Gray scale guide-based multispectral remote sensing image compression method and system

Abstract

The present application provides a multispectral remote sensing image compression method and system based on gray guide, which can effectively reduce the data redundancy of multispectral images and improve the compression quality and multiple. The method comprises the following steps: step 1, data preparation; step 2, building a model for training; step 2-1, performing spectral and spatial down-sampling; step 2-2, extracting spatial features and spectral features; step 2-3, taking the spatial features of band 1 as a guide to predict the gray images of other bands, then taking the spectral features of the spatial down-sampled images as a constraint to make the spectral features of the predicted multispectral images consistent with the original image block; subtracting the predicted image from the original image block to obtain a residual error; step 2-4, encoding and decoding the residual error, and encoding and decoding the spatial feature and spectral feature latent variables; step 3, obtaining a multispectral remote sensing image compression model after training; step 4, inputting the remote sensing image to be compressed into the multispectral remote sensing image compression model for compression.

Description

Technical Field

[0001] This invention belongs to the field of image data processing technology, specifically relating to a method and system for compressing multispectral remote sensing images based on grayscale guidance. Background Technology

[0002] With the continuous development of remote sensing technology, high-resolution multispectral remote sensing imagery has been widely applied to solving various practical problems, such as target detection, disaster early warning, and traffic monitoring. These problems typically require remote sensing satellites to acquire remote sensing data with wide swaths and high spatial resolution. However, timely transmission of images captured by remote sensing satellites to the ground has become a key technical issue in expanding the application scenarios of remote sensing data. Because the bandwidth of the satellite-to-ground transmission link is smaller than that of the ground, and multispectral remote sensing imagery has more bands than typical natural images, the redundant information between multiple bands leads to a surge in the amount of data that needs to be transmitted. Therefore, a novel high-compression-ratio multispectral remote sensing image compression method is urgently needed. Image predictive coding is a lossless compression technique based on image pixel prediction, which can achieve efficient compression by predicting the spatial and temporal domains of the image.

[0003] Predictive coding in the spatial domain is commonly used for still image compression. It utilizes the statistical properties of pixel values ​​and the correlation between neighboring pixels to compress the original image into a smaller data volume for easier storage and transmission. Predictive coding methods include pixel-based linear predictive coding and template-based nonlinear predictive coding.

[0004] Temporal compression primarily refers to inter-frame compression of video, and predictive coding is a crucial technique. The earliest video compression standard, H.261, employed motion estimation-based predictive coding, achieving a high compression ratio by estimating and encoding motion information between frames in an image sequence. Subsequently, with the continuous upgrading of video coding standards, predictive coding technology has further developed. Standards such as MPEG-1 and MPEG-2 use a combination of intra-frame and inter-frame prediction, utilizing spatial correlation within frames and temporal correlation between frames to achieve even higher compression efficiency. However, with the continuous increase in video resolution and bitrate, traditional predictive coding methods are no longer sufficient to meet the demands of modern video compression.

[0005] However, current research on multispectral remote sensing image compression is limited. Commonly used methods are based on sparse representation and principal component analysis (PCA), which suffer from insufficient compression ratios and information loss. Neural network-based methods, on the other hand, use deep learning techniques to construct nonlinear models for prediction and compression of multispectral images. These methods can achieve better compression results and reconstruction quality at higher compression ratios. However, neural network-based multispectral image compression methods still have some problems. Because remote sensing images have more bands than typical natural images, they involve greater computational load and training complexity, requiring large amounts of training data and computational resources, and the models have poor interpretability. Furthermore, these methods require further improvement and optimization to enhance their efficiency and performance in practical applications. Summary of the Invention

[0006] This invention is made to solve the above-mentioned problems, and aims to provide a method and system for compressing multispectral remote sensing images based on grayscale guidance, which can effectively reduce data redundancy in multispectral images and improve compression quality and compression ratio.

[0007] To achieve the above objectives, the present invention adopts the following solution: This invention provides a grayscale-guided multispectral remote sensing image compression method, comprising the following steps: Step 1, Data preparation: Based on the multispectral remote sensing image to be compressed, acquire historical multispectral remote sensing images from the same source, and crop the historical multispectral remote sensing images for model construction. Each historical multispectral remote sensing image is cropped into multiple image patches. Step 2: Build a compressed model for training; Step 2-1: Perform spectral and spatial downsampling on each image patch: extract the grayscale image of band 1 of the image patch to obtain the spectrally downsampled image. Band 1 is the first band with the longest imaging band in the image. Perform spatial downsampling on the image patch to obtain the spatially downsampled image. Step 2-2: Extract spatial and spectral features for each image patch: extract spatial features from the spectral downsampled image, and extract spectral features from the spatial downsampled image; Steps 2-3: For each image patch, guided by the spatial features of band 1, the grayscale images of other bands are predicted. Then, the spectral features of the spatially downsampled image are used as constraints to ensure that the spectral features of the predicted multispectral image are consistent with those of the original image patch. The predicted image is subtracted from the original image patch to obtain the residual. During decoding, the predicted image is added to the residual to obtain the reconstructed multispectral image. This step is used in both the encoding and decoding stages. Steps 2-4 involve encoding and decoding the residuals, as well as encoding and decoding the latent variables of spatial and spectral features. Step 3: Set the loss function and model parameters for model training, and train the compressed model built in Step 2. After training, a multispectral remote sensing image compression model based on grayscale guidance is obtained. Step 4: Input the multispectral remote sensing image to be compressed into the multispectral remote sensing image compression model for compression.

[0008] Preferably, in the grayscale-guided multispectral remote sensing image compression method provided by the present invention, in step 2-2, a spatial feature extractor and a spectral feature extractor are constructed using a shallow convolutional neural network. The spatial feature extractor consists of a 3×3×1 convolutional kernel and is used to extract spatial features from the spectrally downsampled image. The spectral feature extractor consists of a 1×1×3 one-dimensional convolutional kernel and is used to extract spectral features from the spatially downsampled image. The present invention uses a 3×3×1 convolutional kernel to extract spatial features from the spectrally downsampled image and a 1×1×3 convolutional kernel to extract spectral features from the spatially downsampled image, which helps to extract spectral and spatial features separately and in a targeted manner.

[0009] Preferably, in step 2-2, the grayscale-guided multispectral remote sensing image compression method provided by the present invention utilizes a spatial feature extractor to extract spectral downsampled images. Spatial features obtained from U Using a spectral feature extractor to spatially downsample images Spectral characteristics obtained V ; U for tensor, U for tensor, h , w and c It is the original image patch Height, width, and number of channels.

[0010] Preferably, in the grayscale-guided multispectral remote sensing image compression method provided by the present invention, in steps 2-3, spatial features and spectral features are combined to obtain spatial-spectral joint features. F : ; right F Spatial upsampling Then in terms of spectral dimensions ( c +1) Divide the joint features of the upsampled data to obtain: ; In the formula, The first characteristic representing the joint feature i One channel; H This represents the features after upsampling; calculate and of F The inner product under the norm serves as the spatial similarity between two channels. : ; Composed of the joint features of the first channel and the second channel to the third channel c Spatial similarity vector of +1 channel , ; Guided by the grayscale image of band 1, spatial similarity vectors between bands are added. S As an attention mechanism, spatial-spectral joint features are fused: ; In the formula, The output fusion features, The joint features adopted above, This represents the convolution operation. This represents the Hadamard product operation, which ultimately fuses features. Input Image Generation Network G Prediction results for generating multispectral images at original resolution ; Prediction results With the original image patch By subtracting the difference, we obtain the residual (see diagram). .

[0011] Preferably, in the gray-scale guided multispectral remote sensing image compression method provided by the present invention, in steps 2-4, the encoding and decoding of the residual adopts an encoding and decoding method based on wavelet transform; the spatial features and spectral features are encoded and decoded respectively using entropy estimation based on Gaussian mixture model.

[0012] Preferably, in the grayscale-guided multispectral remote sensing image compression method provided by the present invention, in step 4, the multispectral remote sensing image to be compressed is image data that has been cropped in the same manner as in step 1.

[0013] Specifically, the grayscale-guided multispectral remote sensing image compression method provided by this invention, after constructing a multispectral remote sensing image compression model based on steps 1 to 3, allows all source-related multispectral remote sensing images to be directly compressed by inputting into the multispectral remote sensing image compression model in step 4, thereby achieving encoding and decoding.

[0014] Furthermore, the present invention also provides a grayscale-guided multispectral remote sensing image compression system capable of automatically implementing the above-described method of the present invention, comprising: The data preparation department acquires historical multispectral remote sensing images from the same source based on the multispectral remote sensing image to be compressed, and crops the historical multispectral remote sensing images for model construction. Each historical multispectral remote sensing image is cropped into multiple image patches. The training model building department is responsible for building compressed models for training; it includes the following modules: The downsampling module performs spectral and spatial downsampling on each image patch: it extracts the grayscale image of band 1 of the image patch to obtain a spectrally downsampled image, where band 1 is the longest imaging band in the image; and it performs spatial downsampling on the image patch to obtain a spatially downsampled image. The feature extraction module extracts spatial and spectral features for each image patch: spatial features are extracted from the spectrally downsampled image, and spectral features are extracted from the spatially downsampled image. For each image patch, the other band prediction module uses the spatial features of band 1 as a guide to predict the grayscale images of other bands. Then, it uses the spectral features of the spatial downsampled image as a constraint to make the spectral features of the predicted multispectral image consistent with the original image patch. The predicted image is subtracted from the original image patch to obtain the residual. During decoding, the predicted image is added to the residual to obtain the reconstructed multispectral image. The encoding / decoding module encodes and decodes the residuals, as well as the latent variables of spatial and spectral features; In the compression model construction section, the loss function and model parameters for model training are set, and the compression model built in step 2 is trained. After training, a multispectral remote sensing image compression model based on grayscale guidance is obtained. The compression unit inputs the multispectral remote sensing image to be compressed into the multispectral remote sensing image compression model for compression. The control unit is connected to the data preparation unit, training model building unit, compression model building unit, and compression unit, and controls their operation.

[0015] Preferably, the gray-scale-guided multispectral remote sensing image compression system provided by the present invention may further include: a testing module, used to perform gray-scale-guided multispectral remote sensing image testing and verification on the model trained by the compression model construction unit; inputting the multispectral image test data to be compressed into the model, using the gray-scale information of band 1 as guidance, outputting the encoded file and the decoded image, and testing the peak signal-to-noise ratio and structural similarity of the decoded image; if the test results do not meet the requirements, returning to the compression model construction unit for adjustment: based on the peak signal-to-noise ratio and structural similarity, for cases where the quality of the decoded image does not meet the requirements, increasing the number of training iterations, and increasing the weight value of the prediction error of the corresponding original image patch in the loss function. Continue training; compare file sizes before and after compression, and if the compression ratio does not meet the requirements, increase the weight value of the corresponding latent code entropy in the loss function. Continue training.

[0016] The grayscale-guided multispectral remote sensing image compression system provided by the present invention may further include: an input display unit, which is communicatively connected to the control unit, for allowing users to input operation commands and displaying them accordingly.

[0017] The grayscale-guided multispectral remote sensing image compression method and system provided by the present invention have the following advantages compared with the prior art: (1) The present invention can effectively remove spectral redundancy information in multispectral images.

[0018] (2) The present invention is highly robust to changes in the selected reference band and is applicable to the compression of multispectral remote sensing images in most scenarios.

[0019] (3) Compared with other deep learning compression methods, the residual prediction compression method adopted in this invention can achieve flexible rate distortion optimization by adjusting the quantization step size of the compressed residual.

[0020] (4) The present invention has advantages in high compression of multispectral remote sensing images. It can make full use of the spectral redundancy information of multispectral images to improve the compression ratio and decoding quality of images, as well as the clarity of multispectral images under high compression. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the construction of a compression model according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the shallow feature extractor and grayscale prediction process for other bands involved in an embodiment of the present invention. Figure 3 This is a diagram illustrating the compression results of multispectral remote sensing images involved in an embodiment of the present invention. Detailed Implementation

[0022] The following describes in detail, with reference to the accompanying drawings, the specific implementation scheme of the grayscale-guided multispectral remote sensing image compression method and system involved in this invention.

[0023] <Example 1> The grayscale-guided multispectral remote sensing image compression method involved in this embodiment has the following specific steps: I. Data preparation: The original wide-format remote sensing images are cropped, and the cropped image patches (small images) are used as training and testing data.

[0024] In this embodiment, each original image is cropped to generate 128×128 image blocks. After cropping all images, more than 10,000 image blocks are generated. The size of the image blocks depends on the performance of the graphics card used for training. When the graphics card has stronger performance and more video memory, 256×256 blocks can also be used.

[0025] In this embodiment, all the cropped data is randomly divided into training data and test data in a 9:1 ratio.

[0026] II. Figure 1 As shown, a compressed model is built for training.

[0027] The multispectral image covers multiple imaging bands, arranged from longest to shortest wavelength. The first band is selected as band 1, and the imaging result in this band is taken as the grayscale image of band 1. The spatial information based on the grayscale image of band 1 is used to predict and encode the grayscale information of other bands. Then, the predicted residual, the grayscale image of band 1, and the downsampled multispectral image are encoded and decoded.

[0028] This section includes: separating and downsampling the grayscale image of Band 1 to obtain a single-band grayscale image with the original spatial resolution and a low-resolution multispectral image with the original spectral dimensions; performing shallow feature extraction, including spatial and spectral feature extraction; predicting grayscale images of other bands by using a shallow feature extractor to obtain spatial and spectral features, and using a deep generative network to predict grayscale image data of bands other than Band 1; entropy encoding and decoding, encoding and decoding the prediction residual, and entropy encoding and decoding of shallow spatial and spectral features; extracting shallow features from the decoded grayscale image of Band 1 and the low-resolution multispectral image; predicting the original image using spatial and spectral features, and adding the prediction result to the decoded residual to obtain the reconstructed image of the original image.

[0029] (1) First, the shallow feature extraction and predictive coding in the embodiment are introduced: like Figure 2 As shown, spatial features are extracted from the grayscale image of band 1 using two shallow neural networks—a spatial feature extractor and a spectral feature extractor. U Extracting spectral features from downsampled multispectral images V The shallow neural network consists of layers CONV, CONV, POOL, CONV, CONV, POOL, and CONV, where CONV is a convolutional layer and POOL is a pooling layer. In this embodiment, the CONV convolutional kernel of the spatial feature extractor is set to 3×3×1, and the CONV convolutional kernel of the spectral feature extractor is set to 1×1×3.

[0030] Then, predictions are made for grayscale images in other bands: Combining spatial features with spectral features yields spatial-spectral joint features. F As shown in the following formula: ; right F Spatial upsampling Then in terms of spectral dimensions ( c +1) Divide the joint features of the upsampled data to obtain: ; In the formula, The first characteristic representing the joint feature i One channel. Calculation and of F The inner product under the norm serves as the spatial similarity between two channels. : ; Composed of the joint features of the first channel and the second channel to the third channel c Spatial similarity vector of +1 channel ,in .

[0031] Guided by the grayscale image of band 1, spatial similarity vectors between bands are added. S As an attention mechanism, spatial-spectral joint features are fused: ; In the formula, The output fusion features, The joint features adopted above, This represents the convolution operation. This represents the Hadamard product operation. Ultimately, it will fuse the features. Input Image Generation Network G Prediction results for generating multispectral images at original resolution .

[0032] Prediction results With the original image By subtracting, we obtain the residual plot. .

[0033] (2) For the residual plot Encode; Discrete wavelet transform is used to analyze the entire residual map. Perform transform coding, and set a low-pass filter for discrete wavelet transform. h and high-pass filterg for: ; For residual plot The discrete wavelet transform is expressed as: ; In the formula, Residual plot The j -1st floor, 2nd floor k + n The pixel value at each position. h and g For low-pass and high-pass filters, j , k The distribution represents the number and location of the decomposition. N The length of the wavelet filter. and These represent the low-frequency and high-frequency coefficients after decomposition, respectively.

[0034] The residual encoding file can be generated by quantizing and arithmetic encoding the transform coefficients. The compression factor of the residual map can be flexibly adjusted by adjusting the enhancement step size.

[0035] The inverse wavelet transform is used to recover the transform coefficients as the reconstructed residuals. Its formula can be expressed as: ; (3) Entropy coding of spatial and spectral features; Entropy coding and entropy decoding are used for grayscale images of band 1. Spatial features U and downsampled images spectral characteristics V Entropy encoding and decoding are performed. Latent variables are predicted using a Gaussian mixture model, thereby achieving accurate entropy estimation of the latent variables, which can be expressed by the following formula. and spatial features U The entropy encoding and entropy decoding process: ; In the formula, and These represent the grayscale image of band 1 and its entropy-decoded reconstructed image, respectively. U It is a spatial feature. and These represent latent variables and quantified latent variables, respectively, representing spatial characteristics. Spatial feature extractor Q Refers to the quantification process. and These represent the entropy encoder and entropy decoder, respectively. and These are their parameters; It is based on the entropy estimation of the reconstructed image distribution by latent variables, that is, the latent variables of a single-band grayscale image in quantization. Conditional distribution under the following conditions p This represents the conditional distribution of the variable.

[0036] Similarly, I lr and spectral characteristics V The entropy encoding and entropy decoding process can be represented as: ; In the formula, and These represent the downsampled image and its entropy-decoded reconstructed image, respectively. V It is a spectral feature. and These represent latent variables and quantified latent variables, respectively, representing spectral characteristics. Refers to a spectral feature extractor; It is based on the entropy estimation of the reconstructed image distribution by latent variables, that is, the quantized latent variables of the downsampled image. The conditional distribution.

[0037] 3. Set the loss function and model parameters for model training.

[0038] The prediction error of the original image (the cropped image patch, i.e., the training data), the reconstruction loss of the band 1 grayscale image, the reconstruction loss of the downsampled image, and the entropy of the latent variable entropy encoding are used as the loss function for the entire model training. The loss function formula in the example is as follows: ; In the formula, This represents the prediction error of the original image (the cropped image patch). The reconstruction loss of the band 1 grayscale image, To reduce the reconstruction loss of the sampled image, The entropy of the latent code, For the total loss function, , , These are, respectively, an image patch, a band 1 grayscale image, and a downsampled image. This represents the encoder's prediction of the original image. and They are respectively and Entropy decoding to reconstruct the image and For latent variables of the image, The parameters of the latent variable distribution, , , , They are respectively , , , The corresponding weighting coefficients are set as follows in the embodiment: , , , .

[0039] The training generations were set to 300, the batch size to 16, the learning rate to 0.0001, and the Adam adaptive learning rate optimization algorithm was selected as the optimization function.

[0040] Fourth, based on the compression model trained in step three, input the multispectral remote sensing image to be compressed to achieve the encoding and decoding of the multispectral remote sensing image with a high compression ratio.

[0041] The spectral and spatial downsampling module mentioned in step two is used to simply downsample the original image patches. Spatial and spectral features are separated, and a grayscale image containing spatial information is obtained by extracting the first band of the multispectral image and performing spectral downsampling. Low-resolution images containing spectral information are obtained by spatial downsampling using bicubic interpolation. .

[0042] The shallow feature extractor mentioned in step two is a shallow convolutional neural network, whose layers include CONV, CONV, POOL, CONV, CONV, POOL, and CONV in sequence, where CONV is a convolutional layer and POOL is a pooling layer. The CONV convolutional kernel of the spatial feature extractor is 3×3×1, and the CONV convolutional kernel of the spectral feature extractor is 1×1×3.

[0043] In the above process, for the multispectral remote sensing image to be compressed, a portion of historical data from the same source (taken by the same satellite) is used for training, and another portion of historical data is used for testing. After the test meets the requirements, all multispectral remote sensing images from this source can be directly compressed using the "trained grayscale-guided multispectral remote sensing image compression model" from step three. Only when the source changes does it become necessary to reacquire historical data and use steps two and three of the above scheme to train the model and conduct testing.

[0044] like Figure 3 As shown, to facilitate understanding of the technical effects of this invention, it is compared with the existing compression method BPG and the state-of-the-art neural network compression method CAE on high-resolution multi-mode satellite multispectral data (8-band multispectral imagery). The test results are detailed in Table I below: Table I. Comparison Results on High-Resolution Multi-Mode Remote Sensing Image Test Data Table I records the results of each model on the test data of multispectral remote sensing images from the Gaofen-Multi-Mode satellite.

[0045] according to Figure 3 Compared with Table I, the quality of compressed images is measured using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), with higher values ​​for both indicating clearer compressed images. It can be seen that the image quality of remote sensing image compression using this invention is significantly higher than traditional methods and other state-of-the-art compression methods employing neural networks. Compared to BPG, it reduces the bit rate by 35%, and compared to CAE, it reduces the bit rate by 60%.

[0046] <Example 2> This second embodiment provides a grayscale-guided multispectral remote sensing image compression system that can automatically implement the above-mentioned method of the present invention. The system includes a data preparation unit, a training model building unit, a compression model construction unit, a compression unit, an input display unit, and a control unit.

[0047] The data preparation department performs the steps described in step 1 above, acquiring historical multispectral remote sensing images from the same source based on the multispectral remote sensing image to be compressed, and cropping the historical multispectral remote sensing images for model building. Each historical multispectral remote sensing image is cropped into multiple image patches.

[0048] The training model building section performs the steps described in step 2 above to build a compressed model for training; including a downsampling module, a feature extraction module, a prediction module for other bands, and an encoding / decoding module.

[0049] The compression model construction department performs the steps described in step 3 above, sets the loss function and model parameters for model training, trains the compression model built in step 2, and obtains a grayscale-guided multispectral remote sensing image compression model after training.

[0050] The compression unit performs the steps described in step 4 above, inputting the multispectral remote sensing image to be compressed into the multispectral remote sensing image compression model for compression. The compression unit includes a testing module and a compression module. The testing module is used to perform grayscale-guided multispectral remote sensing image testing and verification on the model trained by the compression model construction unit. It inputs the multispectral image test data to be compressed into the model, using the grayscale information of band 1 as a guide, outputting a compressed encoded file and a decoded image, and testing the peak signal-to-noise ratio (PSNR) and structural similarity of the decoded image. If the test results do not meet the requirements, it returns to the compression model construction unit for adjustment: based on the PSNR and structural similarity, for cases where the decoded image quality does not meet the requirements, it increases the number of training iterations and increases the loss function. Adjust the weights and continue training; based on the comparison of file sizes before and after compression, if the compression ratio does not meet the requirements, increase the weights in the loss function. The weight values ​​are adjusted, and training continues until the test results meet the requirements, resulting in the final multispectral remote sensing image compression model. The compression module inputs the multispectral remote sensing image to be compressed into the final multispectral remote sensing image compression model for compression.

[0051] The input display unit is used to allow users to input operation commands and to display the input, output, and intermediate processing data of the corresponding unit or module according to the operation commands.

[0052] The control unit is connected to the data preparation unit, training model building unit, compression model building unit, compression unit, and input display unit, and controls their operation.

[0053] The above embodiments are merely illustrative examples of the technical solutions of the present invention. The grayscale-guided multispectral remote sensing image compression method and system involved in this invention are not limited to the content described in the above embodiments, but are defined by the scope of the claims. Any modifications, additions, or equivalent substitutions made by those skilled in the art based on these embodiments are within the scope of protection claimed by the claims of this invention.

Claims

1. A multispectral remote sensing image compression method based on grayscale guidance, characterized in that, Includes the following steps: Step 1, Data preparation: Based on the multispectral remote sensing image to be compressed, acquire historical multispectral remote sensing images from the same source, and crop the historical multispectral remote sensing images for model construction. Each historical multispectral remote sensing image is cropped into multiple image patches. Step 2: Build a compressed model for training; Step 2-1: Perform spectral and spatial downsampling on each image patch: extract the grayscale image of band 1 of the image patch to obtain the spectrally downsampled image. Band 1 is the first band with the longest imaging band in the image. Perform spatial downsampling on the image patch to obtain the spatially downsampled image. Step 2-2: Extract spatial and spectral features for each image patch: extract spatial features from the spectral downsampled image, and extract spectral features from the spatial downsampled image; Steps 2-3: For each image patch, guided by the spatial features of band 1, the grayscale images of other bands are predicted. Then, the spectral features of the spatial downsampled image are used as constraints to ensure that the spectral features of the predicted multispectral image are consistent with the original image patch. The predicted image is subtracted from the original image patch to obtain the residual. During decoding, the predicted image is added to the residual to obtain the reconstructed multispectral image. Steps 2-4 involve encoding and decoding the residuals, as well as encoding and decoding the latent variables of spatial and spectral features. Step 3: Set the loss function and model parameters for model training, and train the compressed model built in Step 2. After training, a multispectral remote sensing image compression model based on grayscale guidance is obtained. Step 4: Input the multispectral remote sensing image to be compressed into the multispectral remote sensing image compression model for compression.

2. The multispectral remote sensing image compression method based on grayscale guidance according to claim 1, Its features are: In step 2-2, a spatial feature extractor and a spectral feature extractor are constructed using a shallow convolutional neural network. The spatial feature extractor consists of a 3×3×1 convolutional kernel and is used to extract spatial features from the spectral downsampled image. The spectral feature extractor consists of a 1×1×3 one-dimensional convolution kernel and is used to extract spectral features from spatially downsampled images.

3. The multispectral remote sensing image compression method based on grayscale guidance according to claim 1, characterized in that: in, In step 2-2, a spatial feature extractor is used to extract the spectral downsampled image. Spatial features obtained from U Using a spectral feature extractor to spatially downsample images Spectral characteristics obtained V ; U for tensor, U for tensor, h , w and c It is the original image patch Height, width, and number of channels.

4. The multispectral remote sensing image compression method based on grayscale guidance according to claim 3, characterized in that: in, In steps 2-3, spatial features and spectral features are combined to obtain spatial-spectral joint features. F : ； right F Spatial upsampling Then in terms of spectral dimensions ( c +1) Divide the joint features of the upsampled data to obtain: ； In the formula, The first characteristic representing the joint feature i One channel; H This represents the features after upsampling; calculate and of F The inner product under the norm serves as the spatial similarity between two channels. : ； Composed of the joint features of the first channel and the second channel to the third channel c Spatial similarity vector of +1 channel , ; Guided by the grayscale image of band 1, spatial similarity vectors between bands are added. S As an attention mechanism, spatial-spectral joint features are fused: ； In the formula, The output fusion features, The joint features adopted above, This represents the convolution operation. This indicates that the Hadamard product operation will ultimately fuse features. Input Image Generation Network G Prediction results for generating multispectral images at original resolution ; Prediction results With the original image patch Differences are calculated to obtain residuals. .

5. The grayscale-guided multispectral remote sensing image compression method according to claim 1, Its features are: In steps 2-4, the encoding and decoding of the residuals adopts a wavelet transform-based encoding and decoding method. Spatial features and spectral features are encoded and decoded separately using entropy estimation based on a Gaussian mixture model.

6. The multispectral remote sensing image compression method based on grayscale guidance according to claim 1, characterized in that: in, In step 4, the multispectral remote sensing image to be compressed is the image data that has been cropped in the same way as in step 1.

7. The multispectral remote sensing image compression method based on grayscale guidance according to claim 1, characterized in that: in, After constructing the multispectral remote sensing image compression model based on steps 1-3, all multispectral remote sensing images from the same source can be directly compressed by inputting into the multispectral remote sensing image compression model in step 4, thus achieving encoding and decoding.

8. A multispectral remote sensing image compression system based on grayscale guidance, characterized in that, include: The data preparation department acquires historical multispectral remote sensing images from the same source based on the multispectral remote sensing image to be compressed, and crops the historical multispectral remote sensing images for model construction. Each historical multispectral remote sensing image is cropped into multiple image patches. The training model building department is responsible for building compressed models for training; it includes the following modules: The downsampling module performs spectral and spatial downsampling on each image patch: it extracts the grayscale image of band 1 of the image patch to obtain a spectrally downsampled image, where band 1 is the longest imaging band in the image; and it performs spatial downsampling on the image patch to obtain a spatially downsampled image. The feature extraction module extracts spatial and spectral features for each image patch: spatial features are extracted from the spectrally downsampled image, and spectral features are extracted from the spatially downsampled image. For each image patch, the other band prediction module uses the spatial features of band 1 as a guide to predict the grayscale images of other bands. Then, it uses the spectral features of the spatial downsampled image as a constraint to make the spectral features of the predicted multispectral image consistent with the original image patch. The predicted image is subtracted from the original image patch to obtain the residual. During decoding, the predicted image is added to the residual to obtain the reconstructed multispectral image. The encoding / decoding module encodes and decodes the residuals, as well as the latent variables of spatial and spectral features. In the compression model construction section, the loss function and model parameters for model training are set, and the compression model built in step 2 is trained. After training, a multispectral remote sensing image compression model based on grayscale guidance is obtained. The compression unit inputs the multispectral remote sensing image to be compressed into the multispectral remote sensing image compression model for compression. The control unit is connected to the data preparation unit, training model building unit, compression model building unit, and compression unit, and controls their operation.

9. The multispectral remote sensing image compression system based on grayscale guidance according to claim 8, Its features are: The compression section includes: The testing module is used to perform grayscale-guided multispectral remote sensing image testing and verification on the model trained by the compression model construction unit. The model is input with the multispectral image test data to be compressed. Guided by the grayscale information of band 1, it outputs encoded files and decoded images, and tests the peak signal-to-noise ratio (PSNR) and structural similarity of the decoded images. If the test results do not meet the requirements, the module returns to the compression model construction unit for adjustments: based on the PSNR and structural similarity, for cases where the decoded image quality does not meet the requirements, the number of training iterations is increased, and the weight value of the prediction error corresponding to the original image patch in the loss function is increased, and training continues; based on the comparison of file sizes before and after compression, for cases where the compression ratio does not meet the requirements, the weight value of the latent code entropy in the loss function is increased, and training continues until the test results meet the requirements; thus, the final multispectral remote sensing image compression model is obtained. The compression module inputs the multispectral remote sensing image to be compressed into the final multispectral remote sensing image compression model for compression.

10. The multispectral remote sensing image compression system based on grayscale guidance according to claim 8, characterized in that, Also includes: The input display unit is connected in communication with the control unit and is used to allow users to input operation commands and display the corresponding commands.

Images