A method and device for spatial downscaling of marine data based on a double backbone network

Abstract

The application discloses a kind of ocean data space scale-down method and device based on double backbone network, belong to ocean data processing technical field.The method includes: obtaining low-resolution ocean multi-element data containing ocean current elements and auxiliary ocean elements, and carrying out preset quantile maximum minimum value normalization processing;The normalized data is input into the double backbone network model trained in advance, and the model includes main feature backbone network, auxiliary feature backbone network, middle layer feature fusion module, deep reconstruction module, upsampling module and output convolution layer;Two backbone networks with different parameters are used to extract independent features of ocean current elements and auxiliary ocean elements respectively, and high-resolution ocean current scale-down reconstruction results are output after feature fusion, deep reconstruction, upsampling and output convolution processing.The application uses a double backbone differentiation feature extraction mechanism to effectively avoid feature interference between different physical property elements and improve the accuracy of sea surface flow field scale-down reconstruction.

Description

Technical Field

[0001] This invention relates to the field of marine data processing technology, specifically to a method and apparatus for spatial downscaling of marine data based on a dual-backbone network. Background Technology

[0002] Surface currents are an important indicator of ocean dynamic processes and have significant application value in operational ocean forecasting fields such as marine disaster early warning, oil spill drift prediction, and maritime search and rescue. Current operational ocean models typically have a resolution of approximately 1 / 12° (around 8–10 km depending on latitude), which can effectively describe large-scale ocean circulation, but their ability to characterize complex dynamic processes such as sub-mesoscale eddies (1–10 km), fine frontal structures, and small-scale disturbances within 1 km is limited.

[0003] Traditional methods for improving resolution primarily rely on dynamic downscaling, which involves recalculating the entire layer using a higher-resolution ocean numerical model. This approach requires simultaneously solving ocean dynamic equations, thermohaline transport equations, and multiphysics coupling processes, and performing long-term continuous integration of three-dimensional elements such as temperature, salinity, and current fields. As the model's horizontal resolution increases, the number of grid cells, time steps, and the amount of three-dimensional computation all increase significantly, leading to an exponential increase in computational, storage, and data transmission costs. Therefore, this approach is difficult to implement in operational settings and can only be performed using supercomputing, greatly limiting its application scenarios.

[0004] To reduce the cost of generating high-resolution marine products, early research mainly employed statistical downscaling methods, including interpolation reconstruction, empirical statistical relationships, and prior knowledge constraints. However, these methods have limited ability to recover complex marine dynamic structures. In recent years, with the development of deep learning technology, image super-resolution techniques from the field of computer vision have gradually been transferred to marine downscaling tasks. Existing deep learning-based downscaling methods are dominated by single-backbone deep residual networks, represented by Enhanced Deep Super-Resolution (EDSR). These networks directly analogize various marine elements (such as ocean current components, sea surface height, and sea surface temperature) to the red, green, and blue channels of natural images. After channel stitching, a single backbone network performs unified feature extraction and end-to-end optimization reconstruction using a globally weighted loss function.

[0005] However, ocean multi-element fields differ fundamentally from natural images in terms of physical properties, spatial structure, and dynamic mechanisms. In ocean current downscaling tasks, different ocean elements possess different physical meanings and spatial characteristics, and their roles in downscaling reconstruction also vary. Existing methods employ a uniform feature extraction approach for all input elements, failing to consider the differences in physical properties and the varying levels of their roles in the downscaling task, resulting in a lack of targeted feature extraction. Furthermore, the dynamic complexity of ocean current elements exhibits significant spatial differences across different regions. For instance, flow field changes are far more dramatic in land-sea boundary regions, mesoscale eddy activity zones, and frontal regions than in open ocean basins. Existing methods typically apply the same weight to all spatial locations during optimization, making it difficult to specifically improve the reconstruction quality of locally complex regions. These issues, to some extent, limit further improvements in the accuracy of ocean current downscaling reconstruction. Summary of the Invention

[0006] The purpose of this invention is to provide a method, apparatus, electronic device and storage medium for spatial downscaling of marine data based on a dual-backbone network, so as to improve the accuracy of ocean current downscaling reconstruction.

[0007] In a first aspect, embodiments of the present invention provide a method for spatial downscaling of marine data based on a dual-backbone network, comprising: Acquire low-resolution multi-element ocean data to be processed, wherein the low-resolution multi-element ocean data includes ocean current elements as the target of downscaling reconstruction and auxiliary ocean elements as auxiliary constraints. The low-resolution ocean multi-element data is subjected to preset quantile maximum and minimum value normalization processing to obtain normalized low-resolution ocean multi-element data. The normalized low-resolution ocean multi-element data is input into a pre-trained dual-backbone network model, which includes: a main feature backbone network, an auxiliary feature backbone network, a mid-level feature fusion module, a deep reconstruction module, an upsampling module, and an output convolutional layer. The main feature backbone network is used to extract features from the ocean current elements to obtain the main feature tensor; at the same time, the auxiliary feature backbone network is used to extract features from the auxiliary ocean elements to obtain the auxiliary feature tensor; wherein the convolution parameters of the main feature backbone network and the auxiliary feature backbone network are not shared. The main feature tensor and the auxiliary feature tensor are fused by the mid-layer feature fusion module to obtain a fused feature tensor; The deep reconstruction module performs deep feature extraction on the fused feature tensor, and the upsampling module enhances the spatial resolution of the feature tensor after deep feature extraction. The output convolutional layer performs ocean current element channel mapping on the feature tensor after resolution enhancement, and performs inverse normalization on the channel mapping output to output high-resolution ocean current downscaling reconstruction results.

[0008] Furthermore, the pre-training process of the dual-backbone network model includes the following steps: S1. Construction of marine dataset: The original high-resolution marine data is sequentially gridded and quality controlled, and then low-resolution marine data from the same source as the high-resolution marine data is obtained; the low-resolution marine data and the high-resolution marine data are normalized by preset quantile maximum and minimum normalization to obtain training set, test set and validation set; wherein, the dataset contains ocean current elements as the reconstruction target, and auxiliary marine elements as constraints. S2. Construct and train a single backbone network: Input ocean current elements from the training set into a preset single backbone network for downscaling reconstruction, calculate the root mean square error between the reconstruction results and the corresponding high-resolution label data to obtain the global average spatial error distribution under the training set; perform preset quantile max-min normalization on the spatial error distribution to generate an error weight tensor. S3. Construct a dual-backbone network: The dual-backbone network includes an independent main feature backbone network, an auxiliary feature backbone network, a mid-layer feature fusion module, a deep reconstruction module, an upsampling module, and an output convolutional layer; the main feature backbone network is used to extract ocean current feature features, and the auxiliary feature backbone network is used to extract auxiliary ocean feature features. The parameters of the two backbone networks are not shared. S4. Weighted training of dual-backbone network: Input multi-element data containing ocean current elements and auxiliary ocean elements from the training set into the dual-backbone network, and use a weighted loss function that fuses the error weight tensor to iteratively optimize the dual-backbone network until the preset convergence condition is met, and obtain the trained dual-backbone network model.

[0009] Further, in step S2, the single backbone network includes an input convolutional layer, a deep feature extraction backbone composed of multiple stacked residual feature extraction modules, an upsampling module, and an output convolutional layer connected in sequence; the single backbone network is operated using the land-sea mask loss function, with the land area mask value set to 0 and the ocean area mask value set to 1.

[0010] Furthermore, in step S3, the structures of both the main feature backbone network and the auxiliary feature backbone network are as follows: The input convolutional layer uses a 3×3 convolutional kernel followed by a ReLU activation function to map the input data to a high-dimensional feature space; Sixteen residual feature extraction modules are connected in series. The internal structure of each residual feature extraction module is as follows: a first 3×3 convolutional layer, a ReLU activation function, and a second 3×3 convolutional layer. A residual connection is provided between the input of the module and the output of the second 3×3 convolutional layer.

[0011] Furthermore, in step S3, the deep reconstruction module is composed of 16 shared residual feature extraction modules and a 3×3 feature mapping convolutional layer connected in sequence, and a global residual connection is provided between the input of the deep reconstruction module and the output of the feature mapping convolutional layer.

[0012] Furthermore, in step S3, the upsampling module is composed of two levels of subpixel convolutional upsampling units cascaded in sequence; each level of subpixel convolutional upsampling unit includes a 3×3 convolutional layer, a PixelShuffle spatial rearrangement layer with a magnification of 2, and a ReLU activation function. A single level achieves a 2x improvement in spatial resolution, and two levels cascaded achieve a 4x improvement in spatial resolution.

[0013] Furthermore, the ocean current elements include the U and V components of sea surface current velocity; the auxiliary ocean elements include sea surface height data and / or sea surface temperature data.

[0014] Secondly, embodiments of the present invention provide a marine data spatial downscaling device based on a dual-backbone network, comprising: The acquisition module is used to acquire low-resolution ocean multi-element data to be processed, the data including ocean current elements and auxiliary ocean elements, and to perform preset quantile maximum and minimum normalization processing on the data. The processing module is used to input normalized low-resolution ocean multi-element data into a pre-trained dual-backbone network model; through the dual-backbone network model, the dual-backbone feature extraction, feature fusion, deep reconstruction, upsampling and output convolution processing are performed to output high-resolution ocean current downscaling reconstruction results.

[0015] Thirdly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in the first aspect.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having computer-readable instructions stored thereon, the computer-readable instructions being executable by a processor to implement the method as described in the first aspect.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: by setting a main feature backbone network and an auxiliary feature backbone network with non-shared parameters, the target ocean current elements and auxiliary ocean elements are extracted independently, which effectively avoids feature interference between different physical attribute elements and enhances the model's ability to learn the differentiated features of multiple ocean elements; at the same time, by using the spatial error weight tensor generated based on the pre-training error of the single backbone network to weight the loss function in the training process of the dual backbone network, the model can adaptively strengthen its attention to complex dynamic regions with high errors, thereby improving the overall accuracy of the downscaling reconstruction of the sea surface current field. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 The flowchart of a marine data spatial downscaling method based on a dual-backbone network provided by an embodiment of the present invention is shown. Figure 2 The diagram shown is a flowchart of a dual-backbone network model training method provided in an embodiment of the present invention. Figure 3 The diagram shows the structural schematics of the single-backbone network and the dual-backbone network constructed according to embodiments of the present invention. Figure 4 This invention provides a comparison chart of the visualization results of downscaling reconstruction of ocean current elements using multiple schemes, as shown in the embodiments of the present invention. Figure 5 A schematic diagram of a marine data spatial downscaling device based on a dual-backbone network provided in an embodiment of the present invention is shown. Figure 6 A schematic diagram of an electronic device provided by an embodiment of the present invention is shown. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] It should be noted that, unless otherwise stated, the technical or scientific terms used in this invention should have the ordinary meaning as understood by one of ordinary skill in the art.

[0021] Furthermore, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order. Additionally, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to those processes, methods, products, or apparatuses.

[0022] Please refer to Figure 1 , Figure 1 This is a flowchart of a marine data spatial downscaling method based on a dual-backbone network provided by an embodiment of the present invention, including the following steps: S101. Obtain low-resolution ocean multi-element data to be processed, wherein the low-resolution ocean multi-element data includes ocean current elements as the target of downscaling reconstruction and auxiliary ocean elements as auxiliary constraints. The ocean current elements include the U and V components of sea surface current velocity; the auxiliary ocean elements include sea surface height (SSH) data and / or sea surface temperature (SST) data.

[0023] S102. Perform preset quantile maximum and minimum value normalization processing on the low-resolution marine multi-element data to obtain normalized low-resolution marine multi-element data. The preset percentiles can be the 95th percentile, 97th percentile, 99th percentile, etc.

[0024] S103. Input the normalized low-resolution ocean multi-element data into a pre-trained dual-backbone network model. The dual-backbone network model includes: a main feature backbone network, an auxiliary feature backbone network, a mid-level feature fusion module, a deep reconstruction module, an upsampling module, and an output convolutional layer. S104. The ocean current elements are extracted using the main feature backbone network to obtain the main feature tensor; simultaneously, the auxiliary ocean elements are extracted using the auxiliary feature backbone network to obtain the auxiliary feature tensor; wherein, the convolution parameters of the main feature backbone network and the auxiliary feature backbone network are not shared. S105. The main feature tensor and the auxiliary feature tensor are fused by the mid-layer feature fusion module to obtain a fused feature tensor. S106. The deep reconstruction module performs deep feature extraction on the fused feature tensor, and the upsampling module improves the spatial resolution of the feature tensor after deep feature extraction. S107. The output convolutional layer performs ocean current element channel mapping on the feature tensor after resolution improvement, and performs inverse normalization processing on the channel mapping output result to output high-resolution ocean current downscaling reconstruction result.

[0025] To facilitate understanding, the training process of the dual-backbone network model will be described in detail below.

[0026] Figure 2 The diagram shows a flowchart of a dual-backbone network model training method provided in an embodiment of the present invention. Figure 3 The diagram shows the structural schematics of the single-backbone network and the dual-backbone network constructed in the embodiments of the present invention.

[0027] Please refer to Figure 2 , Figure 3 The training process of a dual-backbone network model includes the following steps: Step S201: Construction of the ocean dataset.

[0028] The original high-resolution ocean data is sequentially gridded and quality controlled, and then low-resolution ocean data from the same source as the high-resolution ocean data is obtained. The low-resolution ocean data and the high-resolution ocean data are normalized to a preset quantile maximum and minimum, and then divided into training set, test set and validation set. The dataset includes ocean current elements as the reconstruction target and auxiliary ocean elements as constraints.

[0029] It should be noted that low-resolution ocean data can be directly collected data or data obtained by downsampling high-resolution ocean data; this invention does not limit this.

[0030] Specifically, original high-resolution and low-resolution ocean datasets were obtained. The high-resolution dataset included sea surface velocity components (U and V components), while the low-resolution dataset included sea surface velocity components (U and V components), sea surface height, and sea surface temperature. First, the raw data underwent standard gridding, interpolating the high-resolution data to a uniform 1 / 24° regular latitude and longitude grid, and the low-resolution data to a 1 / 6° regular latitude and longitude grid. Quality control was then performed to remove outliers and missing values. Both the low-resolution and high-resolution data were then normalized using the 99th percentile maximum-minimum method, calculated as follows: ,in and These represent the maximum and minimum values ​​after 99th percentile truncation. The normalized data were divided into training, validation, and test sets in an 8:1:1 ratio. The U and V components of ocean currents were the primary target ocean features for downscaling reconstruction, while sea surface height and sea surface temperature were secondary ocean features.

[0031] Step S202: Construct and train a single backbone network to generate error weight tensors.

[0032] The ocean current elements in the training set are input into a preset single backbone network for downscaling reconstruction. The root mean square error of the reconstruction result is calculated with the corresponding high-resolution label data to obtain the global average spatial error distribution under the training set. The spatial error distribution is then subjected to preset quantile max-min normalization to generate an error weight tensor.

[0033] Specifically, a single backbone network is constructed, with the following structure: Figure 3 As shown, it includes, in sequence: an input convolutional layer Conv (3×3 convolutional kernel, 2 input channels, 64 output channels, followed by a ReLU activation function); a deep feature extraction backbone, consisting of 32 residual feature extraction modules (…). Figure 3 The residual blocks (referred to as residual blocks in Chinese) are stacked sequentially. The internal structure of each residual block is as follows: a first 3×3 convolutional layer - a ReLU activation function - a second 3×3 convolutional layer, with a residual connection between the module input and the output of the second convolutional layer; an upsampling module, which consists of two cascaded sub-pixel convolutional upsampling units, each of which includes a 3×3 convolutional layer, a PixelShuffle spatial rearrangement layer with a magnification of 2, and a ReLU activation function; and an output convolutional layer, which uses a 3×3 convolutional kernel and has 2 output channels.

[0034] Low-resolution data containing only ocean current U and V components from the training set are input into this single backbone network for pre-training. During training, a land-sea masking loss function is used, with land region mask values ​​set to 0 and ocean region mask values ​​set to 1. The mask is multiplied element-wise by the basic L1 loss function to obtain the land-sea masking loss function: L = L1 ⊙ MASK; where Mask∈{0,1}.

[0035] After pre-training, the low-resolution data of ocean current elements in the training set are input again into the single-backbone network to obtain downscaled reconstruction results. The root mean square error between the reconstruction result and the corresponding high-resolution ground truth value is calculated for each grid point to obtain the global spatial error distribution. The spatial error distribution is normalized to the 99th percentile maximum and minimum values ​​to generate the spatial error weight tensor W. The higher the value of W, the greater the reconstruction difficulty of that region under the single-backbone network.

[0036] Step S203: Construction and weighted training of the dual backbone network.

[0037] The dual-backbone network comprises an independent main feature backbone network, an auxiliary feature backbone network, a mid-layer feature fusion module, a deep reconstruction module, an upsampling module, and an output convolutional layer. The main feature backbone network is used to extract ocean current features, and the auxiliary feature backbone network is used to extract auxiliary ocean features. The parameters of the two backbone networks are not shared.

[0038] Specifically, a dual-backbone network is constructed, and its overall architecture is as follows: Figure 3 As shown, it specifically includes the following modules: (1) Main Feature Backbone Network: Used to receive the U and V components of ocean current input. The structure is as follows: input convolutional layer (3×3 convolutional kernel, 2 input channels, 64 output channels, followed by a ReLU activation function); 16 residual feature extraction modules (referred to as residual blocks) are sequentially connected. The internal structure of each residual block is "first 3×3 convolutional layer - ReLU activation function - second 3×3 convolutional layer," with residual connections between the input of the residual block and the output of the second convolutional layer. This network outputs the main feature tensor. .

[0039] (2) Auxiliary Feature Backbone Network: Used to receive sea surface height data input. Its structure is completely consistent with the main feature backbone network, including an input convolutional layer (1 input channel, 64 output channels) and 16 residual feature extraction modules, but all convolutional parameters are stored independently and not shared. This network outputs an auxiliary feature tensor. .

[0040] (3) Mid-level feature fusion module: This module fuses the main feature tensor. and auxiliary feature tensor The features are concatenated along the channel dimension to obtain 128 channels; then, a 1×1 convolutional layer (128 input channels, 64 output channels) is used for channel compression and feature recombination to obtain a fused feature tensor. .

[0041] (4) Deep Reconstruction Module: Consists of 16 shared residual feature extraction modules and one 3×3 feature mapping convolutional layer connected sequentially. Module Input There is a global residual connection between the output of the feature mapping convolutional layer and the output of the feature mapping convolutional layer, i.e., the output... .

[0042] (5) Upsampling module: It consists of two cascaded subpixel convolutional upsampling units. Each stage includes a 3×3 convolutional layer, a PixelShuffle spatial rearrangement layer with a magnification of 2, and a ReLU activation function. A single stage achieves a 2x improvement in spatial resolution, and two cascaded stages achieve a 4x improvement in spatial resolution.

[0043] (6) Output convolutional layer: A 3×3 convolutional kernel is used, and the number of output channels is 2 (corresponding to the U component and V component respectively) to obtain the final high-resolution ocean current downscaling reconstruction result.

[0044] The low-resolution multi-factor data, including ocean current U-components, V-components, and sea surface height data, from the training set are input into the dual-backbone network constructed in step S203. The base loss function is the mean absolute error loss (L1 Loss). The spatial error weight tensor W generated in step S202 is multiplied element-wise with the base loss function to obtain the weighted loss function: L = L1 ⊙ W; where W∈[0,1].

[0045] With the goal of minimizing the weighted loss function, all learnable parameters of the dual-backbone network are iteratively updated using the backpropagation algorithm and the Adam optimizer. During training, the model performance is monitored on the validation set. Training is stopped when the validation set loss no longer decreases for several consecutive epochs, the optimal model weights are saved, and the trained dual-backbone network model is obtained.

[0046] The process of spatial downscaling ocean data using the trained model is as follows: Low-resolution multi-element ocean data to be processed is acquired, including ocean current U-components, ocean current V-components, sea surface height, and sea surface temperature data, with a spatial resolution of 1 / 6°. The data is normalized to the 99th percentile maximum and minimum values, with normalization parameters consistent with those used in the training phase. The normalized ocean current U- and V-component data are input into the main feature backbone of the dual-backbone network model, while the normalized sea surface height and sea surface temperature data are input into the auxiliary feature backbone. The main feature backbone extracts features from the ocean current data, outputting a main feature tensor; the auxiliary feature backbone extracts features from the sea surface height and sea surface temperature data, outputting an auxiliary feature tensor. The convolution parameters of the two backbones are not shared, enabling differentiated modeling of the main and auxiliary features. The mid-layer feature fusion module concatenates the main and auxiliary feature tensors along the channel dimension and then fuses them using a 1×1 convolution to obtain a fused feature tensor. The deep reconstruction module performs deep feature extraction on the fused feature tensor and enhances information transmission through global residual connections. The upsampling module increases the feature tensor spatial resolution by a factor of 4, reaching 1 / 24°, through two-stage PixelShuffle operations. The output convolutional layer maps the upsized feature tensor into a 2-channel output, corresponding to the U and V components of the high-resolution ocean current, respectively. The output is then denormalized to restore the original physical dimensions, yielding the final high-resolution downscaled reconstruction of the ocean current.

[0047] To verify the effectiveness of the method of this invention, a certain ocean reanalysis dataset (a certain ocean forecasting dataset) was used as experimental data. The complete scheme in the above embodiments, which uses 99th percentile normalization and sea surface height and sea surface temperature as auxiliary factors (denoted as the error weighted dual backbone model DBWSR, or Dual Backbone Weighted Super-Resolution Network), was quantitatively compared with the following four schemes: Option 1 (Bicubic): A bicubic interpolation method that does not involve any deep learning model.

[0048] Option 2 (EDSR, Enhanced Deep Super-Resolution): A classic single-backbone depth residual network is used as the baseline model. 32 residual blocks are used. All input elements (sea surface current field, sea surface height and sea surface temperature) are stitched together in a unified channel and then input into the single-backbone network. The land-sea mask loss function is adopted.

[0049] Option 3 (EDSR-W, EDSR-Weighted, Weighted Enhanced Depth Residual Image Super-Resolution Network): Based on EDSR, the loss function is replaced with a weighted loss function generated based on the training error of single bone intervention.

[0050] Option 4 (DBSR, Dual-Backbone Super-Resolution): This method adopts the dual-backbone network structure of the present invention (16 residual blocks in the main backbone + 16 residual blocks in the auxiliary backbone), but does not use prior space weighting, that is, the loss function does not fuse the error weight tensor.

[0051] The evaluation metrics used were root mean square error (RMSE), correlation coefficient (R), accuracy of current direction classification (ACC), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). The experimental results are shown in the table below:

[0052] The experimental results show that the complete scheme (DBWSR) of this invention achieves the best performance in the core metric RMSE, improving by 78% compared to the bicubic interpolation method and by 5% compared to the classic single-backbone residual network (EDSR). The PSNR reaches 41.25 dB, an improvement of 2.89 dB compared to Bicubic and 0.85 dB compared to EDSR, indicating that this invention can more accurately recover the spatial details of ocean current fields. The SSIM reaches 0.9912, which, although a small improvement over existing methods, is still the highest among all comparative models. These results verify the synergistic effect of the dual-backbone differential feature extraction structure and the error weighting mechanism.

[0053] To more intuitively illustrate the technical effects of the present invention, the embodiments of the present invention also provide, as follows: Figure 4 The image shows a comparison of the visualization results of downscaling reconstruction of ocean current elements using multiple schemes.

[0054] The marine data spatial downscaling method based on a dual-backbone network provided in this invention has the following beneficial effects: By setting up a main feature backbone network and an auxiliary feature backbone network with non-shared parameters, the target features of ocean currents and auxiliary ocean features are extracted independently, which effectively avoids feature interference between different physical attribute features and enhances the model's ability to learn the differentiated features of multiple ocean features. At the same time, by using the spatial error weight tensor generated based on the pre-training error of the single backbone network to weight the loss function during the training process of the dual backbone network, the model can adaptively strengthen its focus on complex dynamic regions with high errors, thereby improving the overall accuracy of the downscaling reconstruction of the sea surface current field.

[0055] In the above embodiments, a method for spatial downscaling of marine data based on a dual-backbone network is provided. Correspondingly, the present invention also provides a device for spatial downscaling of marine data based on a dual-backbone network. The device for spatial downscaling of marine data based on a dual-backbone network provided in the embodiments of the present invention can implement the above-described method for spatial downscaling of marine data based on a dual-backbone network. This device can be implemented through software, hardware, or a combination of both. For example, the device can include integrated or separate functional modules or units to perform the corresponding steps in the above methods. Please refer to... Figure 5 As shown, the marine data spatial downscaling device 10 based on a dual-backbone network provided by the present invention includes: The acquisition module 101 is used to acquire low-resolution ocean multi-element data to be processed, the data including ocean current elements and auxiliary ocean elements, and to perform preset quantile maximum and minimum normalization processing on the data. The processing module 102 is used to input normalized low-resolution ocean multi-element data into a pre-trained dual-backbone network model; through the dual-backbone network model's dual-backbone feature extraction, feature fusion, deep reconstruction, upsampling and output convolution processing, high-resolution ocean current downscaling reconstruction results are output.

[0056] In some embodiments of the present invention, the ocean current elements include the U-component and V-component of sea surface current velocity; the auxiliary ocean elements include sea surface height data and / or sea surface temperature data.

[0057] The marine data spatial downscaling device based on a dual-backbone network provided in this invention embodiment and the marine data spatial downscaling method based on a dual-backbone network provided in this invention embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0058] The present invention also provides an electronic device corresponding to the method provided in the foregoing embodiments. The electronic device may be a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the above-described marine data spatial downscaling method based on a dual-backbone network.

[0059] Please refer to Figure 6 This illustrates a schematic diagram of an electronic device provided by some embodiments of the present invention. For example... Figure 6 As shown, the electronic device 20 includes: a processor 200, a memory 201, a bus 202, and a communication interface 203. The processor 200, the communication interface 203, and the memory 201 are connected via the bus 202. The memory 201 stores a computer program that can run on the processor 200. When the processor 200 runs the computer program, it executes the marine data spatial downscaling method based on a dual-backbone network provided in any of the foregoing embodiments of the present invention.

[0060] The memory 201 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 203 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0061] Bus 202 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory 201 is used to store programs. After receiving an execution instruction, the processor 200 executes the program. The marine data spatial downscaling method based on a dual-backbone network disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 200, or implemented by the processor 200.

[0062] The processor 200 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 200 or by instructions in software form. The processor 200 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 201. The processor 200 reads the information in memory 201 and, in conjunction with its hardware, completes the steps of the above method.

[0063] The electronic equipment provided in this embodiment of the invention and the marine data spatial downscaling method based on dual backbone network provided in this embodiment of the invention are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0064] The present invention also provides a computer-readable storage medium corresponding to the marine data spatial downscaling method based on a dual-backbone network provided in the foregoing embodiments, wherein a computer program (i.e., a program product) is stored thereon. When the computer program is run by a processor, it executes the marine data spatial downscaling method based on a dual-backbone network provided in any of the foregoing embodiments.

[0065] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0066] The computer-readable storage medium provided in the above embodiments of the present invention and the marine data spatial downscaling method based on dual backbone network provided in the embodiments of the present invention are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A spatial downscaling method for marine data based on a dual-backbone network, characterized in that, include: Acquire low-resolution multi-element ocean data to be processed, wherein the low-resolution multi-element ocean data includes ocean current elements as the target of downscaling reconstruction and auxiliary ocean elements as auxiliary constraints. The low-resolution ocean multi-element data is subjected to preset quantile maximum and minimum value normalization processing to obtain normalized low-resolution ocean multi-element data. The normalized low-resolution ocean multi-element data is input into a pre-trained dual-backbone network model, which includes: a main feature backbone network, an auxiliary feature backbone network, a mid-level feature fusion module, a deep reconstruction module, an upsampling module, and an output convolutional layer. The main feature backbone network is used to extract features from the ocean current elements to obtain the main feature tensor; at the same time, the auxiliary feature backbone network is used to extract features from the auxiliary ocean elements to obtain the auxiliary feature tensor; wherein the convolution parameters of the main feature backbone network and the auxiliary feature backbone network are not shared. The main feature tensor and the auxiliary feature tensor are fused by the mid-layer feature fusion module to obtain a fused feature tensor; The deep reconstruction module performs deep feature extraction on the fused feature tensor, and the upsampling module enhances the spatial resolution of the feature tensor after deep feature extraction. The output convolutional layer performs ocean current element channel mapping on the feature tensor after resolution enhancement, and performs inverse normalization on the channel mapping output to output high-resolution ocean current downscaling reconstruction results.

2. The marine data spatial downscaling method based on a dual-backbone network according to claim 1, characterized in that, The pre-training process of the dual-backbone network model includes the following steps: S1. Construction of marine dataset: The original high-resolution marine data is sequentially gridded and quality controlled, and then low-resolution marine data from the same source as the high-resolution marine data is obtained; the low-resolution marine data and the high-resolution marine data are normalized by preset quantile maximum and minimum normalization to obtain training set, test set and validation set; wherein, the dataset contains ocean current elements as the reconstruction target, and auxiliary marine elements as constraints. S2. Construct and train a single backbone network: Input ocean current elements from the training set into a preset single backbone network for downscaling reconstruction, calculate the root mean square error between the reconstruction results and the corresponding high-resolution label data to obtain the global average spatial error distribution under the training set; perform preset quantile max-min normalization on the spatial error distribution to generate an error weight tensor. S3. Construct a dual-backbone network: The dual-backbone network includes an independent main feature backbone network, an auxiliary feature backbone network, a mid-layer feature fusion module, a deep reconstruction module, an upsampling module, and an output convolutional layer; the main feature backbone network is used to extract ocean current feature features, and the auxiliary feature backbone network is used to extract auxiliary ocean feature features. The parameters of the two backbone networks are not shared. S4. Weighted training of dual-backbone network: Input multi-element data containing ocean current elements and auxiliary ocean elements from the training set into the dual-backbone network, and use a weighted loss function that fuses the error weight tensor to iteratively optimize the dual-backbone network until the preset convergence condition is met, and obtain the trained dual-backbone network model.

3. The marine data spatial downscaling method based on a dual-backbone network according to claim 2, characterized in that, In step S2, the single backbone network includes an input convolutional layer, a deep feature extraction backbone composed of multiple residual feature extraction modules stacked together, an upsampling module, and an output convolutional layer connected in sequence; the single backbone network is operated using the land-sea mask loss function, with the land area mask value set to 0 and the ocean area mask value set to 1.

4. The marine data spatial downscaling method based on a dual-backbone network according to claim 2, characterized in that, In step S3, the structures of both the main feature backbone network and the auxiliary feature backbone network are as follows: The input convolutional layer uses a 3×3 convolutional kernel followed by a ReLU activation function to map the input data to a high-dimensional feature space; Sixteen residual feature extraction modules are connected in series. The internal structure of each residual feature extraction module is as follows: a first 3×3 convolutional layer, a ReLU activation function, and a second 3×3 convolutional layer. A residual connection is provided between the input of the module and the output of the second 3×3 convolutional layer.

5. The marine data spatial downscaling method based on a dual-backbone network according to claim 2, characterized in that, In step S3, the deep reconstruction module is composed of 16 shared residual feature extraction modules and a 3×3 feature mapping convolutional layer connected in sequence, and a global residual connection is provided between the input of the deep reconstruction module and the output of the feature mapping convolutional layer.

6. The marine data spatial downscaling method based on a dual-backbone network according to claim 2, characterized in that, In step S3, the upsampling module is composed of two levels of subpixel convolutional upsampling units cascaded in sequence; each level of subpixel convolutional upsampling unit includes a 3×3 convolutional layer, a PixelShuffle spatial rearrangement layer with a magnification of 2, and a ReLU activation function. A single level achieves a 2x improvement in spatial resolution, and two levels cascaded achieve a 4x improvement in spatial resolution.

7. The marine data spatial downscaling method based on a dual-backbone network according to any one of claims 1 to 6, characterized in that, The ocean current elements include the U and V components of sea surface current velocity; the auxiliary ocean elements include sea surface height data and / or sea surface temperature data.

8. A marine data spatial downscaling device based on a dual-backbone network, characterized in that, include: The acquisition module is used to acquire low-resolution ocean multi-element data to be processed, the data including ocean current elements and auxiliary ocean elements, and to perform preset quantile maximum and minimum normalization processing on the data. The processing module is used to input normalized low-resolution ocean multi-element data into a pre-trained dual-backbone network model; through the dual-backbone network model, the dual-backbone feature extraction, feature fusion, deep reconstruction, upsampling and output convolution processing are performed to output high-resolution ocean current downscaling reconstruction results.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-readable instructions that can be executed by a processor to implement the method as described in any one of claims 1 to 7.

Images