully connected layers to build the MCA-EfficientNet B0 network model; S5. Load the pre-trained weights on the ImageNetlK dataset into the MCA-EfficientNet B0 network model, then input the training samples into the MCA-EfficientNet B0 network model, continuously optimize the weighted cross entropy loss function through the back propagation algorithm, adjust the network parameters, and finally obtain the trained network model; S6. Input the test sample into the trained network model to obtain the remote sensing image scene classification result.
2. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 1 is characterized by: In the step S1, the three remote sensing image scene classification datasets, UC Merced Land-Use, AID and NWPU-RESISC45, are divided into training samples and test samples according to the ratios of 8:2, 5:5 and 2:8 respectively.
3. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 1 is characterized in that: In the step S2, the images in the remote sensing image dataset are cropped to a resolution of 224×224, and then the images are horizontally flipped with a probability p, and finally the image data is standardized.
4. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 3 is characterized by: In step S2, the probability p is set to 0.
5.
5. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 1 is characterized by: In step S3, in the coordinate information embedding stage, for the input X, firstly, average pooling and maximum pooling encoding operations are performed on each channel along the horizontal coordinate and the vertical coordinate, and then 1×1 convolution and ReLu activation function are used to process each channel independently; finally, the results of the two pooling operations are added element by element; therefore, the output of the cth channel with a height of h is: Similarly, the output of the cth channel with width w is: Among them, x c (h, i) represents the feature of the cth channel with height h, x c (j, w) represents the feature of the cth channel with a width of w; W represents the width of the feature map, H represents the height of the feature map; F1 is the convolution operation; δ is the activation function.
6. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 5 is characterized by: In step S3, in the coordinate attention generation stage, the outputs of the two transformations are concatenated, and then the concatenated features are transformed using a 1×1 convolution transformation function F1: f=δ(F1([z h ,z w ])) Among them, [·,·] represents the concatenation operation along the spatial dimension; z h is the feature in the vertical direction; z w is the feature in the horizontal direction; δ represents the nonlinear activation function; f represents the intermediate feature map that encodes the spatial information in the horizontal and vertical directions; r is the reduction rate in the SE block, and the number of channels of f can be increased or decreased by controlling r; Then, decompose f into two tensors along the spatial dimension and Where C / r represents the number of channels after the reduction using SE blocks, is a real vector space, f h is the tensor in the vertical direction, f w is the tensor in the horizontal direction; Then use another two 1×1 convolution transformations F h and F w Separately put f h and f w Transformed into a tensor with the same number of channels as the input X, we get: g h =σ(F h (f h )) g w =σ(F w (f w )) Among them, σ is the Sigmoid activation function, g h and g w They are C×H×1 and C×1×W feature maps respectively; they are then expanded and used as attention weights; the final output Y is expressed as: Among them, x c (i, j) is the input of the cth channel, is the attention weight in the vertical direction, is the attention weight in the horizontal direction.
7. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 1 is characterized in that: The step S5 specifically includes the following sub-steps: S5.1, extract shallow information through 3×3 convolution layer; S5.2, extract deep features through 7 MBConv modules in sequence; S5.3, through a 3×3 convolutional layer, further extract features; S5.4, aggregate key information through MCA attention module; S5.
5. Finally, the classification results of the training samples are obtained through average pooling and fully connected layers.
8. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 2 is characterized by: In step S1, the UC Merced Land-Use dataset contains remote sensing images of 21 scene categories, including airplanes, beaches, baseball fields, and buildings. Each scene category includes 100 images of 256×256 pixels, totaling 2100 images, and the spatial resolution of the images is 0.3 m.
9. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 2 is characterized by: In step S1, the AID data set contains remote sensing images of 30 scene categories, including airports, bridges, churches, and densely populated residential areas. Each scene category includes 220 to 420 images of 600×600 pixels, totaling 10,000 images, and the spatial resolution of the images varies between 0.5 and 8 meters.
10. The remote sensing image scene classification method based on the MCA attention mechanism efficient network according to claim 2 is characterized by: In step S1, the NWPU-RESISC45 dataset contains remote sensing images of 45 scene categories, each scene category includes 700 images of 256×256 pixels, totaling 31,500 images, and the spatial resolution of the images varies between 0.2 and 30 meters.
Citation Information
Cited By
Urban management case classification method based on images
CN120656002A
An image-based urban management case classification method
CN120656002B
