An ionospheric super-resolution refinement reconstruction method and device based on an improved ViT model, equipment and storage medium
By using an improved ViT model, combined with multi-channel tensors and mask layers, the problems of ground site dependence and smoothing in traditional ionospheric modeling are solved, enabling high-resolution ionospheric image reconstruction and improving model accuracy.
Patent Information
- Application Number
- CN202510286920.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-11-07
- Estimated Expiration
- 2045-03-12
AI Technical Summary
Traditional ionospheric modeling relies on a large number of ground stations and suffers from excessive image smoothing, which affects model accuracy.
By employing an improved ViT model, combining multi-channel tensors and mask layers, and utilizing a Transformer encoder and upsampling module, high-resolution reconstruction of ionospheric images is achieved through self-attention mechanism and convolutional layer processing.
It effectively reduces the dependence on the number of ground stations, improves the spatial resolution and model accuracy of ionospheric images, and overcomes the drawback of overly smooth images.
Smart Images
Figure CN120219168B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to an ionospheric super-resolution refinement reconstruction method and device based on an improved ViT model, an ionospheric modeling device, and a storage medium. BACKGROUND
[0002] Traditional ionospheric TEC modeling usually relies on satellite observation data of ground GNSS stations and is based on mathematical models (such as spherical harmonics) for modeling, which often relies on a large number of ground stations and is based on mathematical functions (such as spherical harmonics) for modeling, which also faces the disadvantage of excessive smoothing.
[0003] With the development of artificial intelligence, in computer image processing, super-resolution is a technology for converting low-resolution blurred images into high-resolution clear images. Introducing the super-resolution technology into ionospheric model construction is beneficial to overcome the shortcomings of traditional models relying on a large number of ground stations and excessive smoothing of the model. SUMMARY
[0004] The purpose of the present application is to overcome the shortcomings of the prior art, provide an ionospheric super-resolution refinement reconstruction method based on an improved ViT model, apply ViT to the super-resolution reconstruction of the ionosphere, effectively reduce the dependence on the number of ground stations, overcome the disadvantage of excessive smoothing of the image, and improve the spatial resolution of the ionospheric image, thereby improving the accuracy of the ionospheric model.
[0005] To achieve the above purpose, the application is implemented by using the following technical scheme:
[0006] In a first aspect, the application provides an ionospheric super-resolution refinement reconstruction method based on an improved ViT model, comprising:
[0007] Obtaining ionospheric TEC data and VTEC data in a to-be-measured time period;
[0008] The ionospheric TEC data is used as a low-resolution image, the VTEC data is used as a pixel real value, and a mask layer is generated according to whether the pixel real value includes an observation value;
[0009] The low-resolution image, the pixel real value, and the mask layer are combined to obtain a multi-channel tensor;
[0010] The multi-channel tensor is used as an input, and a global ionospheric map is output based on the improved ViT model;
[0011] The improved ViT model is improved based on a ViT model, the ViT model comprises a single-channel input layer, an image feature embedding module, a Transformer encoder, an MLP classification module and an output layer connected in sequence, the single-channel input layer is replaced by a multi-channel input layer, and the MLP classification module is replaced by an up-sampling module to obtain the improved ViT model.
[0012] Further, the ionospheric TEC data is provided by an IRI2020 model, and the VTEC data is provided by a ground GNSS station.
[0013] Further, the mask layer is generated according to whether the pixel real value includes an observation value, and the method comprises the following steps:
[0014] The pixel real value including the observation value is marked as valid, and the pixel real value not including the observation value is marked as invalid, the valid is represented by 1, and the invalid is represented by 0.
[0015] The marking result of the pixel real value constitutes a binary mask layer.
[0016] Further, the Transformer encoder comprises a plurality of Transformer encoding layers, each of the Transformer encoding layers is used to take the pixel real value and the mask layer as additional channel information, and dynamically adjusts global information capture to obtain low-resolution features by using the additional channel information as weights through a multi-head self-attention mechanism.
[0017] Further, the up-sampling module comprises an up-sampling unit and a convolution layer connected in sequence, the up-sampling unit adopts a deconvolution layer or pixel recombination, and the up-sampling module is used to map the low-resolution features to a high-resolution image.
[0018] Further, the improved ViT model is pre-trained, and the pre-training method comprises the following steps:
[0019] Obtaining historical ionospheric TEC data and VTEC data;
[0020] Dividing the ionospheric TEC data and the VTEC data corresponding to the time to obtain a training set and a test set;
[0021] Taking the training set data as input, training the improved ViT model, determining the optimal parameters of the improved ViT model, and obtaining a preliminary improved ViT model;
[0022] Taking the test set data as input, testing the improved ViT model to obtain a test result;
[0023] According to the test result, an evaluation index is calculated, and the pre-trained improved ViT model is obtained based on the evaluation index.
[0024] In a second aspect, the present application also provides an ionosphere super-resolution refinement reconstruction device based on an improved ViT model, comprising:
[0025] a data acquisition module configured to acquire ionosphere TEC data and VTEC data in a to-be-tested time period;
[0026] a data processing module configured to take the ionosphere TEC data as a low-resolution image, take the VTEC data as a pixel real value, generate a mask layer according to whether the pixel real value includes an observation value, and combine the low-resolution image, the pixel real value and the mask layer to obtain a multi-channel tensor;
[0027] an ionosphere reconstruction module configured to take the multi-channel tensor as an input and output a global ionosphere map based on the improved ViT model.
[0028] In a third aspect, the present application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the ionosphere super-resolution refinement reconstruction method based on the improved ViT model according to any one of the above aspects.
[0029] In a fourth aspect, the present application also provides a device, comprising:
[0030] a memory configured to store instructions;
[0031] a processor configured to execute the instructions so that the device implements the ionosphere super-resolution refinement reconstruction method based on the improved ViT model according to any one of the above aspects.
[0032] Compared with the prior art, the present application has the following beneficial effects:
[0033] The present application effectively captures the spatial dependence of the ionosphere image by using the global self-attention mechanism of the ViT model, and introduces the pixel real value and the mask layer to enhance the reconstruction ability of the model for different data regions, and adopts multi-channel fusion, upsampling operation and convolution layer processing to ensure the high resolution and detail accuracy of the reconstructed image.
[0034] The present application applies the ViT model to the super-resolution reconstruction of the ionosphere, which can effectively reduce the dependence on the number of ground stations, overcome the defect of excessive smoothness of the image, and improve the spatial resolution of the ionosphere image, thereby improving the accuracy of the ionosphere model. BRIEF DESCRIPTION OF DRAWINGS
[0035] Figure 1 FIG. 1 is a flowchart of an ionosphere super-resolution refinement reconstruction method based on an improved ViT model according to an embodiment of the present application. DETAILED DESCRIPTION
[0036] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention. Example 1
[0037] like Figure 1 As shown, this embodiment of the invention provides a method for super-resolution refinement reconstruction of the ionosphere based on an improved ViT model, including the following steps:
[0038] First, acquire ionospheric TEC data from the IRI2020 model and VTEC data from ground GNSS stations within the time period to be measured.
[0039] Ionospheric TEC data is used as a low-resolution image, and VTEC data is used as the pixel ground truth. Then, a mask layer is generated based on whether the pixel ground truth includes the observation value. Pixel ground truth containing the observation value is marked as valid (1), and pixel ground truth not containing the observation value is marked as invalid (0). The marking results of all pixel ground truth values are combined into a binary mask layer.
[0040] The low-resolution image, the actual pixel value, and the mask layer are merged to form a multi-channel tensor. The shape of the multi-channel tensor is [B,C,H,W], where B represents the batch size, C represents the number of channels, H represents the image height, and W represents the image width.
[0041] An improved ViT model is constructed, which is an improvement upon the existing ViT model. The ViT model consists of a single-channel input layer, an image feature embedding module, a Transformer encoder, an MLP classification module, and an output layer connected in sequence. In this embodiment, the single-channel input layer is replaced with a multi-channel input layer, and the MLP classification module is replaced with an upsampling module, resulting in the improved ViT model. The specific data processing steps include:
[0042] The multichannel tensor is input from the multichannel input layer. The image embedding module divides the multichannel tensor into several small patches. Each patch is P×P in size. Each flattened patch is embedded into a high-dimensional space to form the input sequence.
[0043] The Transformer encoder consists of multiple Transformer coding layers. Each Transformer coding layer uses the real pixel value and mask layer as additional channel information, and uses the additional channel information as weights to dynamically adjust the global information of the input sequence through a multi-head self-attention mechanism to capture low-resolution features.
[0044] The upsampling module adopts an upsampling unit and a convolution layer connected in sequence, the upsampling unit adopts a deconvolution layer or pixel recombination, the low-resolution feature is mapped to a high-resolution image through the upsampling module, local detail information is retained, further fine reconstruction is performed, and finally a high-resolution global ionospheric map is output through an output layer.
[0045] The improved ViT model is pre-trained, and the pre-training method comprises the following steps:
[0046] IRI2020 model provides ionospheric TEC data as low-resolution images for 11 years (1 solar activity cycle) from 2010 to 2020, VTEC data obtained from the ground GNSS station during the same period as the pixel true value, and a mask layer is generated according to whether the pixel true value includes an observation value, and the low-resolution image, the pixel true value and the mask layer during the same period constitute a data set.
[0047] The data set is divided into a training set and a test set according to a certain proportion.
[0048] The training set data is used as input to train the improved ViT model, determine the optimal parameters of the improved ViT model, and obtain a preliminary improved ViT model.
[0049] The test set data is used as input to test the improved ViT model to obtain a test result.
[0050] According to the test result, an evaluation index is calculated, and a pre-trained improved ViT model is obtained based on the evaluation index.
[0051] The multi-channel tensor composed of the low-resolution image, the pixel true value and the mask layer is used as input, and the global ionospheric map (GIM) of the European Center for Medium-Range Weather Forecasts (CODE) can be output based on the pre-trained improved ViT model. Embodiment 2
[0052] Based on the embodiment 1, the embodiment further provides an ionospheric super-resolution refinement reconstruction device based on an improved ViT model, comprising:
[0053] The data acquisition module is configured to acquire ionospheric TEC data and VTEC data in a to-be-tested time period.
[0054] The data processing module is configured to use the ionospheric TEC data as a low-resolution image, the VTEC data as a pixel true value, and generate a mask layer according to whether the pixel true value includes an observation value, and combine the low-resolution image, the pixel true value and the mask layer to obtain a multi-channel tensor.
[0055] The ionospheric reconstruction module is configured to take a multi-channel tensor as input, and output a global ionospheric map based on the improved ViT model. Embodiment 3
[0056] Based on the embodiment 1, the embodiment further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the ionospheric super-resolution refinement reconstruction method based on the improved ViT model. Embodiment 4
[0057] Based on the embodiment 1, the embodiment further provides a device, which is characterized in that comprising:
[0058] a memory for storing instructions;
[0059] a processor for executing the instructions, so that the device executes the ionospheric super-resolution refinement reconstruction method based on the improved ViT model.
[0060] The above only describes the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the technical principles of the present application, a number of improvements and modifications can be made, and these improvements and modifications should be considered as the protection scope of the present application.
Claims
1. An ionospheric super-resolution refinement reconstruction method based on an improved ViT model, characterized in that, The method comprises the following steps: obtaining ionospheric TEC data and VTEC data in a to-be-tested time period; regarding the ionospheric TEC data as a low-resolution image, regarding the VTEC data as a pixel real value, and generating a mask layer according to whether the pixel real value includes an observation value; merging the low-resolution image, the pixel real value and the mask layer to obtain a multi-channel tensor; inputting the multi-channel tensor, and outputting a global ionospheric map based on an improved ViT model; the improved ViT model is obtained by improving a ViT model, wherein the ViT model comprises a single-channel input layer, an image feature embedding module, a Transformer encoder, an MLP classification module and an output layer connected in sequence, the single-channel input layer is replaced by a multi-channel input layer, and the MLP classification module is replaced by an up-sampling module to obtain the improved ViT model; the Transformer encoder comprises a plurality of Transformer encoding layers, each of the Transformer encoding layers is used for taking the pixel real value and the mask layer as additional channel information, and dynamically adjusting global information capture by using the additional channel information as weight through a multi-head self-attention mechanism to obtain a low-resolution feature; the up-sampling module comprises an up-sampling unit and a convolution layer connected in sequence, the up-sampling unit adopts a deconvolution layer or pixel recombination, and the up-sampling module is used for mapping the low-resolution feature to a high-resolution image.
2. The ionospheric super-resolution refinement reconstruction method based on the improved ViT model according to claim 1, characterized in that, The ionospheric TEC data is provided by an IRI2020 model, and the VTEC data is provided by a ground GNSS station.
3. The ionospheric super-resolution refinement reconstruction method based on the improved ViT model according to claim 1, characterized in that, The mask layer is generated according to whether the pixel real value includes the observation value, and the method comprises the following steps: marking the pixel real value containing the observation value as valid, and marking the pixel real value not containing the observation value as invalid, wherein the valid is represented by 1, and the invalid is represented by 0; the marking result of the pixel real value constitutes a binary mask layer.
4. The ionospheric super-resolution refinement reconstruction method based on the improved ViT model according to claim 1, characterized in that, The method further comprises pre-training the improved ViT model, and the pre-training method comprises the following steps: obtaining historical ionospheric TEC data and VTEC data; dividing the ionospheric TEC data and the VTEC data corresponding to the time to obtain a training set and a test set; inputting the training set data to train the improved ViT model, determining optimal parameters of the improved ViT model, and obtaining a preliminary improved ViT model; inputting the test set data to test the improved ViT model to obtain a test result; calculating an evaluation index according to the test result, and obtaining the improved ViT model after pre-training based on the evaluation index.
5. An ionospheric super-resolution refinement reconstruction device based on an improved ViT model, characterized in that, The method comprises the following steps: a data acquisition module configured to obtain ionospheric TEC data and VTEC data in a to-be-tested time period; a data processing module configured to regard the ionospheric TEC data as a low-resolution image, regard the VTEC data as a pixel real value, generate a mask layer according to whether the pixel real value includes an observation value, and merge the low-resolution image, the pixel real value and the mask layer to obtain a multi-channel tensor; an ionospheric reconstruction module configured to input the multi-channel tensor, and output a global ionospheric map based on an improved ViT model. The improved ViT model is improved based on a ViT model, the ViT model comprises a single-channel input layer, an image feature embedding module, a Transformer encoder, an MLP classification module and an output layer connected in sequence, the single-channel input layer is replaced by a multi-channel input layer, and the MLP classification module is replaced by an up-sampling module to obtain the improved ViT model; The Transformer encoder comprises a plurality of Transformer encoding layers, each Transformer encoding layer is used for taking the pixel real value and the mask layer as additional channel information, and dynamically adjusting global information capture by a multi-head self-attention mechanism to obtain a low-resolution feature by taking the additional channel information as a weight; The up-sampling module comprises an up-sampling unit and a convolution layer connected in sequence, the up-sampling unit adopts a deconvolution layer or pixel recombination, and the up-sampling module is used for mapping the low-resolution feature to a high-resolution image.
6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor, and the ionospheric super-resolution refinement reconstruction method based on the improved ViT model is realized.
7. An apparatus, comprising: Comprise: a memory for storing instructions; a processor for executing the instructions, so that the device executes the ionospheric super-resolution refinement reconstruction method based on the improved ViT model.