Ionized layer TEC image modeling method and system based on DCGAN and Poisson fusion

Abstract

The invention belongs to the technical field of ionosphere TEC data processing and global ionosphere modeling, and discloses an ionosphere TEC image modeling method and system based on DCGAN and Poisson fusion, focused on an image completion technology, observable global ionosphere data is converted into a TEC image by using an image processing method, data set division is carried out on the image data, and the data set division is carried out on the TEC image. According to the method, a global ionosphere TEC image data set is obtained through a global ionosphere TEC model, a DCGAN model and a Poisson fusion algorithm are used for training the data set and input random vectors, a final TEC image completing image completion and boundary elimination is obtained, and therefore ionosphere modeling completion and high-precision global ionosphere TEC modeling are achieved. The invention provides a possible method for solving the problems that more ionosphere local information is difficult to effectively capture when global ionosphere modeling is carried out by a traditional mathematical method, and the time consumption is long and the precision is low when a complete TEC map is obtained due to limited ground-based GNSS coverage range and data integration requirements.

Description

Technical Field

[0001] The present invention belongs to the technical field of ionospheric TEC data processing and global ionospheric modeling, and in particular relates to an ionospheric TEC image modeling method and system based on DCGAN and Poisson fusion. Background Art

[0002] The ionosphere usually refers to the upper atmosphere of the Earth from about 60 km above the ground to the top of the magnetosphere. Its activity level is gradually increasing as the current solar activity enters a new round of active cycle. At present, the intense activity of the ionosphere and the ionospheric delay error have brought serious impacts on the services and applications of satellite navigation systems. In this context, it is of great significance to establish a high-precision global ionospheric TEC model and predict the spatiotemporal distribution and change state of the ionospheric total electron content (TEC) to mitigate its impact on the performance of high-precision navigation and positioning services.

[0003] For a long time, the global ionospheric TEC modeling research has been carried out around empirical models and mathematical models. However, global empirical models are often constructed through the understanding and simulation of physical processes. Although the calculation is simple, the model correction accuracy is low; the ionospheric mathematical model uses global observation data to fit the plane or surface to obtain the overall optimal result, which leads to the results of the model in some local areas that are seriously inconsistent with the actual situation. In addition, the mathematical fitting model is often based on empirical understanding and cognition, and does not deeply explore the potential characteristics of the ionospheric structure and changes from the perspective of big data, resulting in low reliability of the modeling results. At the same time, the main source of observation data for traditional modeling methods is ground-based GNSS observation data, but in fact, GNSS ground tracking stations are unevenly distributed around the world, especially in land with harsh natural conditions and ocean areas that account for 70% of the earth's surface. There is basically no observation data, resulting in low modeling accuracy. In short, due to the irregular characteristics of the ionosphere, the complex mechanism of local changes, and the problem of missing observation data, mathematical models and empirical models cannot establish a high-precision global ionospheric TEC model, and cannot accurately describe the spatiotemporal changes of the ionosphere and its complex large-scale anomalies.

[0004] In recent years, with the continuous iteration and update of computer hardware, the computing power has been enhanced, and the development of artificial intelligence (AI) has made great progress. The sources of ionospheric TEC data are rich and the accumulated amount is huge, which has created good conditions for the application of AI technology. Deep learning technology has high application value in areas where accurate models cannot be established. It is one of the important possible methods to solve the problem of global ionospheric modeling and data fusion. Since the changes in the ionosphere are greatly affected by various environmental factors, the advantage of deep learning technology is that it does not require manual design of rules. The use of neural network models can mine more potential features of the data, which is of great significance for obtaining a global ionospheric TEC model with higher accuracy and reliability.

[0005] After the generative adversarial network (GAN) was proposed, some scholars gradually tried to use related technologies to complete the areas where ionospheric observations were missing to model the global ionosphere. For example, some scholars proposed the global and local generative adversarial network (GLGAN) method and applied it to global TEC modeling. This model consists of a global discriminator and a local discriminator, which has a good effect on the irregular local repair of TEC images, but its disadvantage is that it cannot repair large areas of missing TEC data. In short, how to combine multi-source ionospheric observation data and today's many cutting-edge AI technologies to model the global ionosphere with high precision is the current research focus and key. Summary of the invention

[0006] In view of the above problems, the present invention focuses on image completion technology, and uses the globally observable ionospheric data to provide an ionospheric TEC image modeling method and system based on DCGAN and Poisson fusion to achieve ionospheric modeling completion and high-precision global ionospheric TEC modeling.

[0007] According to one aspect of the present invention, there is provided an ionospheric TEC image modeling method based on DCGAN and Poisson fusion, comprising: Obtain observable global ionospheric TEC data and convert them into TEC images; The converted TEC image is input into the trained deep convolutional adversarial generative network model for image completion; Based on the constructed Poisson equation, the TEC values ​​of the image completion area and the TEC values ​​around the missing data area are Poisson fused to obtain the TEC values ​​of the missing data area, and the completed TEC image is smoothed to obtain the final TEC image. The final TEC image is used to perform ionospheric TEC modeling; wherein the training of the deep convolution adversarial generation network model comprises: Construct TEC image dataset; A deep convolutional generative adversarial network model is constructed to input a random vector into the generator to obtain an initial TEC image, and the generated initial TEC image and the TEC image of the dataset are input into the discriminator to extract high-dimensional features and judge the authenticity, and then the completed TEC image is obtained through mapping processing; Use the test set and validation set in the dataset for testing and verification, and output the trained deep convolutional adversarial generation network model.

[0008] As a further technical solution, a TEC image dataset is constructed, including: Get TEC image datasets of global ionospheric TEC data; Based on the solar activity index F10.7, the test set was divided; Based on the K-fold cross-validation method, the training set and the validation set are obtained.

[0009] As a further technical solution, a deep convolutional adversarial generative network model is constructed, including: Following the architectural guidelines of Stable Deep Convolutional GANs, replace all pooling layers in the discriminator and generator, and use batch normalization in both the generator and the discriminator; Remove all hidden layers in the fully connected layer; In the generator, all layers except the output layer use the ReLU activation function, and the Tanh activation function is used in the output layer; In the discriminator, all layers except the output layer use the LeakyReLU activation function, and the Sigmoid activation function is used in the output layer.

[0010] As a further technical solution, a random vector is input into a generator to obtain an initial TEC image, and the generated initial TEC image and the TEC image of the data set are input into a discriminator to perform high-dimensional feature extraction and authenticity judgment, including: Input the random vector into the generator, use the input layer of the generator to project and reshape the random vector into high-dimensional features, use three hidden layers to convert the high-dimensional features into details of the TEC image, and use the output layer to output the initial TEC image; The generated initial TEC image and the TEC image of the dataset are input into the discriminator, and the TEC image is read using the input layer and three hidden layers, and high-dimensional features are extracted from the TEC image. The high-dimensional features are reshaped into a one-dimensional vector using the output layer, and then the discriminant value is generated using the fully connected layer and the Sigmoid activation function for judging the authenticity of the TEC image.

[0011] As a further technical solution, a complete TEC image is obtained through mapping processing, including: The TEC values ​​of the generated initial TEC image are mapped to the area in the dataset where the real TEC image corresponds to and where the data is missing, and the TEC values ​​corresponding to the pixels around the missing area are matched.

[0012] As a further technical solution, the mapped values ​​processed by the mapping and the TEC values ​​around the missing data area are Poisson fused to obtain the TEC values ​​of the missing data area, including: Based on the constructed Poisson equation, the gradient of the TEC value around the missing data area is calculated, and the gradient of the TEC value of the completed image is calculated; Based on the minimization of Poisson's equation, the TEC value of the missing data area is calculated using the obtained gradient information.

[0013] According to one aspect of the present invention, there is provided an ionospheric TEC image modeling system based on DCGAN and Poisson fusion, comprising: An image acquisition module is used to acquire observable global ionospheric TEC data and convert it into TEC images; The image completion module is used to input the converted TEC image into the trained deep convolutional adversarial generative network model for image completion; An image fusion module is used to perform Poisson fusion on the TEC values ​​of the image completion area and the TEC values ​​around the missing data area based on the constructed Poisson equation to obtain the TEC values ​​of the missing data area, and use this to smooth the completed TEC image to obtain the final TEC image; A TEC modeling module is used to perform ionospheric TEC modeling using the final TEC image; wherein the training of the deep convolution adversarial generative network model includes: Construct TEC image dataset; A deep convolutional generative adversarial network model is constructed to input a random vector into the generator to obtain an initial TEC image, and the generated initial TEC image and the TEC image of the dataset are input into the discriminator to extract high-dimensional features and judge the authenticity, and then the completed TEC image is obtained through mapping processing; Use the test set and validation set in the dataset for testing and verification, and output the trained deep convolutional adversarial generation network model.

[0014] According to one aspect of the present invention, there is provided an ionospheric TEC image modeling device based on DCGAN and Poisson fusion, comprising a memory and a processor. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the ionospheric TEC image modeling method based on DCGAN and Poisson fusion.

[0015] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the steps of the ionospheric TEC image modeling method based on DCGAN and Poisson fusion.

[0016] The ionospheric TEC image modeling method based on DCGAN and Poisson fusion proposed in the present invention focuses on TEC image completion, combines the neural network model with the image processing technology, considers image regeneration while taking into account image fusion, so as to achieve TEC image completion and high-precision global ionospheric TEC modeling. Compared with the existing technology, it has the following advantages: 1) Focusing on image completion, the TEC images are processed and analyzed as two-dimensional images, and the neural network model and image fusion technology in deep learning are used to complete the TEC images, which can solve the problem of low accuracy and low reliability of global ionospheric TEC modeling caused by missing data in some areas of the TEC images to a certain extent; 2) The deep convolutional generative adversarial network (DCGAN) model is used, which completes a number of key architectural improvements based on GAN, making the model more powerful in image processing, generating better image quality, and more stable in training; in addition, in order to achieve a more stable training process, the present invention follows the "Architecture Guidelines for Stable Deep Convolutional GANs" when building the model; 3) After the DCGAN model is randomly initialized to generate a global TEC image, the TEC values ​​of the generated TEC image are mapped to the corresponding data-missing regions of the real TEC image using a joint context loss and prior loss method, ensuring that the global appearance of the completed TEC image is maintained while maintaining high mapping accuracy; 4) After the DCGAN model completes the TEC image, in order to further optimize the image quality, the Poisson fusion algorithm is used for post-processing. By solving the partial differential equation, the image is smoothed and the boundaries are visually eliminated under the guidance of the image gradient information to solve the problem of mismatch between the boundary gradient of the TEC image completion area and the original observation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction is given below to the drawings used in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic flow chart of an ionospheric TEC image modeling method based on DCGAN and Poisson fusion provided in an embodiment of the present invention.

[0019] Figure 2 A schematic diagram of the structure of a deep convolutional generative adversarial network (DCGAN) model provided for an embodiment of the present invention, wherein the top number represents the number of channels of the convolution and the bottom number represents the dimension of the output matrix.

[0020] Figure 3 A flow chart of ionospheric modeling based on the DCGAN model and Poisson fusion algorithm provided in an embodiment of the present invention.

[0021] FIG4 (a) is a schematic diagram of a TEC image drawn by CODE 2014 DOY 292 GIMs provided by an embodiment of the present invention.

[0022] FIG4( b ) is a schematic diagram showing the comparison of the effects of redrawing the CODE TEC image with local time as the horizontal coordinate provided by an embodiment of the present invention, wherein the horizontal coordinate of the left image is longitude, and the horizontal coordinate of the right image is local time.

[0023] FIG4( c ) is a schematic diagram of TEC images to be completed after being covered with different types of masks in different years in the CODE TEC image test set provided by an embodiment of the present invention.

[0024] Fig. 4 (d) is a schematic diagram of a global TEC image obtained by using a DCGAN model to complete the image in combination with Poisson fusion post-processing under simulation data provided by an embodiment of the present invention, wherein from left to right are images of different years: CODE TEC image; image after completion of the first type of mask data; image after completion of the second type of mask data; image after completion of the third type of mask data.

[0025] FIG4( e ) is a schematic diagram of TEC images of the CODE TEC image test set provided by an embodiment of the present invention at 18:00 UTC on DOY 297 in different years based on MIT real data masks.

[0026] FIG4( f ) is a schematic diagram of simulated data provided by an embodiment of the present invention, which includes, from top to bottom, CODE TEC images of different years, TEC images of different years completed using the DCGAN model based on real data, and TEC images of different years completed using the DCGAN model based on real data and subjected to Poisson fusion. DETAILED DESCRIPTION

[0027] The present invention provides a global ionosphere modeling method based on a deep convolutional generative adversarial network and Poisson fusion. The method can solve the limitations of traditional mathematical methods for global ionosphere modeling, such as difficulty in effectively capturing more local ionosphere information, low modeling efficiency and low precision, and realize ionosphere modeling completion and high-precision global ionosphere TEC modeling. The constructed ionosphere modeling model based on the DCGAN model and the Poisson fusion algorithm has strong practicality and high generalization, and can be applied to other observable global TEC data.

[0028] The overall process of the present invention is as follows: first, the observable global ionospheric TEC data is obtained and converted into a TEC image; then, the converted TEC image is input into the trained deep convolutional generative adversarial network model for image completion; then, based on the constructed Poisson equation, the TEC values ​​of the image completion area and the TEC values ​​around the missing data area are Poisson fused to obtain the TEC values ​​of the missing data area, and the completed TEC image is smoothed to obtain the final TEC image; finally, the final TEC image is used for ionospheric TEC modeling.

[0029] The training and Poisson fusion of the deep convolutional adversarial generative network model include the following processes: (1) Obtain observable global ionospheric TEC data, use image processing methods to plot the raw data into stable and effective TEC images, and perform preprocessing.

[0030] (2) All TEC images in step (1) are combined to form a complete data set, and the data set is divided based on the solar activity index F10.7.

[0031] (3) A deep convolutional generative adversarial network (DCGAN) model is constructed. A random vector is input into the generator to obtain an initial TEC image. The TEC image generated by the generator and the TEC image of the data set in step (2) are then input into the discriminator for high-dimensional feature extraction and authenticity judgment. The completed TEC image is then obtained through mapping processing.

[0032] (4) Construct the Poisson equation. The TEC values ​​of the missing data area are obtained by using the mapped values ​​in step (3) and the TEC values ​​around the missing data area through the Poisson fusion algorithm. The generated TEC image is smoothed to obtain the final TEC image for ionospheric completion modeling.

[0033] (5) In the test set of the dataset, the final TEC image in step (4) is evaluated and verified using both simulated data and real data.

[0034] In order to make the purpose, technical scheme and advantages of the embodiment of the present invention clearer, the technical scheme in the embodiment of the present invention will be clearly and completely described in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments or single embodiments provided by the present invention are arbitrarily combined with each other to form a new technical solution. This combination is not subject to the constraints of the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in this field to achieve. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that this combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0035] The overall process of generating global ionospheric TEC images based on deep convolutional generative adversarial networks and Poisson fusion is as follows: Figure 1 As shown, the specific steps of the present invention will be described in detail below in conjunction with specific implementation methods.

[0036] In order to illustrate the feasibility and strong generalization of this method more specifically, the GIMs file released by the CODE Analysis Center was selected as the original global ionospheric TEC data, and the corresponding TEC image was drawn using GIMs. The DCGAN model image completion and Poisson fusion algorithm were used to process it, and the results were evaluated and verified.

[0037] 1. Data input and preprocessing After obtaining observable global ionospheric TEC data, the following steps are performed: Step (1): Use image processing methods to draw the original data into a stable and effective TEC image. Common methods include: using TEC data released by the organization to draw the image or directly using a pre-drawn TEC image.

[0038] Taking the GIMs file released by the CODE Analysis Center as an example, we use it as the original global ionospheric TEC data, and use GIMs to draw the corresponding TEC image to obtain the CODE TEC image set. The schematic diagram of the TEC image drawn by the 2014 DOY 292 GIMs is shown in Figure 4 (a).

[0039] Step (2): Preprocessing the TEC image, usually including: using interpolation methods to adjust the size of the image, specifying an outlier processing method, optimizing the image's horizontal and vertical coordinate representation according to the actual distribution of the data, etc.

[0040] Taking the CODE TEC image set as an example, when drawing the image, the 73 × 71 matrix is ​​adjusted to a 64 × 64 matrix using the two-dimensional cubic interpolation method, and the data of the outliers (greater than 999TECU) in the GIMs are used as blanks; after the drawing is completed, in order to ensure the spatial consistency and stability of the TEC value and the local time in the image and to achieve effective DCGAN model training, the global ionospheric TEC image is redrawn with the local time as the horizontal coordinate. The comparison of the redrawing effect is shown in Figure 4 (b).

[0041] 2. Dataset Division All preprocessed TEC images are combined to form a complete data set, and the data set is divided based on the solar activity index F10.7. The specific implementation steps of the data set division are as follows: For the test set, under the premise of taking the solar activity index F10.7 as the indicator, it is necessary to ensure that the test set contains data representing moderate solar activity, intense solar activity, and calm solar activity; for the training set and validation set, the K-fold cross-validation method is used for division, where K is selected as 10.

[0042] Taking the CODE image set as an example, from the 18-year data set, half of the data (July to December) in 2012, 2014 and 2018 were selected as the test set, representing the years with moderate solar activity, intense solar activity and calm solar activity, respectively. The other 15 years of data were used as the training set for training and verification. During the training process, the K-fold cross-validation (K=10) method was used to evaluate and optimize the model.

[0043] (III) Building a DCGAN model and performing image completion The specific implementation steps are as follows: Step (1): Build a DCGAN model.

[0044] When building the deep convolutional generative adversarial network (DCGAN) model, follow the "Architecture guidelines for stable Deep Convolutional GANs", including: replacing all pooling layers with convolution (discriminator) and deconvolution (generator); using batch normalization in both the generator and the discriminator; removing all hidden layers in the fully connected layer; using the ReLU activation function in all layers except the output layer in the generator, and the Tanh activation function in the output layer; using the LeakyReLU activation function in all layers except the output layer in the discriminator to solve the gradient vanishing problem of ReLU, and using the Sigmoid activation function in the output layer. Among them, the generator is responsible for converting the input random vector or noise into a realistic image. This step involves converting the input feature vector into a feature map through a series of convolution and pooling operations, and then converting this feature map into the final generated image; but since the convolution and pooling operations will cause the size of the feature map to be reduced, it is necessary to use a deconvolution operation to restore the extracted feature map to the same size as the original image, so that the generator can maintain the same spatial resolution as the original image when generating the image, and produce a result that is more similar to the reference true value.

[0045] According to the above guidelines, the generated DCGAN model structure is as follows Figure 2 As shown, the top number indicates the number of channels of the convolution, and the bottom number indicates the dimension of the output matrix, as follows: For the generator, the input layer projects and reshapes the random vector z(1×100) into a 4×4×512 high-dimensional feature matrix, which is then batch normalized to improve the stability of training and the ReLU activation function to introduce nonlinearity; the high-dimensional features are then converted into details of the TEC image through three hidden layers with a similar structure, including a DeConv layer, a Norm layer, and a ReLU layer; the output layer consists of another DeConv layer and a Tanh layer to obtain the final TEC image. The four DeConv layers in the generator are transposed convolution layers, which transform the original 4×4×512 feature matrix into the final 64×64 pixel global TEC image. Therefore, the generator produces a 64×64 TEC image G(z) from a 1×100 random vector z.

[0046] For the discriminator, both the CODG TEC images and the TEC images generated by the generator are input into four convolutional layers to extract high-dimensional features from the TEC images; the input layer is a Conv layer and a LeakyReLU, followed by three hidden layers each consisting of a Conv layer, a Norm layer, and a LeakyReLU layer; finally, the output layer reshapes the 4×4×512 feature matrix into a 1×4,096 vector, followed by a fully connected layer and a Sigmoid activation function to generate a value between 0 and 1. Therefore, the discriminator reads a TEC image and outputs a single value D(G(z)), based on which it determines whether the TEC image is real (D(G(z))>0.5) or wrong (D(G(z))<0.5). Among them, the activation function sigmoid unit , tanh unit , ReLU unit , LeakyReLU unit The calculation formulas are as follows:

[0047] Among them, y represents the element value of the feature tensor that is output to this activation function layer after the original input image is processed by the previous layer and is used to represent the image features and information; It is usually 0.25 and can also be trained as a hyperparameter.

[0048] Step (2): Input the dataset and random vector into the constructed DCGAN model for training, and initialize the generation of the global TEC image. In the specific training process, in order to reflect the model optimization performance and parameter training quality results, we selected the FID (Fréchet Inception Distance) indicator for measurement. It is currently a more recognized indicator for measuring the quality of generated images. It not only focuses on the quality of generated images, but also reflects the diversity of images, and therefore can better reflect the state achieved by the generator during the training process. Its calculation formula is as follows:

[0049] in: and are the mean vectors of the real image and the generated image in the feature space respectively; and are the covariance matrices of the real image and the generated image, respectively.

[0050] The calculated result is the Frechet distance between the two distributions. We can use this distance to measure the difference in image distribution. The smaller the FID value, the better the quality of the generated image and the closer it is to the real image.

[0051] Step (3): Mapping is performed to map the TEC value of the generated TEC image to the area in the dataset where the real TEC image corresponds to and where data is missing. At the same time, the TEC values ​​corresponding to the pixels around the missing area need to be matched. The specific instructions are as follows: The mapping value is obtained by combining the context loss and the prior loss method. , given the random initialization z, the TEC image x to be completed, and the DCGAN output G(z). From G(z) to The mapping can be obtained by minimizing the joint loss function:

[0052] Right now

[0053] in, Represents element-by-element multiplication of matrices; is the contextual loss, used to help data completion for local improvement; is the prior loss, which is the discriminator loss of DCGAN training, used to maintain the global appearance of the completed TEC map; M is the grayscale mask image generated by x, with white at the missing data locations and black at the remaining locations; the factor Used to adjust the ratio of the two loss functions; the weight parameter W is:

[0054] Where i and j are the indices of the pixels, is the neighborhood of pixel i, and N is the number of neighborhood pixels. Pixels with observed data close to the missing data area have a larger weight value, while pixels with missing data or surrounded by observed data have a weight of zero.

[0055] 4. Image Fusion Using Poisson Algorithm The Poisson equation is constructed, and the TEC values ​​of the missing data area are obtained by the mapping value obtained by the mapping process and the TEC values ​​around the missing data area through the Poisson fusion algorithm. The generated TEC image is smoothed to obtain the final TEC image for ionosphere modeling. The specific instructions and implementation steps are as follows: Poisson fusion algorithm is a widely used technique in image processing, suitable for seamless fusion of image regions in image editing and restoration tasks. Poisson fusion algorithm is based on Poisson equation, and achieves smooth transition of image regions by solving partial differential equations, thereby visually eliminating boundaries; its key idea is to use the gradient information of the image to smooth the image, so that the area to be fused naturally matches the surrounding pixels.

[0056] Given the mapping value of formula (2) or formula (3) ,set up is the area with missing data, that is , The final TEC value f within can be obtained by minimizing as follows:

[0057] in, is the gradient operator, The f value outside keeps the original TEC value ( hour).

[0058] After applying Poisson fusion, the TEC value smoothly transitions from the data area to the gap area, and the problem of mismatch between the boundary gradient of the processed filled area and the original observed data is greatly suppressed.

[0059] The ionosphere modeling flow chart based on the DCGAN model and Poisson fusion algorithm is as follows: Figure 3 As shown, the blue arrow represents the training process. The generator (G) generates TEC images through random vectors. The discriminator (D) accepts the TEC images generated by the generator and the CODE TEC data of the training set to determine whether they are true or false TEC images. The blue dotted arrow represents an image that is judged to be a false image and returns to the generator to generate a new image. The gray arrow represents the mapping method of the joint context loss and the prior loss to obtain the mapped result. The gray dotted arrow represents the Poisson fusion post-processing algorithm to obtain the final image.

[0060] 5. Result evaluation and verification In the test set of the dataset, simulated data and real data were used to evaluate and verify the final TEC image obtained by image completion using the DCGAN model and image fusion using the Poisson algorithm.

[0061] Taking the final TEC image obtained from the CODE image set as an example, the specific steps are as follows: In the simulation data verification, the matrix mask verification method is adopted to construct three types of random masks for testing. The mask design is as follows, where each pixel represents 5.6° (longitude) × 2.8° (latitude): 10% of the data is missing, the mask is composed of a random size of 2×2 pixel matrix; 30% of the data is missing, the mask is composed of a random size of 2×2 pixel matrix; 10% of the data is missing, the mask is composed of a random size of 4×4 pixel matrix, which are called the first type, second type and third type of masks, respectively. Figure 4 (c) shows the TEC images to be completed after the test set is covered with different types of masks in different years. The test data is input into the trained DCGAN model for image completion, and then processed by Poisson fusion. The comparison effect after processing is shown in Figure 4 (d). In the obtained results: there is no area with missing data, and the problem of mismatch between the boundary gradient of the completed area and the original observed data is effectively solved. It can be seen that the completed TEC image is close to the CODE TEC image under different mask coverage conditions. In conclusion, the DCGAN model combined with the Poisson fusion algorithm has excellent image completion performance, which is conducive to high-precision global ionospheric TEC image modeling.

[0062] In the real data verification, the TEC data provided by MIT was used for verification. For the MIT data missing area, the MIT real mask was constructed to further verify the model performance. Figure 4 (e) is a schematic diagram of using MIT real data to generate a mask image. The test data was input into the trained DCGAN model for image completion and Poisson fusion post-processing. The schematic diagram of the comparison effect after processing is shown in Figure 4 (f). From the output results, it can be seen that the overall completed TEC trend and size are good, and the unfused effect is removed. In general, the DCGAN and Poisson fusion algorithms still have very good completion effects under the MIT real mask; considering that the CODE TEC data itself has certain errors in the ocean area, the global TEC image obtained by using the DCGAN and Poisson fusion algorithm is acceptable. Using this data for global ionosphere modeling is a feasible method with great development potential.

[0063] The implementation basis of each embodiment of the present invention is to implement programmed processing through a device with a processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this reality, on the basis of the above embodiments, an embodiment of the present invention provides an ionospheric TEC image modeling system based on DCGAN and Poisson fusion, which is used to execute the ionospheric TEC image modeling method based on DCGAN and Poisson fusion in the above method embodiment.

[0064] The system includes: an image acquisition module, which is used to acquire observable global ionospheric TEC data and convert it into a TEC image; an image completion module, which is used to input the converted TEC image into a trained deep convolutional generative adversarial network model for image completion; an image fusion module, which is used to Poisson-fuse the TEC values ​​of the image completion area and the TEC values ​​around the missing data area based on the constructed Poisson equation to obtain the TEC values ​​of the missing data area, and smooth the completed TEC image to obtain the final TEC image; a TEC modeling module, which is used to use the final TEC image to perform ionospheric TEC modeling. Wherein, the training of the deep convolutional generative adversarial network model and the Poisson fusion based on the training results include: constructing a TEC image data set; constructing a deep convolutional generative adversarial network model, which is used to input a random vector into a generator to obtain an initial TEC image, and input the generated initial TEC image and the TEC image of the data set into a discriminator to extract high-dimensional features and judge the authenticity, and then obtain the completed TEC image through mapping processing; using the test set and verification set in the data set for testing and verification, and outputting the trained deep convolutional generative adversarial network model.

[0065] The ionospheric TEC image modeling system based on DCGAN and Poisson fusion provided in the embodiment of the present invention aims at the current situation of repairing missing TEC data. It adopts the aforementioned several modules, combines the neural network model with the image processing technology, considers image regeneration while taking into account image fusion, so as to realize the completion of TEC images and high-precision global ionospheric TEC modeling.

[0066] It should be noted that the system embodiment provided by the present invention is used to implement the method in the above method embodiment, and is also used to implement the method in other method embodiments provided by the present invention. The only difference is that the corresponding functional modules are set, and the principle is basically the same as the principle of the above system embodiment provided by the present invention. As long as the technical personnel in the field refer to the specific technical solutions in other method embodiments on the basis of the above system embodiment, obtain the corresponding technical means and the technical solutions composed of these technical means by combining technical features, and on the premise of ensuring the practicality of the technical solution, the modules in the above system embodiment can be improved to obtain the corresponding system class embodiments, which are used to implement the methods in other method class embodiments.

[0067] Based on the same inventive concept as the above embodiment, an embodiment of the present invention further provides an ionospheric TEC image modeling device based on DCGAN and Poisson fusion, comprising a memory and a processor. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the ionospheric TEC image modeling method based on DCGAN and Poisson fusion.

[0068] Based on the same inventive concept as the above embodiment, the embodiment of the present invention further provides a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the steps of the ionospheric TEC image modeling method based on DCGAN and Poisson fusion.

[0069] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0070] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0071] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0072] In summary, the present invention focuses on image completion technology, uses image processing methods to convert observable global ionosphere data into TEC images, divides the image data into data sets, obtains a global ionosphere TEC image data set, and uses the DCGAN model and Poisson fusion algorithm to train the data set and the input random vector to obtain the final TEC image with image completion and boundary elimination, thereby achieving ionosphere modeling completion and high-precision global ionosphere TEC modeling. The present invention proposes a possible solution to the problem that traditional mathematical methods are difficult to effectively capture more local ionosphere information for global ionosphere modeling, and that it takes a long time and has low accuracy to obtain a complete TEC map due to limited ground-based GNSS coverage and data integration requirements. With the development of artificial intelligence technology, deep learning technology in the field of image restoration and completion is becoming increasingly mature, and this modeling method is expected to provide new inspiration and profound thinking for global high-precision and high-resolution ionosphere modeling.

[0073] The terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to the steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. An ionospheric TEC image modeling method based on DCGAN and Poisson fusion, characterized in that: include: Obtain observable global ionospheric TEC data and convert them into TEC images; The converted TEC image is input into the trained deep convolutional adversarial generative network model for image completion; Based on the constructed Poisson equation, the TEC values ​​of the image completion area and the TEC values ​​around the missing data area are Poisson fused to obtain the TEC values ​​of the missing data area, and the completed TEC image is smoothed to obtain the final TEC image. The final TEC image is used to perform ionospheric TEC modeling; wherein the training of the deep convolution adversarial generation network model comprises: Construct TEC image dataset; A deep convolutional generative adversarial network model is constructed to input a random vector into the generator to obtain an initial TEC image, and the generated initial TEC image and the TEC image of the dataset are input into the discriminator to extract high-dimensional features and judge the authenticity, and then the completed TEC image is obtained through mapping processing; Use the test set and validation set in the dataset for testing and verification, and output the trained deep convolutional adversarial generation network model.

2. The ionospheric TEC image modeling method based on DCGAN and Poisson fusion according to claim 1 is characterized in that: Construct a TEC image dataset, including: Get TEC image datasets of global ionospheric TEC data; Based on the solar activity index F10.7, the test set was divided; Based on the K-fold cross-validation method, the training set and the validation set are obtained.

3. The ionospheric TEC image modeling method based on DCGAN and Poisson fusion according to claim 1 is characterized in that: Build a deep convolutional adversarial network model, including: Following the architectural guidelines of Stable Deep Convolutional GANs, replace all pooling layers in the discriminator and generator, and use batch normalization in both the generator and the discriminator; Remove all hidden layers in the fully connected layer; In the generator, all layers except the output layer use the ReLU activation function, and the Tanh activation function is used in the output layer; In the discriminator, all layers except the output layer use the LeakyReLU activation function, and the Sigmoid activation function is used in the output layer.

4. The ionospheric TEC image modeling method based on DCGAN and Poisson fusion according to claim 3 is characterized in that: The random vector is input into the generator to obtain the initial TEC image, and the generated initial TEC image and the TEC image of the data set are input into the discriminator for high-dimensional feature extraction and authenticity judgment, including: Input the random vector into the generator, use the input layer of the generator to project and reshape the random vector into high-dimensional features, use three hidden layers to convert the high-dimensional features into details of the TEC image, and use the output layer to output the initial TEC image; The generated initial TEC image and the TEC image of the dataset are input into the discriminator, and the TEC image is read using the input layer and three hidden layers, and high-dimensional features are extracted from the TEC image. The high-dimensional features are reshaped into a one-dimensional vector using the output layer, and then the discriminant value is generated using the fully connected layer and the Sigmoid activation function for judging the authenticity of the TEC image.

5. The ionospheric TEC image modeling method based on DCGAN and Poisson fusion according to claim 4 is characterized in that: The completed TEC image is obtained through mapping, including: The TEC values ​​of the generated initial TEC image are mapped to the area in the dataset where the real TEC image corresponds to and where the data is missing, and the TEC values ​​corresponding to the pixels around the missing area are matched.

6. The ionospheric TEC image modeling method based on DCGAN and Poisson fusion according to claim 1, characterized in that: The mapped values ​​and the TEC values ​​around the missing data area are Poisson fused to obtain the TEC values ​​of the missing data area, including: Based on the constructed Poisson equation, the gradient of the TEC value around the missing data area is calculated, and the gradient of the TEC value of the completed image is calculated; Based on the minimization of Poisson's equation, the TEC value of the missing data area is calculated using the obtained gradient information.

7. Ionospheric TEC image modeling system based on DCGAN and Poisson fusion, characterized by: include: An image acquisition module is used to acquire observable global ionospheric TEC data and convert it into TEC images; The image completion module is used to input the converted TEC image into the trained deep convolutional adversarial generative network model for image completion; An image fusion module is used to perform Poisson fusion on the TEC values ​​of the image completion area and the TEC values ​​around the missing data area based on the constructed Poisson equation to obtain the TEC values ​​of the missing data area, and use this to smooth the completed TEC image to obtain the final TEC image; A TEC modeling module is used to perform ionospheric TEC modeling using the final TEC image; wherein the training of the deep convolution adversarial generative network model includes: Construct TEC image dataset; A deep convolutional generative adversarial network model is constructed to input a random vector into the generator to obtain an initial TEC image, and the generated initial TEC image and the TEC image of the dataset are input into the discriminator to extract high-dimensional features and judge the authenticity, and then the completed TEC image is obtained through mapping processing; Use the test set and validation set in the dataset for testing and verification, and output the trained deep convolutional adversarial generation network model.

8. An ionospheric TEC image modeling device based on DCGAN and Poisson fusion, characterized in that: It includes a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the ionospheric TEC image modeling method based on DCGAN and Poisson fusion as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, which enable the computer to execute the steps of the ionospheric TEC image modeling method based on DCGAN and Poisson fusion as described in any one of claims 1 to 6.

Images