Remote sensing image scene classification method and device based on compression domain, and medium

Abstract

This application provides a method, device, and medium for scene classification of remote sensing images based on the compression domain, relating to the field of remote sensing images. The method includes: constructing an unsupervised scene classification model on the compression domain of a remote sensing image, including a multi-scale compression module, a multi-scale decompression module, and a classification module; inputting the remote sensing image to be classified into the multi-scale image compression module for compression to obtain compression domain information at different levels; inputting the compression domain information into the multi-scale decompression module to obtain first decoded feature information; and inputting the first decoded feature information into the classification module for scene classification to obtain the classification result of the remote sensing image. By classifying scene of remote sensing images based on compression domain information, the computational resources and time overhead of re-decompressing images required in remote sensing image applications are reduced.

Description

Technical Field

[0001] This application relates to the field of remote sensing images, and in particular to a method, device and medium for classifying remote sensing image scenes based on compressed domain. Background Technology

[0002] In the development of image compression technology, two main categories of methods can be clearly distinguished: traditional compression techniques and deep learning-based compression methods. Traditional compression methods, such as JPEG, PNG, JPEG2000, and JPEGXL, have been widely applied in various scenarios (see Roelofs G. PNG Lossless Image Compression. Lossless Compression Handbook, 2002: 371-390). These techniques typically achieve high compression ratios and effectively reduce storage space requirements. However, these methods may be insufficient in processing extremely large datasets or scenarios requiring rapid processing. On the other hand, deep learning-based compression methods, such as the L3C model, represent a new direction in compression technology. These methods utilize the powerful learning capabilities of deep neural networks, not only compressing data while maintaining image quality but also significantly improving processing speed. Especially in applications requiring direct image analysis or other advanced processing in the compression domain, deep learning-based methods can provide faster processing speeds, thereby reducing time overhead and improving overall efficiency. Therefore, when choosing a suitable compression method, we need to decide based on the actual application requirements and environment. If the focus is on reducing storage space, traditional methods may be a better choice; however, in scenarios with high demands for processing speed and immediacy, deep learning-based compression techniques demonstrate unique advantages. This comparative analysis in Reference 2 (Q. Zhu, Y. Zhong, B. Zhao, G. Xia, and L. Zhang, “Bag-of-visualwords scene classifier with local and global features for high spatial resolution remote sensing imagery,” IEEE Geosci. Remote Sens. Lett., vol. 13, no. 6, pp. 747–751, Jun. 2016.) helps us better understand and select suitable compression strategies to address different technical challenges and business needs.

[0003] While remote sensing image compression shares significant similarities with conventional image compression in many techniques and objectives, the specific needs of remote sensing images have also given rise to some unique technical considerations. Remote sensing images typically possess higher resolution and dynamic range, and often contain multispectral or hyperspectral data, meaning their data volume far exceeds that of conventional images, thus placing higher demands on the efficiency and effectiveness of compression algorithms. Although traditional compression techniques and deep learning-based methods perform well in the field of conventional image compression, these techniques require specific adjustments and optimizations when applied to remote sensing images. For example, considering the multi-band characteristics of remote sensing images, the compression algorithm in reference 3 (Y. Chen, H. Jiang, C. Li, X. Jia, and P. Ghamisi, “Deep feature extraction and classification of hyperspectral images based on convolutional neural networks,” IEEE Trans. Geosci. RemoteSens., vol. 54, no. 10, pp. 6232–6251, Oct. 2016.) needs to be able to effectively process and preserve key information between bands, achieving compression without sacrificing image quality. For three-channel remote sensing images, the processing is more similar to that of ordinary RGB images, allowing for the direct application of conventional image compression techniques. Furthermore, by utilizing deep learning-based compression methods, we can train these models on remote sensing image datasets to better adapt to the characteristics of remote sensing images, thereby improving compression efficiency and image restoration quality. This approach not only enhances the versatility of compression techniques but also brings a new direction to the field of remote sensing image compression.

[0004] In remote sensing image processing, due to the massive size of image data, storing this data usually requires compression. However, when this compressed data needs to be used in a specific application, it must first be decompressed, a process that not only consumes a lot of computing resources but also takes up valuable time.

[0005] To address this issue, an effective strategy is to process image data without fully decompressing it, significantly reducing the demands on computing resources and time. This method allows for the direct extraction of necessary information from compressed data for analysis and processing, avoiding the overhead of decompressing the entire dataset. This not only optimizes the processing flow but also greatly improves the efficiency of remote sensing image data applications, enabling the processing of more data in a shorter time and thus maximizing its value in various remote sensing applications. This strategy opens up new possibilities for the efficient utilization of remote sensing images and demonstrates the innovative prospects of the integration of information technology and remote sensing technology in the era of big data.

[0006] The core objective of scene classification is to assign class labels to remote sensing images (RS) based on feature descriptors that describe the semantic content of the images. Over the past decade, although various handcrafted features have been introduced into the research literature in the field of remote sensing, these features have proven inefficient, time-consuming, and computationally complex when processing large archives of remote sensing images containing petabytes of data. Furthermore, these handcrafted features often have weak discriminative power and rely on further processing using artificial intelligence techniques, which increases labor costs.

[0007] To overcome these limitations, deep learning (DL) algorithms, especially convolutional neural networks (CNNs), have been widely adopted in remote sensing, as seen in reference 4 (X. Chen, S. Xiang, C.-L. Liu, and C.-H. Pan, “Vehicle detection in satellite images by hybrid deepconvolutional neural networks,” IEEE Geosci. Remote Sens. Lett., vol. 11, no. 10, pp. 1797–1801, Oct. 2014.). CNNs learn the inherent patterns of images layer by layer and acquire high-level semantic content through multiple convolution and pooling operations, thus achieving higher classification accuracy. Studies have shown that CNNs perform exceptionally well in various remote sensing applications, including scene classification. Initially, CNN models typically required training from scratch using a large amount of training data. However, recent research has demonstrated that applying pre-trained deep learning networks (such as ImageNet, VGG16, GoogLeNet, and CaffeNet) in remote sensing can significantly improve classification performance. These pre-trained networks typically outperform models trained from scratch in terms of classification accuracy, especially when training samples are limited. They effectively utilize the prior knowledge of the pre-trained models, reducing the need for training resources and improving classification accuracy. However, even so, most remote sensing image scene classifications require complete image decompression, making methods based on compression domain information for direct classification extremely rare. Summary of the Invention

[0008] The purpose of this invention is to address the problem that existing remote sensing image scene classification methods are difficult to directly classify based on compressed domain information, and to provide a remote sensing image scene classification method, device, and medium based on compressed domain.

[0009] The above-mentioned objective of this application is achieved through the following technical solution: S1: Construct a scene classification model in the unsupervised remote sensing image compression domain, including: a multi-scale compression module, a multi-scale decompression module, and a classification module; input the remote sensing image to be classified into the multi-scale image compression module for compression to obtain compression domain information at different levels; S2: Input the compressed domain information into the multi-scale decompression module to obtain the first decoded feature information; S3: Input the first decoded feature information into the classification module to perform scene classification and obtain the classification result of the remote sensing image.

[0010] Optionally, step S1 includes: S11: The multi-scale compression module includes three compression layers: a first compression layer, a second compression layer, and a third compression layer; each compression layer includes a convolutional layer and a feature learning module; the feature learning module includes eight residual blocks and a convolutional layer. S12: Downsample the remote sensing image and input it into the first compression layer to obtain the first extracted features; quantize the first extracted features to obtain the first compressed domain information and the first unquantized features; S13: Input the first unquantized feature into the second compression layer to obtain the second compression domain information and the second unquantized feature; S14: Input the second unquantized feature into the third compression layer to obtain the third compression domain information; the compression domain information at different levels includes: the first compression domain information, the second compression domain information, and the third compression domain information.

[0011] Optionally, step S2 includes: S21: The multi-scale decompression module includes three decompression layers and a feature fusion module. The three decompression layers are: a first decompression layer, a second decompression layer, and a third decompression layer. Each decompression layer includes: a convolutional layer, a feature learning module, and an upsampling module. The decoded feature information includes: first decoded feature information, second decoded feature information, and third decoded feature information. S22: Introduce a normal distribution, and input the normal distribution and the third compression domain information into the third decompression layer to obtain the third decoding feature information; S23: Generate a priori probability distribution based on the third decoded feature information; S24: Input the prior probability distribution and the second compression domain information into the second decompression layer to obtain the second decoding feature information; S25: The third decoded feature information and the second decoded feature information are fused together by the feature fusion module to obtain the first fused feature; S26: Input the first compression domain information and the first fusion feature into the first decompression layer to obtain the first decoding feature information.

[0012] Optionally, step S3 includes: S31: The classification module includes: a reconstruction module and a classification network; S32: Input the first decoded feature information into the reconstruction module to obtain the reconstructed features; S33: Input the reconstructed features into the classification network to classify the scene and obtain the classification results of the remote sensing image.

[0013] Optionally, the reconstruction module includes: three convolutional layers, two ReLU activation layers, and one Sigmoid activation layer.

[0014] Optionally, the classification network uses pre-trained models of AlexNet and ShuffleNet.

[0015] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a remote sensing image scene classification method based on a compressed domain.

[0016] A computer-readable storage medium storing instructions that, when executed, perform a remote sensing image scene classification method based on a compressed domain.

[0017] The beneficial effects of the technical solution provided in this application are: Compared with existing technologies, the advantages of this invention are: This application constructs an unsupervised scene classification model in the compressed domain of remote sensing images. By classifying scenes in remote sensing images based on compressed domain information, it reduces the computational resources and time overhead required for recompressing images in remote sensing image applications. The model exhibits excellent generalization ability, demonstrating good performance in both general image domains and remote sensing image domains. Attached Figure Description

[0018] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart illustrating the steps of the remote sensing image scene classification method based on the compressed domain in the embodiments of this application; Figure 2 This is a network model structure diagram of the remote sensing image scene classification method based on the compressed domain in the embodiments of this application; Figure 3 This is a schematic diagram of the electronic device structure of the remote sensing image scene classification method based on the compressed domain in the embodiments of this application. Detailed Implementation

[0019] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0020] Embodiments of this application provide a method for remote sensing image scene classification based on compressed domain.

[0021] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a remote sensing image scene classification method based on the compressed domain, as described in an embodiment of this application, including: S1: Construct a scene classification model in the unsupervised remote sensing image compression domain, including: a multi-scale compression module, a multi-scale decompression module, and a classification module; input the remote sensing image to be classified into the multi-scale image compression module for compression to obtain compression domain information at different levels; Step S1 includes: S11: The multi-scale compression module includes three compression layers: a first compression layer, a second compression layer, and a third compression layer; each compression layer includes a convolutional layer and a feature learning module; the feature learning module includes eight residual blocks and a convolutional layer. Specifically, the Feature Learning Module (FLM) consists of eight ResBlocks and one convolutional layer. Each ResBlock connects two convolutional layers through a residual structure.

[0022] S12: Downsample the remote sensing image and input it into the first compression layer to obtain the first extracted features; quantize the first extracted features to obtain the first compressed domain information and the first unquantized features; S13: Input the first unquantized feature into the second compression layer to obtain the second compression domain information and the second unquantized feature; Specifically, the process proceeds to the Feature Learning Module (FLM) in the second compression layer. The FLM structure of the current layer is similar to the previous layer; this design allows for better feature extraction. After feature extraction via FLM, these features are also quantized to obtain the compressed domain information for the current layer. Through this layer-by-layer feature extraction and quantization process, the compressed domain information for each layer is obtained.

[0023] S14: Input the second unquantized feature into the third compression layer to obtain the third compression domain information; the compression domain information at different levels includes: the first compression domain information, the second compression domain information, and the third compression domain information.

[0024] Specifically, the multi-scale compression module includes the following sub-modules: a feature learning module, consisting of eight residual blocks (ResBlocks) and one convolutional layer. Residual blocks are a structure in deep neural networks that effectively alleviates the vanishing gradient problem, making the training of deeper networks more stable. Each residual block includes two convolutional layers and a skip connection, allowing the input to be directly added to the output, preserving more original information. The eight residual blocks are concatenated to progressively extract and enhance image features. Subsequently, the convolutional layer integrates the outputs of the residual blocks to further optimize the feature map. This design fully utilizes the advantages of deep networks, improves feature extraction efficiency, and provides richer feature representations for subsequent classification tasks. The quantization module is used to quantize the features generated by the multi-scale compression module. The quantization process converts continuous feature values ​​into discrete values, reducing the number of bits for storage and transmission, thereby improving the system's compression efficiency. The values ​​are discretized into 25 discrete values ​​between -1 and 1. The design of the quantization module effectively reduces data redundancy while preserving the representational power of the features, ensuring the accuracy of subsequent decompression and classification.

[0025] S2: Input the compressed domain information into the multi-scale decompression module to obtain the first decoded feature information; Step S2 includes: S21: The multi-scale decompression module includes three decompression layers and a feature fusion module. The three decompression layers are: a first decompression layer, a second decompression layer, and a third decompression layer. Each decompression layer includes: a convolutional layer, a feature learning module, and an upsampling module. The decoded feature information includes: first decoded feature information, second decoded feature information, and third decoded feature information. Specifically, the module consists of a decompression layer (Feature Learning Module, FLM, with the same structure as the multi-scale compression module) and an upsampling module. The upsampling module comprises a convolutional layer and a pixelShuffle layer. After processing by the decoder, the decoded feature information of the current layer can be obtained. Unlike the encoding order in the multi-scale compression module (from the first compression layer to the third compression layer), the decoding order in the multi-scale decompression module is exactly the opposite, from the bottom layer (the third decompression layer) to the top layer (the first decompression layer).

[0026] S22: Introduce a normal distribution, and input the normal distribution and the third compression domain information into the third decompression layer to obtain the third decoding feature information; S23: Generate a priori probability distribution based on the third decoded feature information; S24: Input the prior probability distribution and the second compression domain information into the second decompression layer to obtain the second decoding feature information; S25: The third decoded feature information and the second decoded feature information are fused together by the feature fusion module to obtain the first fused feature; S26: Input the first compression domain information and the first fusion feature into the first decompression layer to obtain the first decoding feature information.

[0027] Specifically, the multi-scale decompression module includes the following sub-modules: The feature learning module in the multi-scale decompression module is similar to that in the multi-scale compression module. This symmetrical structure design helps the system better restore and recover image features. Specifically, this symmetrical structure ensures the alignment and matching of feature information during compression and decompression, thereby maximizing the preservation of the original image information and details and improving the quality of the decompressed image.

[0028] The upsampling module, consisting of convolutional layers and pixelShuffle layers, is responsible for upsampling low-resolution feature images to high resolution. The convolutional layers are used for feature extraction and transformation, adjusting the number of channels and spatial distribution of the input feature map through learned weights. The pixelShuffle layers rearrange the low-resolution feature maps output by the convolutional layers into a high-resolution image, thereby improving spatial resolution. This design effectively enhances image details and avoids the blurring and artifacts caused by traditional upsampling methods (such as nearest neighbor interpolation or bilinear interpolation). Simultaneously, this combination better preserves and restores image details, ensuring that the feature maps received by the classification network contain more high-frequency information, thus improving classification accuracy.

[0029] The feature fusion module comprises ResBlocks (residual module), convolutional layers, and an upsampling module. This design aims to learn more deeply from the coarse feature information before feature fusion, then use upsampling to match the feature dimensions and fuse them with the feature information from the previous layer. The fusion strategy here uses tensor channel concatenation, followed by a convolutional layer to adjust the number of channels and learn information from both feature channels. The introduction of this module not only achieves better performance in lossless compression but also yields superior results in classification tasks.

[0030] S3: Input the first decoded feature information into the classification module to perform scene classification and obtain the classification result of the remote sensing image.

[0031] Step S3 includes: S31: The classification module includes: a reconstruction module and a classification network; S32: Input the first decoded feature information into the reconstruction module to obtain the reconstructed features; S33: Input the reconstructed features into the classification network to classify the scene and obtain the classification results of the remote sensing image.

[0032] The reconstruction module includes: three convolutional layers, two ReLU activation layers, and one Sigmoid activation layer.

[0033] Specifically, the reconstruction module contains three convolutional layers, nested with PReLU and sigmoid activation functions. This design allows for further recovery and enhancement of the decompressed features. The PReLU activation function introduces non-linear features while avoiding the "neuron death" problem that may occur with the traditional ReLU activation function. The sigmoid activation function is used in the output layer, compressing the output values ​​to between 0 and 1, thereby enhancing the contrast and detail of the feature map and providing more accurate input for subsequent classification tasks.

[0034] The classification network uses pre-trained models of AlexNet and ShuffleNet.

[0035] Specifically, the classification network is responsible for classifying the reconstructed features. Pre-trained models from AlexNet and ShuffleNet are used; these models have been thoroughly trained on large datasets and offer efficient feature extraction and classification performance. AlexNet excels at handling high-resolution images and boasts high classification accuracy; ShuffleNet, on the other hand, is characterized by its high computational efficiency and low resource consumption, making it suitable for resource-constrained scenarios. By combining these two models, system efficiency can be improved while maintaining classification accuracy.

[0036] Specifically, the system comprises the following modules: a multi-scale compression module, a multi-scale decompression module, and a multi-scale decompression module. The multi-scale decompression module uses the compressed domain information generated by the multi-scale compression module to recover features, resulting in three levels of decompressed features. It also generates a prior distribution based on the current features as input to the next level. The classification module consists of feature reconstruction and classification. First, the decompressed features are further recovered and enhanced to ensure feature integrity and adaptability to subsequent classification tasks. Then, a pre-trained classification network provided by the PyTorch library is used for classification. These pre-trained models have been trained on datasets with millions of records, achieving optimal classification performance. Through the collaborative work of these modules, the overall system can efficiently compress and decompress remote sensing images, ensuring feature accuracy and integrity, thus providing a foundation for efficient scene classification. The design of the multi-scale feature extraction and reconstruction module enables the system to capture details and global information in the image, combined with the pre-trained classification network, to achieve efficient and accurate scene classification.

[0037] To illustrate the effectiveness of this invention, the following specific datasets are provided for experimental comparison.

[0038] The experiments used the NWPU-RESISC45 and AID datasets for validation. These two datasets are widely used in remote sensing image scene classification research, providing abundant training and testing samples, and are crucial for evaluating and improving remote sensing image classification algorithms. This section will verify two effects of the invention.

[0039] First, there's the compression ratio. Bpsp effectively reflects the model's lossless compression efficiency; the smaller the value, the better the compression effect. The model not only compares its compression performance on remote sensing datasets but also achieves superior performance on traditional images, demonstrating its excellent compression capabilities.

[0040] Secondly, there's classification accuracy, which is the most widely used and objective indicator for evaluating classification methods. A higher value indicates better classification performance. The formula for calculating classification accuracy is: , The number of test samples, This indicates whether the predicted label of the i-th sample is correct. Here, the accuracy of classification based on different proportions of information in the decompressed domain is compared with that of the original image classification. Tables 2 and 3 show the experimental results using NWPU-RESISC45 and AID as datasets, respectively.

[0041] Meanwhile, the data compression of this method on the two datasets is shown in Table 4. It can be seen that it can achieve a result close to or even surpassing the original image using less than half the amount of data in the original image.

[0042] Table 1. Comparison of compression performance of this method and existing technologies in different databases.

[0043] Table 2. Classification accuracy of the present invention on the NWPU-RESISC45 dataset.

[0044] Table 3. Classification accuracy of the present invention on the AID dataset

[0045] Table 4. Data volume of this invention based on compressed domain information application

[0046] In Table 1, Open Images, DIV2K, RAISE-1K, Koda, NWPU-RESISC45, and AID are datasets, while PNG, WebP, and L3C are three models. The values ​​are the compression performance (bpsp) metrics of the corresponding models on their respective datasets.

[0047] In Table 2, AlexNet and ShuffleNet are two classification networks. Original image as input, Z1, Z2, Z3 as input, Z2, Z3 as input, and Z3 as input represent the classification results when the original image is used as input and Z1, Z2, Z3 (information from the compressed domain) are used as network inputs (NWPU-RESISC45 dataset).

[0048] In Table 3, AlexNet and ShuffleNet are two classification networks. Original image as input, Z1, Z2, Z3 as input, Z2, Z3 as input, and Z3 as input represent the classification results when the original image is used as input and Z1, Z2, Z3 (information from the compressed domain) are used as network inputs (AID dataset).

[0049] The Data Volume (%) in Table 4 represents the amount of data. For example, Z1, Z2, and Z3 represent 35.06% of the information in the original image (on the NWPU-RESISC45 dataset), and so on. This application also discloses an electronic device. (See reference...) Figure 3 , Figure 3This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0050] The communication bus 502 is used to enable communication between these components.

[0051] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0052] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0053] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the aforementioned method for classifying remote sensing image scenes based on a compressed domain.

[0054] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will readily conceive of those skilled in the art upon consideration of the specification and the disclosure of practical truths.

[0055] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A remote sensing image scene classification method based on compressed domain, characterized in that, The method includes the following steps: S1: Construct a scene classification model in the unsupervised remote sensing image compression domain, including: a multi-scale compression module, a multi-scale decompression module, and a classification module; input the remote sensing image to be classified into the multi-scale image compression module for compression to obtain compression domain information at different levels; Step S1 includes: S11: The multi-scale compression module includes three compression layers: a first compression layer, a second compression layer, and a third compression layer; each compression layer includes a convolutional layer and a feature learning module; the feature learning module includes eight residual blocks and a convolutional layer. S12: Downsample the remote sensing image and input it into the first compression layer to obtain the first extracted features; quantize the first extracted features to obtain the first compressed domain information and the first unquantized features; S13: Input the first unquantized feature into the second compression layer to obtain the second compression domain information and the second unquantized feature; S14: Input the second unquantized feature into the third compression layer to obtain the third compression domain information; the compression domain information at different levels includes: the first compression domain information, the second compression domain information, and the third compression domain information; S2: Input the compressed domain information into the multi-scale decompression module to obtain the first decoded feature information; Step S2 includes: S21: The multi-scale decompression module includes three decompression layers and a feature fusion module. The three decompression layers are: a first decompression layer, a second decompression layer, and a third decompression layer. Each decompression layer includes: a convolutional layer, a feature learning module, and an upsampling module. The decoded feature information includes: first decoded feature information, second decoded feature information, and third decoded feature information. S22: Introduce a normal distribution, and input the normal distribution and the third compression domain information into the third decompression layer to obtain the third decoding feature information; S23: Generate a priori probability distribution based on the third decoded feature information; S24: Input the prior probability distribution and the second compression domain information into the second decompression layer to obtain the second decoding feature information; S25: The third decoded feature information and the second decoded feature information are fused together by the feature fusion module to obtain the first fused feature; S26: Input the first compression domain information and the first fusion feature into the first decompression layer to obtain the first decoding feature information; S3: Input the first decoded feature information into the classification module to perform scene classification and obtain the classification result of the remote sensing image.

2. The remote sensing image scene classification method based on compressed domain as described in claim 1, characterized in that, Step S3 includes: S31: The classification module includes: a reconstruction module and a classification network; S32: Input the first decoded feature information into the reconstruction module to obtain the reconstructed features; S33: Input the reconstructed features into the classification network to classify the scene and obtain the classification results of the remote sensing image.

3. The remote sensing image scene classification method based on compressed domain as described in claim 2, characterized in that, The reconstruction module includes: three convolutional layers, two ReLU activation layers, and one Sigmoid activation layer.

4. The remote sensing image scene classification method based on compressed domain as described in claim 2, characterized in that, The classification network uses pre-trained models of AlexNet and ShuffleNet.

5. An electronic device, characterized in that, The device includes a processor (501), a memory (505), a user interface (503), and a network interface (504). The memory (505) is used to store instructions. The user interface (503) and the network interface (504) are used to communicate with other devices. The processor (501) is used to execute the instructions stored in the memory (505) to cause the electronic device to perform the method as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the method as described in any one of claims 1-4.

Images