Marine environment data compression method and system based on deep learning

Abstract

The invention belongs to the technical field of ocean observation data compression, and discloses an ocean environment data compression method and system based on deep learning. The method comprises the following steps: converting processed marine environment data into a tensor format to form a model training data set; constructing a deep learning model adopting an encoder-decoder structure, training the constructed deep learning model by using the model training data set, compressing the input model training data set through an encoder, and then restoring the compressed model training data set into data the same as input data through a decoder; and comparing the restored data with the original marine environment data which is not trained by the deep learning model by using an evaluation index to verify the restoration precision. According to the method, attention weights for different regions or variables can be dynamically distributed, so that key features are extracted more accurately. The layered hidden space structure designed by the invention can separate and represent multi-scale feature information, and ensures the effectiveness and robustness in a complex data environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean observation data compression, and particularly relates to a method and system for compressing ocean environment data based on deep learning. Background Art

[0002] Ocean environment data has the characteristics of multi-dimension, multi-scale and multi-type, including sea surface temperature, ocean three-dimensional temperature field, salinity field, sea current field, wave data, and seabed topography and geomorphology, etc. These data are interrelated and highly coupled, and the scale is often extremely large, such as high-resolution ocean remote sensing images, underwater detection data, and seabed acoustic detection data, etc. The overall data volume increases at the level of TB or even PB. For ships operating at sea, the real-time or quasi-real-time acquisition of ocean environment data is of great significance for route planning, risk prediction, and ensuring operation safety. However, the satellite communication links relied on by ships have limitations in terms of bandwidth and stability, making it difficult to meet the long-term and high-speed transmission requirements of massive data. Therefore, it is of great engineering value and social significance to achieve efficient compression and transmission of massive ocean environment data in a limited bandwidth environment.

[0003] Traditional data compression methods (such as lossy or lossless compression techniques based on wavelet, wavelet packet, discrete cosine transform, etc.) have achieved good results in image processing, video compression, and partial spatio-temporal data compression. For example, multi-resolution analysis based on Fourier transform or wavelet transform can effectively compress the spatio-temporal distribution data of ocean meteorological fields and ocean wave fields. However, when facing three-dimensional or multi-source heterogeneous ocean data (such as high-resolution underwater terrain data that integrates multiple sensor information), traditional methods often have difficulty in achieving both a high compression ratio and good fidelity. In addition, with the continuous improvement of ocean observation accuracy and frequency, the internal structure of massive data has become increasingly complex, and traditional feature extraction and coding methods may not be able to fully capture the important detailed features in the data, resulting in the data quality after decompression being difficult to meet the needs of ocean monitoring, forecasting, and decision-making.

[0004] A patent for invention, an adaptive compression method for ocean observation data (publication number CN117997351A, published on May 7, 2024), discloses that ocean observation data is rearranged into an approximate square matrix, divided into different block patterns through an automatic block division method, and an improved singular value decomposition technique is used to compress each square matrix.

[0005] Through the above analysis, the problems and defects existing in the prior art are as follows:

[0006] (1)Insufficient adaptability to multi-source heterogeneous data. Traditional compression methods (such as wavelet transform, DCT) and the singular value decomposition (SVD) block technology in Patent CN117997351A are mainly designed for structured or regularized data. However, ocean data has characteristics such as multi-dimensions (such as three-dimensional temperature field) and multi-types (such as the fusion of remote sensing images and acoustic detection data). Its unstructured features (such as seabed terrain point clouds) are difficult to be effectively characterized by fixed blocks or single transformation basis functions, resulting in the loss of detailed features or the decrease of fidelity after compression.

[0007] (2)Limited feature extraction ability. Traditional methods rely on manually designed transformation bases (such as Fourier basis) or fixed block patterns, and it is difficult to adaptively capture non-linear features (such as ocean current field vortices, sudden changes in seabed landforms) in complex ocean data. Especially, the cross-variable coupling relationships (such as the co-variation of temperature-salinity) in multi-sensor fusion data are difficult to be accurately modeled by linear decomposition techniques (such as SVD), which affects the scientific analysis value of the compressed data.

[0008] (3)Computational efficiency and real-time bottlenecks. The method based on singular value decomposition needs to perform matrix decomposition on each block, and the computational complexity increases cubically with the data scale (O(n3)), which is difficult to meet the real-time compression requirements of TB / PB-level data. At the same time, traditional methods lack optimized support for hardware acceleration (such as GPU parallelization) and are difficult to run efficiently on ship edge computing devices. Summary of the Invention

[0009] To overcome the problems existing in the related technologies, the disclosed embodiments of the present invention provide a deep learning-based ocean environment data compression method and system. The purpose of the present invention is to propose an autoencoder deep learning model integrating an attention mechanism to better adapt to the complex characteristics of high-dimensional ocean environment data. Based on the traditional autoencoder framework, this model integrates a multi-head attention mechanism to capture the long-range dependence relationships among different spatial dimensions, time dimensions, and cross-variables in the data. By introducing the attention mechanism, the model can dynamically allocate attention weights to different regions or variables, thereby more accurately extracting key features. In addition, to further improve the compression efficiency and reconstruction accuracy of the model, the present invention designs a hierarchical latent space structure to separate and characterize multi-scale feature information, ensuring effectiveness and robustness in complex data environments.

[0010] The technical solution is as follows: The deep learning-based ocean environment data compression method includes the following steps:

[0011] S1, obtain ocean environment data, arrange the ocean environment data in dimensions of time, depth, longitude, and latitude, and set the values of land areas to 0 if there are land areas; convert the processed ocean environment data into a tensor format to form a model training dataset;

[0012] S2. Construct a deep learning model with an encoder-decoder structure. The encoder gradually compresses the input multi-channel data field into a low-dimensional latent space through multiple layers of convolution and pooling. In the decoder part, transposed convolution is used to gradually decompress the latent vector or feature map back to the same resolution as the input, and the reconstructed ocean environment data is output.

[0013] S3. Use the model training dataset to train the constructed deep learning model. After the input model training dataset is compressed by the encoder, it is restored to the same data as the input through the decoder. The restored data is compared with the original ocean environment data that has not been trained by the deep learning model using evaluation metrics to verify the restoration accuracy. The evaluation metrics include: mean relative error, mean absolute error.

[0014] In step S2, the encoder gradually compresses the input multi-channel data field into a low-dimensional latent space through multiple layers of convolution and pooling, including: the encoder consists of n convolutional layers with a stride of 2 and a number of channels of m i , where n is the number of convolutional layers; and an attention layer is added before the last convolutional layer; the ReLU function is used for activation after each convolution; the preprocessed data enters the encoder, and after each convolution in the encoder, the data volume becomes 1 / 2 of the original, and finally the compressed data is formed; the dimensions of the preprocessed data are:

[0015] B×D×H×W

[0016] In the formula, B is the batch size, D is the number of depth layers, H is the meridional length, and W is the zonal length.

[0017] Furthermore, the expression of convolution is:

[0018]

[0019] In the formula, out(B i , K j ) is the convolution result of the i-th sample under the j-th convolution kernel, bias(K j ) is the bias term of the j-th convolution kernel, h, w are the indices of H, W, weight(K j , h, w) is the weight value of the j-th convolution kernel at h, w, input(B i , h, w) is the data value of the i-th sample at h, w;

[0020] The expression of pooling is:

[0021]

[0022] In the formula, yo,p is the value at position (o, p) in the output feature map after pooling. max is the maximum value in the selection window as the output value, and x o+m,p+n is the value at position (o + m, p + n) in the input feature map after pooling. pool is the range of the pooling window, m and n are the index ranges of the pooling window, x is the input feature map, and o and p are the positions of the pooling window in the input feature map.

[0023] Furthermore, the expression of the ReLU function is:

[0024] ReLU(x) = max(0, x)

[0025] In the formula, x is the input feature map, and max is the maximum value in the selection window as the output value.

[0026] Furthermore, the number of layers n of the convolutional layer is:

[0027] n = N / 2

[0028] In the formula, N is the compression ratio required;

[0029] The number of channels m of the i-th convolutional layer i is:

[0030] m i = D × 2 × i, (i < n)

[0031] In the formula, D is the number of depth layers.

[0032] Furthermore, the attention mechanism of the attention layer is:

[0033]

[0034] In the formula, a i is the i-th input vector, q i is the i-th query vector, k i is the i-th key vector, v i is the i-th value vector, Attention(q, k, v) is the attention value, softmax() is the probability distribution function, T is the matrix transpose, and W q , W k , W v are the trainable weight matrices corresponding to q, k, and v. q is the query vector, k is the key vector, and v is the value vector. is the square root of the dimension of k.

[0035] In step S2, the expression of the transposed convolution is:

[0036] I = C T * O

[0037] Wherein, I is the output vector, O is the input vector, C is the convolution kernel matrix, and T is the matrix transpose.

[0038] In step S3, training the constructed deep learning model using the model training dataset includes:

[0039] Use the PyTorch framework to train the deep learning model. The deep learning model is trained for 100 rounds in total, and the batch size is 32. In each round of training, update the parameters of the deep learning model according to the gradient of the mean absolute error loss function, which measures the difference between the predicted value and the true value. The learning rate scheduler dynamically adjusts the learning rate in each round.

[0040] Furthermore, the learning rate follows the cosine annealing algorithm throughout the training process. The cosine annealing is as follows:

[0041]

[0042] Where: lr(α) is the learning rate of the α-th training round, X is the total number of training rounds, lrf is the learning rate decay factor, and lr0 is the initial learning rate.

[0043] Another object of the present invention is to provide a deep learning-based marine environment data compression system, which implements the deep learning-based marine environment data compression method. The system includes:

[0044] A model training dataset acquisition module for acquiring marine environment data, arranging the marine environment data in dimensions of time, depth, longitude, and latitude, and setting the values of land areas to 0 if there are land areas; converting the processed marine environment data into a tensor format to form a model training dataset;

[0045] A deep learning model construction module for constructing a deep learning model using an encoder-decoder structure. The encoder gradually compresses the input multi-channel data field into a low-dimensional latent space through multiple layers of convolution and pooling. In the decoder part, use transposed convolution to gradually decompress the latent vector or feature map back to the same resolution as the input, and output the reconstructed marine environment data;

[0046] A verification reduction accuracy module for using the deep learning model constructed by training the model training dataset. After compressing the input model training dataset through the encoder, restore it to the same data as the input through the decoder; compare the restored data with the original marine environment data without deep learning model training using evaluation indicators to verify the reduction accuracy.

[0047] Combining all the above technical solutions, the beneficial effects of the present invention are as follows: The present invention proposes a deep learning model of an autoencoder integrating an attention mechanism. On the basis of the traditional autoencoder framework, a multi-head attention mechanism is integrated to capture long-range dependencies among different spatial dimensions, temporal dimensions, and cross-variables in the data. The model can dynamically allocate attention weights to different regions or variables, thereby more accurately extracting key features, achieving a balance between high compression ratio and key scientific feature fidelity under limited satellite bandwidth, and significantly improving the compression efficiency and reconstruction accuracy of high-dimensional ocean environment data.

[0048] Compared with traditional fixed basis function methods such as wavelet transform and DCT, and block SVD technology, the deep learning model integrating a multi-head attention mechanism in the present invention can perform high-rate compression on high-dimensional ocean environment data, avoid the destruction of the topological structure caused by data rearrangement, and greatly improve the compression fidelity of unregularized data. The model can dynamically allocate attention weights to different regions or variables, thereby more accurately extracting key features. In addition, in order to further improve the compression efficiency and reconstruction accuracy of the model, the hierarchical latent space structure designed in the present invention can separate and characterize multi-scale feature information, ensuring effectiveness and robustness in complex data environments.

[0049] Through high-fidelity compression and real-time transmission capabilities, the satellite communication costs of ocean-going ships and ocean monitoring platforms can be significantly reduced, while improving the data timeliness in scenarios such as ocean disaster early warning and resource exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings herein are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure;

[0051] Figure 1 is a flowchart of a method for compressing ocean environment data based on deep learning provided by an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of the principle of a method for compressing ocean environment data based on deep learning provided by an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of the principle of a deep learning model provided by an embodiment of the present invention;

[0054] Figure 4 is a MAE graph of ocean temperature restoration at different depths provided by an embodiment of the present invention;

[0055] Figure 5 is a MAPE graph of ocean temperature restoration at different depths provided by an embodiment of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0057] The innovation of the present invention lies in: innovatively integrating the multi-head attention mechanism in the traditional autoencoder framework to dynamically allocate the coding weights of different regions or variables and accurately extract key features; at the same time, constructing a hierarchical latent space structure to explicitly separate multi-level features in ocean data through multiple sub-networks, achieving high-fidelity compression of ocean environmental data.

[0058] Example 1, the present invention proposes a deep learning model of an autoencoder enhanced by an attention mechanism to achieve efficient compression and feature extraction of high-dimensional ocean environmental data. This model innovatively integrates the multi-head cross-dimensional attention mechanism in the traditional autoencoder framework, breaking through the limitation of the local receptive field of traditional convolutional operations. At the same time, the model adopts a hierarchical latent space decoupling architecture, decomposing the encoding process into multiple sub-networks to extract different features of environmental data respectively.

[0059] Specifically, as Figure 1 shown, the method for compressing ocean environmental data based on deep learning provided by the embodiments of the present invention includes:

[0060] S1, obtaining ocean environmental data, arranging the ocean environmental data in dimensions of time, depth, longitude, and latitude, and setting the values of land areas to 0 if there are land areas; converting the processed ocean environmental data into a tensor format to form a model training dataset;

[0061] S2, constructing a deep learning model with an encoder-decoder structure. The encoder gradually compresses the input multi-channel data field into a low-dimensional latent space through multiple layers of convolution and pooling; in the decoder part, using transposed convolution, the latent vector or feature map is gradually decompressed back to the same resolution as the input, and the reconstructed ocean environmental data is output;

[0062] wherein, the multi-channel data field includes temperature, salinity, sea surface height, etc.;

[0063] S3. Use the model training dataset to train the constructed deep learning model. After compressing the input model training dataset through the encoder, restore it to the same data as the input through the decoder. Compare the restored data with the original ocean environment data that has not been trained by the deep learning model using evaluation metrics to verify the restoration accuracy. The evaluation metrics include: mean relative error and mean absolute error.

[0064] Example 2. As another implementation manner of the present invention, as Figure 2 shown, the deep learning-based ocean environment data compression method provided by the embodiment of the present invention includes:

[0065] (1) Data acquisition.

[0066] Download the open-source ocean environment dataset and preprocess the data.

[0067] (2) Data encoding.

[0068] Construct an encoder for compressing the original data. The encoder consists of n convolutional layers with a stride of 2 and a number of channels of m i . And add an attention layer before the last convolutional layer. Use the ReLU function for activation after each convolution. The preprocessed data (with dimensions B×D×H×W, where B is the batch size, D is the number of depth layers, H is the meridional length, and W is the zonal length) will enter the encoder. After each convolution in the encoder, the data volume will become 1 / 2 of the original, and finally form the compressed data.

[0069] (3) Data decoding.

[0070] Construct a decoder for restoring the compressed data. The decoder consists of n transposed convolutional layers with a stride of 2 and a number of channels of m i . And add an attention layer after the first transposed convolutional layer. Use the ReLU function for activation after each transposed convolution. The compressed data will enter the decoder. After each convolution in the decoder, the data volume will become 2 times the original, and finally form the restored data.

[0071] Among them, a deep learning model is constructed by building an encoder-decoder structure.

[0072] (4) Deep learning model training.

[0073] The deep learning model is trained using the PyTorch framework. The deep learning model is trained for a total of 100 rounds, with a batch size of 32. In each round of training, the model parameters are updated according to the gradient of the mean absolute error loss function, which measures the difference between the predicted value and the true value. The learning rate scheduler dynamically adjusts the learning rate in each round to ensure efficient learning throughout the training process. To monitor the training progress, the model is evaluated on the test set after each round. If the test loss improves, the current model weights are saved as the best-performing model to prevent overfitting and ensure optimal performance.

[0074] (5) Data restoration accuracy evaluation.

[0075] The accuracy of the restoration result is verified using data that has not been trained by the deep learning model.

[0076] (6) Data compression and restoration.

[0077] The trained deep learning model is split into an encoder and a decoder. The encoder is used onshore or in the cloud to compress the original data. Subsequently, the compressed data is sent to the application side such as a marine vessel. The decoder is used on the application side to restore the compressed data.

[0078] Exemplarily, the marine environmental data downloaded in step (1) is in netCDF format.

[0079] Exemplarily, in step (1), the data preprocessing is to arrange the data dimensions as [time, depth, longitude, latitude]. If there is a land area, the values in the land area are set to 0. Finally, the processed data is converted into a tensor format.

[0080] Exemplarily, step (2) is specifically as follows:

[0081] (2.1) The specific expression of convolution is:

[0082]

[0083] In the formula, out(B i ,K j ) is the convolution result of the i-th sample under the j-th convolution kernel, bias(K j ) is the bias term of the j-th convolution kernel, h and w are the indices of H and W, weight(K j ,h,w) is the weight value of the j-th convolution kernel at h and w, and input(B i ,h,w) is the data value of the i-th sample at h and w;

[0084] (2.2) The specific expression of pooling is:

[0085]

[0086] In the formula, y o,p is the value at position (o, p) in the output feature map after pooling. max selects the maximum value in the selection window as the output value, and x o+m,p+n is the value at position (o + m, p + n) in the input feature map after pooling. pool is the range of the pooling window, and m, n are the index ranges of the pooling window. x is the input feature map, and o, p represent the position of the pooling window in the input feature map.

[0087] (2.3) The specific expression of the ReLU activation function is:

[0088] ReLU(x) = max(0, x)

[0089] In the formula, x is the input feature map, and max selects the maximum value in the selection window as the output value.

[0090] (2.4) Innovatively proposed in the present invention, the number of convolutional layers n is:

[0091] n = N / 2

[0092] In the formula, N is the compression ratio required;

[0093] (2.5) Innovatively proposed in the present invention, the number of channels m of the i-th convolutional layer i is:

[0094] m i = D × 2 × i, (i < n)

[0095] In the formula, D is the depth layer number.

[0096] (2.6) Since the attention weights in different regions are different, the present invention adopts an attention mechanism to extract this mapping relationship. The attention mechanism mimics human selective attention, gradually selects more critical information during the training process, assigns different weights according to the importance of different information, so as to better learn data features. The specific expression of the attention mechanism is:

[0097]

[0098] In the formula, a i is the i-th input vector, q i is the i-th query vector, k i is the i-th key vector, v i is the i-th value vector, Attention(q, k, v) is the attention value, softmax() is the probability distribution function, T is the matrix transpose, W q , W k , W vare trainable weight matrices corresponding to q, k, and v, where q is the query vector, k is the key vector, and v is the value vector. is the square root of the dimension of k.

[0099] Exemplarily, in step (3), the specific expression of the transposed convolution is:

[0100] I = C T * O

[0101] In the formula, I is the output vector, O is the input vector, C is the convolution kernel matrix, and T is the matrix transpose.

[0102] Exemplarily, in step (4), to improve convergence, the learning rate follows the cosine annealing algorithm throughout the training process. The specific expression of cosine annealing is:

[0103]

[0104] In the formula: lr(α) is the learning rate at the α-th training epoch, X is the total number of training epochs, lrf is the learning rate decay factor, and lr0 is the initial learning rate.

[0105] Exemplarily, in step (5), the selected accuracy verification metrics are the mean absolute error MAE and the mean relative error MAPE:

[0106]

[0107] In the formula, y true is the true value, y predict is the predicted value, and NN is the total amount of data.

[0108] Example 3, The marine environment data compression system based on deep learning provided by the embodiments of the present invention includes:

[0109] A model training dataset acquisition module, which is used to acquire marine environment data, arrange the dimensions of the marine environment data as [time, depth, longitude, latitude], and if there is a land area, set the value of the land area to 0; finally, convert the processed marine environment data into a tensor format to form a model training dataset.

[0110] A deep learning model construction module, which is used to construct a deep learning model adopting an encoder-decoder structure, including: the encoder gradually compresses the input multi-channel data field (such as temperature, salinity, sea surface height, etc.) into a low-dimensional latent space through multiple layers of convolution and pooling; then, in the decoder part, using transposed convolution, gradually "uncompress" the latent vector or feature map back to the same resolution as the input, and finally output the reconstructed marine environment data;

[0111] A verification reduction accuracy module is used to train and construct a deep learning model using a model training dataset, including: compressing the input model training dataset through an encoder and then restoring it to the same data as the input through a decoder; comparing the restored data with the original ocean environment data that has not been trained by the deep learning model using evaluation metrics to verify the reduction accuracy; the evaluation metrics include: mean relative error, mean absolute error.

[0112] To further illustrate the relevant effects of the embodiments of the present invention, the following experiments are conducted. The present invention calculates the reduction errors at different compression ratios, and different compression ratios can be selected according to actual accuracy requirements.

[0113] Table 1 Error comparison at different compression ratios

[0114]

[0115]

[0116] The present invention calculates the reduction effects of the deep learning model at different depths (as Figure 4 - Figure 5 shown). The deep learning model maintains a high data fidelity at different depths. In the shallow water area (0–200m), the RMSE reaches a peak near the 50m depth, approaching 0.55°C. Since the shallow sea temperature value is relatively high, the influence of the absolute error on the MAPE is relatively small. Therefore, the MAPE in the shallow area is relatively low and fluctuates smoothly, remaining within the range of 1% - 4%. The RMSE in the middle layer area (200–600m) gradually decreases and stabilizes near the 400m depth, and the error value remains between 0.35–0.40°C. In the deep area (600–1000m), the RMSE further decreases and reaches the lowest value in the depth range of 600–800m, approaching 0.25°C.

[0117] The above is only a relatively preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for compressing marine environmental data based on deep learning, characterized in that, The method includes the following steps: S1. Obtain marine environmental data, arrange the marine environmental data in dimensions of time, depth, longitude, and latitude, and set the values of land areas to 0 if there are land areas; convert the processed marine environmental data into a tensor format to form a model training dataset; S2. Construct a deep learning model using an encoder-decoder structure. The encoder gradually compresses the input multi-channel data field into a low-dimensional latent space through multiple layers of convolution and pooling; in the decoder part, use transposed convolution to gradually decompress the latent vector or feature map back to the same resolution as the input and output the reconstructed marine environmental data; S3. Use the model training dataset to train the constructed deep learning model. After the input model training dataset is compressed by the encoder, it is then restored to the same data as the input through the decoder; compare the restored data with the original marine environmental data that has not been trained by the deep learning model using evaluation metrics to verify the restoration accuracy; the evaluation metrics include: mean relative error, mean absolute error.

2. The method for compressing marine environment data based on deep learning according to claim 1, wherein In step S2, the encoder gradually compresses the input multi-channel data field into a low-dimensional latent space through multiple layers of convolution and pooling, including: the encoder consists of n convolutional layers with a stride of 2 and a number of channels of m i where n is the number of convolutional layers; and an attention layer is added before the last convolutional layer, and the ReLU function is used for activation after each convolution; the preprocessed data enters the encoder, and after each convolution in the encoder, the data volume becomes 1 / 2 of the original, and finally the compressed data is formed; the dimension of the preprocessed data is: B×D×H×W In the formula, B is the batch size, D is the number of depth layers, H is the meridional length, and W is the zonal length.

3. The method for compressing marine environmental data based on deep learning according to claim 2, wherein The expression of convolution is: where out(B i , K j ) is the convolution result of the i-th sample under the j-th convolution kernel, bias(K j ) is the bias term of the j-th convolution kernel, h and w are the indices of H and W, weight(K j , h, w) is the weight value of the j-th convolution kernel at h and w, and input(B i , h, w) is the data value of the i-th sample at h and w; The expression of pooling is: where y o,p is the value at position (o, p) in the output feature map after pooling, max selects the maximum value in the selection window as the output value, and x o+m,p+n is the value at position (o + m, p + n) in the input feature map after pooling, pool is the range of the pooling window, m and n are the index ranges of the pooling window, x is the input feature map, and o and p are the positions of the pooling window in the input feature map.

4. The method for compressing marine environmental data based on deep learning according to claim 2, wherein, The expression of the ReLU function is: ReLU(x) = max(0, x) In the formula, x is the input feature map, and max selects the maximum value in the window as the output value.

5. The method for compressing marine environment data based on deep learning according to claim 2, wherein The number of layers n of the convolutional layer is: n = N / 2 In the formula, N is the compression ratio required; The number of channels m of the i-th convolutional layer i is as follows: m i = D × 2 × i, (i < n) In the formula, D is the number of depth layers.

6. The method for compressing marine environment data based on deep learning according to claim 2, wherein The attention mechanism of the attention layer is: Where a i is the i-th input vector, q i is the i-th query vector, k i is the i-th key vector, v i is the i-th value vector, Attention(q, k, v) is the attention value, softmax() is the probability distribution function, T is the matrix transpose, W q , W k , W v are the trainable weight matrices corresponding to q, k, v, q is the query vector, k is the key vector, v is the value vector, is the square root of the dimension of k.

7. The method for compressing marine environmental data based on deep learning according to claim 1, wherein In step S2, the expression of transposed convolution is: I = C T *O In the formula, I is the output vector, O is the input vector, C is the convolution kernel matrix, and T is the matrix transpose.

8. The method for compressing marine environment data based on deep learning according to claim 1, characterized in that In step S3, using the model training dataset to train the constructed deep learning model includes: Use the PyTorch framework to train the deep learning model. The deep learning model is trained for 100 rounds in total, and the batch size is 32; in each round of training, update the parameters of the deep learning model according to the gradient of the mean absolute error loss function, which measures the difference between the predicted value and the true value, and the learning rate scheduler dynamically adjusts the learning rate in each round.

9. The deep learning-based marine environment data compression method according to claim 8, characterized in that, The learning rate follows the cosine annealing algorithm throughout the training process, and the cosine annealing is: In the formula: lr(α) is the learning rate of the α-th training round, X is the total number of training rounds, lrf is the learning rate decay factor, and lr0 is the initial learning rate.

10. A marine environment data compression system based on deep learning, characterized in that, The system implements the deep learning-based marine environmental data compression method according to any one of claims 1-9. The system includes: A model training dataset acquisition module for obtaining marine environmental data, arranging the marine environmental data in dimensions of time, depth, longitude, and latitude, and setting the values of land areas to 0 if there are land areas; converting the processed marine environmental data into a tensor format to form a model training dataset; A deep learning model construction module, which is used to construct a deep learning model adopting an encoder-decoder structure. The encoder gradually compresses the input multi-channel data field into a low-dimensional latent space through multiple layers of convolution and pooling. In the decoder part, transposed convolution is used to gradually decompress the latent vector or feature map back to the same resolution as the input, and output the reconstructed ocean environment data; A verification of restoration accuracy module, which is used to train the constructed deep learning model using the model training data set. After the input model training data set is compressed by the encoder, it is restored to the same data as the input through the decoder; the restored data is compared with the original ocean environment data that has not been trained by the deep learning model using evaluation metrics to verify the restoration accuracy.

Images