A method and system for enhancing ionospheric frequency and height images based on time-frequency coordination

Abstract

A method and system for enhancing ionospheric frequency-height maps based on time-frequency collaboration is provided, relating to the technical field of frequency-height map processing. The method comprises: Step 1: Acquiring first and second frequency-height map sequences; Step 2: Preprocessing these sequences to generate first and second low-spectrum image features, latent codes, and mask templates; Step 3: Processing the low-spectrum image features using an improved context modeling method to obtain a time-frequency planning control signal; Step 4: Processing the first frequency-height map sequence based on the time-frequency planning control signal to obtain a low-spectrum interpolated image; Step 5: Processing the low-spectrum interpolated image or the second frequency-height map sequence based on an improved residual migration and repair strategy to obtain a high-spectrum resolution frequency-height map sequence; Step 6: Super-reconstructing the low-spectrum interpolated image or the second frequency-height map sequence based on the high-spectrum resolution frequency-height map sequence and the mask template to obtain a target frequency-height map sequence. The method and system collaboratively process time-domain and frequency-domain information, optimizing the quality of the frequency-height map without modifying the altimeter hardware.

Description

Technical Field

[0001] The present invention relates to the technical field of frequency-height map processing, and in particular to a method and system for enhancing ionospheric frequency-height map based on time-frequency collaboration. Background Art

[0002] The ionosphere is a critical region of Earth's atmosphere, located at an altitude of approximately 60 to 1000 kilometers. It contains a large number of free electrons and ions. Space weather research focuses on changes in plasma density in the ionosphere. These changes significantly affect the propagation of radio waves and, in turn, modern technological systems such as satellite communications, the Global Positioning System (GPS), and navigation.

[0003] The ionospheric vertical altimeter (ionosonde, also known as an altimeter) is one of the primary instruments for studying the ionosphere. By sweeping high-frequency radio waves into the ionosphere and recording the echoes returning from different altitudes, an ionogram (also known as an ionogram) is generated. An ionogram is a two-dimensional frequency-altitude image that shows how radio waves are reflected at different altitudes in the ionosphere at different frequencies. Because the frequency of the reflected radio waves is related to the characteristic frequency corresponding to the plasma density, the ionosonde provides important information about the structure of the ionospheric plasma density, particularly the electron density distribution of the F layer (the ionosphere is primarily divided into the D, E, and F layers from bottom to top; under solar ultraviolet radiation, the F layer can further split into the F1 and F2 layers). This information is crucial for understanding ionospheric disturbances and their impact on space weather.

[0004] By interpreting frequency-height maps, scientists can study various ionospheric phenomena, including spread F (a phenomenon caused by inhomogeneities in the F layer, typically manifested as irregular spatial and temporal fluctuations in the electron density of the ionosphere), the critical frequency of the F2 layer (foF2) (the lowest frequency at which the F2 layer can reflect electromagnetic waves during or near daytime), and plasmaspheric disturbances. Due to the complexity of the ionosphere, frequency-height maps contain a wealth of implicit information that helps understand ionospheric behavior. Using modern data analysis techniques, researchers can automatically identify and analyze key patterns in frequency-height maps, thereby better understanding ionospheric disturbances and improving space weather monitoring and forecasting.

[0005] In ionospheric sounding, an ionospheric altimeter collects reflection data by sequentially transmitting electromagnetic waves of different frequencies within a specific frequency range. The frequency step size determines the number of sampling points and the scanning time. Common ionospheric altimeters include the DPS-4D (Digisonde-Portable-Sounder-4D). Using a smaller frequency step size (e.g., 0.025 MHz) increases the number of sampling points, thereby extending the scanning time. Therefore, when the acquisition interval is short (e.g., 5 or 10 minutes), a larger frequency step size (e.g., 0.05 MHz) is necessary to complete the scan within the limited time and capture the dynamic changes of the ionosphere. However, this results in a loss of spectral detail and even makes it difficult to calculate the plasma drift velocity using the Doppler effect. Conversely, when the acquisition interval is longer (e.g., 30 minutes), the instrument has ample time for a more detailed scan. Using a smaller frequency step size improves the frequency resolution and captures more subtle structural information in the ionosphere.

[0006] Currently, the traditional frequency-height mapping method is widely used in the field of ionospheric sounding. Ionospheric altimeters sequentially transmit electromagnetic waves within a preset frequency range and scan the ionospheric reflection data using a fixed frequency step size (such as 0.025 MHz or 0.05 MHz). To improve temporal resolution, some systems use a larger frequency step size for short acquisition intervals and a smaller frequency step size for longer acquisition intervals. Furthermore, research has attempted to improve the quality of frequency-height mapping by optimizing the time and frequency domain information through traditional interpolation (such as linear interpolation or spline interpolation) and neural network-based super-resolution reconstruction techniques, respectively.

[0007] Traditional interpolation methods are primarily used to complete the temporal information of missing frames in frequency-height image sequences to improve temporal resolution. For example, linear interpolation methods generate intermediate frames by directly calculating the linear relationship between adjacent frames. This approach is computationally simple and fast. However, these methods often assume smooth image changes. When ionospheric reflection data undergoes rapid and nonlinear changes, they can easily produce blurring, ghosting, or unnatural transitions, making it difficult to accurately recover the true dynamics. Spline interpolation (such as cubic spline interpolation) uses fitting higher-order polynomial functions to create smoother transitions between intermediate frames, better preserving the continuity of image structure. However, it can still introduce unnatural fluctuations or transition distortion when dealing with drastic changes. These limitations limit the effectiveness of these methods in high-precision ionospheric sounding. Additionally, some techniques have attempted to post-process frequency-height images using traditional image interpolation and super-resolution methods, but these techniques often focus on optimizing a single domain and lack the coordinated processing of time and frequency domain information.

[0008] The main functions of the altimeter are to measure 1) the vertical profile of plasma density (electron density) and 2) the occurrence and development of density inhomogeneities and traveling ionospheric disturbances (TIDs). These two natural phenomena have the following characteristics: (1) the change in the vertical profile of plasma density is usually continuous, and even during large events such as magnetic storms, the change rate is usually still on the order of hours; (2) the duration of plasma density inhomogeneities is usually greater than 2 hours, and the movement speed is usually tens of meters per second. It is difficult for them to change significantly or move out of the observation area within 5 minutes. The period of TID is usually between 0.5 hours and 1.5 hours. Some studies have shown that the shortest period of mesoscale TID may be around 6 minutes. Therefore, the extrapolation of 15 minutes of data is sufficient to infer the changes of these phenomena. However, regardless of the time interval, the frequency-height maps actually collected by the altimeter have room for improvement in both time domain resolution and frequency domain resolution.

[0009] Taking into account the heat dissipation and functionality of the equipment, as well as the characteristics of ionospheric phenomena, most international altimeters are set to collect data at a 15-minute interval. Intensive observations are only performed for 5 minutes during seasons with a high incidence of disturbances such as plasma density inhomogeneities, to better monitor their development. Some research institutions (such as the Institute of Radio Waves and the Institute of Earth Sciences of the Chinese Academy of Sciences) are developing altimeters with "ultra-high temporal resolution" capable of collecting data at intervals approaching 1 minute. This aims to capture scientific information such as high-frequency fluctuations in the ionosphere and rapid changes in density inhomogeneities. However, as mentioned above, focusing on improving temporal resolution can result in a loss of spectral detail and the altimeter's ability to observe drift velocities.

[0010] Additionally, existing technologies for enhancing altimeter observation capabilities typically involve increasing transmit power to obtain greater image detail and shortening acquisition intervals to improve temporal resolution. These methods are hardware-level improvements, but increasing transmit power increases the equipment load and can also result in information loss during the filtering process. Shortening the acquisition interval typically uses a larger frequency step (e.g., 0.05 MHz), but this can result in a loss of spectral detail and even make it difficult to calculate plasma drift velocity using the Doppler effect. Summary of the Invention

[0011] To solve the above problems, the present invention provides an ionospheric frequency-height map enhancement method and system based on time-frequency collaboration, which collaboratively processes time domain information and frequency domain information, thereby shortening the acquisition interval of the frequency-height map and obtaining more detailed frequency-height map information. Therefore, the present invention can significantly optimize the quality of the frequency-height map without changing the hardware structure of the altimeter.

[0012] The ionospheric frequency-height image enhancement method based on time-frequency coordination includes the following steps:

[0013] Step 1: Acquire a frequency-height map sequence, and determine whether it is a first frequency-height map sequence or a second frequency-height map sequence based on a time threshold and a spectrum resolution threshold;

[0014] Step 2: Preprocess the first high-frequency image sequence using a pretrained encoder to generate a first low-frequency image feature, a first latent code, and a first mask template corresponding to the first high-frequency image sequence; or preprocess the second high-frequency image sequence to generate a second low-frequency image feature, a second latent code, and a second mask template corresponding to the second high-frequency image sequence;

[0015] Step 3: Processing the first low-frequency spectrum image features using the improved context modeling method to obtain a first time-frequency planning control signal corresponding to the first frequency-height map sequence; or processing the second low-frequency spectrum image features using the improved context modeling method to obtain a second time-frequency planning control signal corresponding to the second frequency-height map sequence;

[0016] Step 4: Under the guidance of the first time-frequency planning control signal, the first frequency-height image sequence is processed to obtain a low-frequency spectrum interpolation image;

[0017] Step 5: Under the guidance of the time-frequency planning control signal, the low-spectrum interpolation image or the second frequency-height map sequence is processed based on the improved residual migration restoration strategy to obtain a high-spectrum resolution frequency-height map sequence;

[0018] Step 6: Based on the high-spectral-resolution frequency-height map sequence and the mask template, super-resolution reconstruction is performed on the low-spectral-interpolation image or the second frequency-height map sequence to obtain a target frequency-height map sequence.

[0019] Preferably, step 2 specifically includes:

[0020] Downsampling the first frequency-height image sequence to generate a low-frequency spectrum image and a first mask template, processing the low-frequency spectrum image using a pre-trained encoder to generate a first low-frequency spectrum image feature, and compressing the first low-frequency spectrum image feature into a low-dimensional continuous latent space using the pre-trained encoder to obtain a first latent code;

[0021] Alternatively, the second high-frequency map sequence is upsampled to generate a second mask template, the second high-frequency map sequence is processed using a pre-trained encoder to generate a second low-frequency spectrum image feature, and the second low-frequency spectrum image feature is compressed into a low-dimensional continuous latent space using the pre-trained encoder to obtain a second latent code.

[0022] Preferably, step 3 specifically includes:

[0023] Context modeling is performed on the first low-spectrum image features and the second low-spectrum image features respectively to capture the inter-frame dynamic changes and dependencies of the first low-spectrum image features and the second low-spectrum image features. The first time-frequency planning control signal is composed of the inter-frame dynamic changes and dependencies corresponding to the first low-spectrum image features, and the second time-frequency planning control signal is composed of the inter-frame dynamic changes and dependencies corresponding to the second low-spectrum image features.

[0024] Preferably, step 4 specifically includes:

[0025] Under the guidance of the first time-frequency planning control signal, a noise frame is applied to the first frequency-height image sequence, and the first frequency-height image sequence with the noise frame applied is reconstructed by a stepwise denoising method to generate a low-frequency spectrum interpolation image.

[0026] Preferably, step 4 further specifically includes:

[0027] Adaptive instance normalization technology is used to integrate the diffusion step information in the reconstruction process into a conditional signal. The first frequency-height map sequence is restored based on the conditional signal so that the low-frequency interpolation image contains the detailed features of the first frequency-height map sequence. The cross-attention mechanism is used to fuse the global planning information so that each frame of the low-frequency interpolation image remains coherent in time and spectrum.

[0028] Preferably, step 5 specifically includes:

[0029] Under the guidance of the first time-frequency planning control signal, the residual between the low-frequency interpolation image and the target frequency-height map sequence is gradually predicted through a diffusion model. The predicted residual is gradually corrected through a multi-step iterative diffusion process to obtain the high-frequency details in the low-frequency interpolation image. Based on the low-frequency interpolation image and the high-frequency details, a high-frequency-height map sequence with high spectral resolution is generated.

[0030] Alternatively, under the guidance of the second time-frequency planning control signal, the residuals between the second frequency height map sequence and the target frequency height map sequence are gradually predicted through a diffusion model, and the predicted residuals are gradually corrected through a multi-step iterative diffusion process to obtain high-frequency details in the second frequency height map sequence, and a high-spectral resolution frequency height map sequence is generated based on the second frequency height map sequence and the high-frequency details.

[0031] Preferably, in step 1, the frequency-height graph sequence whose acquisition interval is not less than the time threshold and whose spectral resolution is less than the spectral resolution threshold is used as the first frequency-height graph sequence, and the frequency-height graph sequence whose acquisition interval is less than the time threshold and whose spectral resolution is not less than the spectral resolution threshold is used as the second frequency-height graph sequence.

[0032] The ionospheric frequency-height map enhancement system based on time-frequency coordination provided in the present invention implements an ionospheric frequency-height map enhancement method based on time-frequency coordination, and the system includes:

[0033] an acquisition module, configured to acquire a frequency-height map sequence and determine whether the sequence is a first frequency-height map sequence or a second frequency-height map sequence based on a time threshold and a spectrum resolution threshold;

[0034] a preprocessing module, connected to the acquisition module, configured to preprocess the first frequency-height map sequence using a pretrained encoder to generate a first low-frequency spectrum image feature, a first latent code, and a first mask template corresponding to the first frequency-height map sequence; or to preprocess the second frequency-height map sequence to generate a second low-frequency spectrum image feature, a second latent code, and a second mask template corresponding to the second frequency-height map sequence;

[0035] A time series prediction module, connected to the preprocessing module, is used to process the first low-frequency spectrum image features using an improved context modeling method to obtain a first time-frequency planning control signal corresponding to the first frequency-height map sequence; or to process the second low-frequency spectrum image features using an improved context modeling method to obtain a second time-frequency planning control signal corresponding to the second frequency-height map sequence;

[0036] a low-spectrum-resolution frequency-height map generation module, connected to the time series prediction module, for processing the first frequency-height map sequence under the guidance of the first time-frequency planning control signal to obtain a low-spectrum interpolation image;

[0037] a high-spectral-resolution frequency-height map generation module, connected to the time series prediction module and the high-spectral-resolution frequency-height map generation module, for processing the low-spectral-resolution interpolation image or the second frequency-height map sequence based on an improved residual migration and repair strategy under the guidance of the time-frequency planning control signal to obtain a high-spectral-resolution frequency-height map sequence;

[0038] The frequency-height map fusion reconstruction module is connected to the high-spectral-resolution frequency-height map generation module, and is used to perform super-resolution reconstruction on the low-spectral interpolation image or the second frequency-height map sequence based on the high-spectral-resolution frequency-height map sequence and the mask template to obtain the target frequency-height map sequence.

[0039] Preferably, the system also includes a pre-training module; the pre-training module is connected to the pre-processing module, the time series prediction module, the low spectrum resolution frequency height map generation module and the high spectrum resolution frequency height map generation module, and is used to pre-train the pre-processing module, the time series prediction module, the low spectrum resolution frequency height map generation module and the high spectrum resolution frequency height map generation module, and is used to jointly train the time series prediction module and the low spectrum resolution frequency height map generation module, and to jointly train the time series prediction module and the high spectrum resolution frequency height map generation module.

[0040] Preferably, the pre-training module pre-trains the pre-processing module, including:

[0041] The preprocessing module is set to 8 times the spatial compression rate and 16 channel potential dimensions, and the learning rate is set to 10 -4 , batch size is 32, codebook size is 512;

[0042] The reconstruction loss function of the preprocessing module is set to ,in, Indicates the first frequency-height graph sequence or the second frequency-height graph sequence, Represents the frequency-height map sequence reconstructed and output by the preprocessing module;

[0043] The codebook loss function of the preprocessing module is set to ,in, represents the potential representation generated by the preprocessing module for the first high-frequency map sequence or the second high-frequency map sequence, represents the most recent codebook vector, Indicates stopping the gradient operation;

[0044] The commitment loss function of the preprocessing module is set as ;

[0045] Total loss function of the preprocessing module Expressed as ,in, represents the balance coefficient;

[0046] When the curve of the total loss function of the preprocessing module converges, the pre-training of the preprocessing module is completed;

[0047] The pre-training module pre-trains the time series prediction module including:

[0048] The reconstruction loss function of the time series prediction module is set to ,in, Represents a binary function, representing whether it is a mask, Represents the prediction function used by the time series prediction module, represents the low-frequency image features after mask processing, Represents the true time-frequency planning information corresponding to the low-frequency image feature. The reconstruction loss is used to encourage the difference between the output of the time series prediction module and the true time-frequency planning information to be no greater than a preset threshold, thereby improving the accuracy of the time series prediction module in capturing dynamic changes and dependencies between frames.

[0049] When the curve of the reconstruction loss function of the time series prediction module converges, the pre-training of the time series prediction module is completed;

[0050] The pre-training module pre-trains the low-spectral-resolution frequency-height map generation module, including:

[0051] The diffusion loss function of the low spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, Indicates conditional signal, Represents the time step of the low-frequency part in the current diffusion process The target residual under represents the first frequency-height graph sequence with added noise, represents the low-spectral-resolution frequency-height map generation module, whose parameters are ;

[0052] When the diffusion loss curve of the low-spectral-resolution frequency-height map generation module converges, the pre-training of the low-spectral-resolution frequency-height map generation module is completed;

[0053] The pre-training module pre-trains the high-resolution spectrum image generation module, including:

[0054] The diffusion loss function of the high spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, Indicates conditional signal, represents the time step of the high-frequency part in the current diffusion process The target residual under Represents the residual between the low-frequency spectrum interpolation image with added noise and the target frequency-high image sequence or the residual between the second frequency-high image sequence and the target frequency-high image sequence, It represents the high-resolution spectrum-height map generation module, and its parameters are ;

[0055] When the diffusion loss function curve of the high-spectral-resolution frequency-height map generation module converges, the pre-training of the high-spectral-resolution frequency-height map generation module is completed;

[0056] The pre-training module jointly trains the time series prediction module and the low-spectral resolution frequency-height map generation module, including:

[0057] The mask diffusion loss function for the joint training of the time series prediction module and the low-spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, Represents the time step of the low-frequency part in the current diffusion process The target residual under represents the low-spectral-resolution frequency-height map generation module, whose parameters are , represents low spectral resolution data with added noise, Represents the time-frequency planning control signal output by the time series prediction module, , Represents the prediction function used by the time series prediction module, Represents the low-frequency image features after mask processing;

[0058] When the curve of the mask diffusion loss function converges, the joint training of the time series prediction module and the low-spectral resolution frequency-height map generation module is completed;

[0059] The pre-training module jointly trains the time series prediction module and the high-spectral resolution frequency map generation module, including:

[0060] The loss function for the joint training of the time series prediction module and the high-spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, represents the time step of the high-frequency part in the current diffusion process The target residual under Represents the residual between the low-frequency spectrum interpolation image with added noise and the target frequency-high image sequence or the residual between the second frequency-high image sequence and the target frequency-high image sequence, It represents the high-resolution spectrum-height map generation module, and its parameters are , Represents the time-frequency planning control signal output by the time series prediction module, , Represents the prediction function used by the time series prediction module, Represents the low-frequency image features after mask processing;

[0061] When the curve of the loss function of the joint training of the time series prediction module and the high spectral resolution frequency height map generation module converges, the joint training of the time series prediction module and the high spectral resolution frequency height map generation module is completed.

[0062] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0063] 1) The present invention can significantly improve the time domain continuity and frequency domain detail expression capabilities of the ionospheric frequency-height map. Specifically, by introducing a strategy that combines time series prediction with residual migration, it is possible to achieve accurate interpolation of intermediate frames of the frequency-height map and super-resolution reconstruction of the frequency-height map with low spectral resolution, effectively improving the data accuracy of the ionospheric altimeter due to long acquisition intervals and low spectral resolution during detection. By adopting a technical path of multi-stage progressive generation and incremental correction, and fully integrating inter-frame temporal information and spectral details, it is possible to achieve global coherence and local fine restoration of image reconstruction, thereby providing high-quality and reliable data support for the accurate monitoring of ionospheric structure and dynamic changes;

[0064] 2) This invention breaks through the traditional limitation of focusing on a single domain and can achieve coordinated processing of time and frequency domain information. This not only improves temporal resolution when the acquisition interval is long, but also recovers high-frequency details in the spectrum under limited sampling conditions, fundamentally resolving the contradiction between time and frequency acquisition. It adopts a multi-stage progressive generation and incremental correction strategy, effectively utilizing time series prediction and residual migration mechanisms, and can achieve global coherence in image generation and fine recovery of local details.

[0065] 3) Existing technologies usually process time domain optimization and frequency domain optimization separately when optimizing frequency-height maps. In the time domain, traditional methods mainly use linear interpolation, spline interpolation and other technologies to fill in the missing intermediate frames within the acquisition interval. However, when faced with rapidly changing ionospheric data or nonlinear dynamic scenes, such methods are prone to problems such as blurring and ghosting. In the frequency domain, a common practice is to use super-resolution reconstruction technology based on deep learning, such as convolutional neural networks or generative adversarial networks, to restore high-frequency details by learning the mapping relationship between low-spectral resolution images and high-spectral resolution images. However, these methods often rely on large-scale training data, and have deficiencies in real-time and noise suppression, making it difficult to balance overall performance. The present invention can simultaneously optimize time domain data and frequency domain data, which can not only complete the intermediate frames to improve time resolution when the acquisition interval is long, but also restore high-frequency details in the spectrum under low sampling conditions, and can fundamentally solve the contradiction between time and frequency acquisition;

[0066] 4) This invention adopts a multi-stage progressive generation and incremental correction strategy, effectively utilizing time series prediction and residual migration mechanisms, which can achieve global coherence and local fine restoration of image generation, significantly improving the blurring and ghosting phenomena that occur in traditional methods;

[0067] 5) The present invention can flexibly adapt to different acquisition conditions (e.g., acquisition intervals of 15 minutes and a spectral resolution of 0.025 MHz, and acquisition intervals of 5 minutes and a spectral resolution of 0.05 MHz), ensuring high detection accuracy while also taking into account system real-time and wide applicability, and its overall performance is significantly superior to traditional technologies;

[0068] 6) The present invention does not modify the existing altimeter hardware itself, which not only does not cause loss of existing data, but also can effectively enhance the information content of scientific data;

[0069] 7) The present invention utilizes a multi-stage progressive generation and incremental correction strategy (the present invention divides the process of time interpolation and spectrum refinement into several steps. Each step first generates a time-frequency diagram that is closer to the actual situation based on the results of the previous stage; then, through small "corrections" or "adjustments," possible deviations are gradually corrected. This not only "fills in" intermediate frames under conditions of insufficient sampling (missing some time frames), but also gradually improves the frequency domain resolution of a single image, thereby ensuring time domain coverage and obtaining finer and clearer spectral details). It can reconstruct missing intermediate frames under limited sampling conditions and refine spectral resolution. This not only improves time resolution without changing existing hardware, but also captures richer spectral details. Without changing the altimeter hardware conditions, it can significantly enhance the ability to accurately track and identify dynamic changes in the ionosphere. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only 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.

[0071] Figure 1 This is a schematic diagram of the ionospheric frequency-height image enhancement method based on time-frequency coordination in Example 1 of the present invention;

[0072] Figure 2 This is a schematic diagram of processing a first frequency-height map sequence using the ionospheric frequency-height map enhancement method based on time-frequency coordination in Example 1 of the present invention;

[0073] Figure 3 This is a schematic diagram of processing a second frequency-height map sequence using the ionospheric frequency-height map enhancement method based on time-frequency coordination in Example 1 of the present invention;

[0074] Figure 4 A schematic diagram of a network structure for processing the first frequency-height graph sequence;

[0075] Figure 5 Schematic diagram of the network structure for processing the second frequency-height graph sequence;

[0076] Figure 6 Schematic diagram of image fusion in Example 1 of the present invention;

[0077] Figure 7 Schematic diagram of the structure of the ionospheric frequency image enhancement system based on time-frequency coordination in Example 2 of the present invention;

[0078] Figure 8 This is a diagram showing the training principles of each module in the pre-training phase in Example 2 of the present invention;

[0079] Figure 9 This is a diagram of the training principles of each module in the joint training stage in Example 2 of the present invention. DETAILED DESCRIPTION

[0080] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0081] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] Example 1:

[0083] like Figure 1 Figure 2 shows the principle diagram of an ionospheric frequency-height image enhancement method based on time-frequency coordination. This method can be used to process a first frequency-height image sequence (in this embodiment, a high-spectral-resolution frequency-height image sequence with a long acquisition interval) to increase the time-domain acquisition density by generating interpolated images. Alternatively, a second frequency-height image sequence (in this embodiment, a low-spectral-resolution frequency-height image sequence with a short acquisition interval) can be processed to increase the frequency-domain acquisition density through frequency super-resolution. In both cases, the quality of the frequency-height images can be optimized to obtain the target frequency-height image sequence (in this embodiment, a high-spectral-resolution image sequence with a short acquisition interval).

[0084] It should be noted that the present invention distinguishes between the first frequency height map sequence and the second frequency height map sequence based on the time threshold and the spectrum resolution threshold. Specifically, the frequency height map sequence whose acquisition interval is not less than the time threshold and whose spectrum resolution is less than the spectrum resolution threshold is used as the first frequency height map sequence, and the frequency height map sequence whose acquisition interval is less than the time threshold and whose spectrum resolution is not less than the spectrum resolution threshold is used as the second frequency height map sequence. In the present invention, the time threshold is generally 15 minutes, and the spectrum resolution threshold is generally 0.05MHz. In actual operation, the first frequency height map sequence with an acquisition interval of 15 minutes and a spectrum resolution of 0.025MHz, or the second frequency height map sequence with an acquisition interval of 5 minutes and a spectrum resolution of 0.05MHz is generally processed. It should be noted that the acquisition interval of the first frequency height map sequence can be set to 30 minutes, and the acquisition interval of the second frequency height map sequence can be set to 10 minutes according to actual needs, which are not listed here one by one.

[0085] In addition, if Figure 1 As shown in the figure, the method used for frequency-frequency image sequences with different acquisition intervals is different. Specifically, when a high-spectral-resolution frequency-frequency image sequence with a long acquisition interval is input, it is necessary to downsample the frequency-frequency image sequence to generate a low-spectral image and a mask template. The encoder then extracts the low-spectral image features of the low-spectral image and generates a latent code through context modeling. The latent code is then decoded to generate a low-spectral interpolated image. The low-spectral interpolated image is then super-reconstructed to obtain a high-spectral-resolution frequency-frequency image sequence with a short acquisition interval.

[0086] When a low-spectral-resolution high-frequency map sequence with a short acquisition interval is input, an upsampled mask template of the low-spectral-resolution high-frequency map sequence is directly generated. Then, the low-spectral image features of the low-spectral-resolution high-frequency map sequence are extracted through an encoder, and latent coding is generated through context modeling. After that, the low-spectral-resolution high-frequency map sequence is directly super-reconstructed to obtain a high-spectral-resolution high-frequency map sequence with a short acquisition interval.

[0087] like Figure 2 and Figure 3 As shown, it is a flowchart of processing a high spectral resolution frequency height map sequence with a long acquisition interval (corresponding to the first frequency height map sequence) or a low spectral resolution frequency height map sequence with a short acquisition interval (corresponding to the second frequency height map sequence) using the ionospheric frequency height map enhancement method based on time-frequency coordination in Example 1 of the present invention.

[0088] like Figure 4 and Figure 5 As shown, the method is used to analyze the first frequency graph sequence (such as Figure 4 As shown, it is a high spectrum resolution frequency chart sequence with a long acquisition interval) or a second frequency chart sequence (in Figure 5The target frequency-height map sequence (shown in FIG) is obtained by processing a low-spectrum-resolution frequency-height map sequence with a short acquisition interval (shown in FIG). Figure 5 As shown in FIG, a schematic diagram of the network structure for collecting high spectral resolution frequency graph sequences with short acquisition intervals.

[0089] like Figure 4 As shown, there are three main steps in processing the first frequency graph sequence (time series prediction , low spectral resolution frequency map generation and high spectral resolution frequency map generation ). Specifically, in time series forecasting There are multiple RMS normalization processes in the algorithm. The output of the encoder is first subjected to RMS normalization and spatial multi-head attention layer processing, and position encoding is applied to it. Then the processing result is subjected to a second RMS normalization process and spatiotemporal multi-head attention layer processing, and a second position encoding is applied to it. Finally, the processing result is subjected to a third RMS normalization process and processed by a fully connected layer. The obtained time series prediction results are used to generate low spectral resolution frequency-height maps and high spectral resolution frequency-height maps, respectively.

[0090] During the generation of low-spectral-resolution frequency-pitch maps, noise is applied to the low-spectral-resolution frequency-pitch map sequence. Specifically, the noise is processed sequentially through three processing layers (including a first layer consisting of a spatial multi-head attention layer and an RMS normalization layer, a second layer consisting of a temporal multi-head attention layer and an RMS normalization layer, and a third layer consisting of a cross multi-head attention layer and an RMS normalization layer). This process includes position encoding and the application of diffusion time steps. The noise is then fed into a decoder along with the output from the time series prediction process to produce a low-spectral-resolution frequency-pitch map sequence, which includes multiple low-spectral-resolution frequency-pitch maps.

[0091] In the generation of high-spectral-resolution frequency-height maps, the low-spectral-resolution frequency-height map obtained in the previous step will be processed in sequence through three processing layers (including the first layer composed of a spatial multi-head attention layer and a root mean square normalization layer, the second layer composed of a temporal multi-head attention layer and a root mean square normalization layer, and the third layer composed of a cross-multi-head attention layer and a root mean square normalization layer), which includes the application of position encoding and diffusion time steps. Finally, the cross-multi-head attention layer of the third layer outputs multiple high-spectral-resolution frequency-height maps with short acquisition intervals. These high-spectral-resolution frequency-height maps with short acquisition intervals constitute a high-spectral-resolution frequency-height map sequence with short acquisition intervals, that is, the target frequency-height map sequence is obtained.

[0092] However, when processing the second frequency map sequence, it is not necessary to Figure 4 The low-resolution frequency-height map generation process shown in the figure is as follows: Figure 5 shown.

[0093] like Figure 5 As shown in the figure, for the low spectral resolution frequency sequence with short acquisition interval, there are only two main steps (time series prediction and high spectral resolution frequency map generation ).like Figure 5 The network structure required for time series prediction and the network structure required for generating high spectral resolution frequency maps are shown in Figure 4 The corresponding structures shown are similar and are not correct here. Figure 5 The two structures are described in detail.

[0094] Specifically, if Figure 5 As shown, an encoder is used to encode multiple low-spectral resolution frequency maps in a low-spectral resolution frequency map sequence. These maps are then processed using the three layers of time series prediction to extract contextual information from the low-spectral resolution frequency maps and obtain a time-frequency planning control signal. Subsequently, the original low-spectral resolution frequency map is processed under the guidance of this time-frequency planning control signal. Finally, a mask template is used to fuse the processed high-frequency details onto the original low-spectral resolution frequency map, thereby obtaining a high-spectral resolution frequency map with a short acquisition interval. Multiple high-spectral resolution frequency maps with a short acquisition interval constitute a high-spectral resolution frequency map sequence with a short acquisition interval.

[0095] In the above, the present invention uses a Vector Quantized Variational Autoencoder (VQ-VAE) as a codec to compress the original frequency-frequency map into a low-dimensional continuous latent space, thereby improving training and inference efficiency. Specifically, VQ-VAE consists of an encoder and a decoder. The encoder uses an 8x spatial compression ratio and a latent dimension of 16 channels to map video frames into an N×16 tile structure (N represents the number of tokens after segmentation), effectively reducing data dimensionality while preserving spatial detail as much as possible. The decoder uses this low-dimensional representation to reconstruct high-quality video frames, providing auxiliary conditions and prior information for the subsequent frequency-frequency map generation step.

[0096] In addition, the three-layer network architecture mentioned above for achieving time series prediction and generating low-spectral resolution high-frequency graph sequences and high-spectral resolution high-frequency graph sequences, as well as the two-layer network architecture required for achieving time series prediction and generating high-spectral resolution high-frequency graph sequences are both based on the Transformer design, which focuses on the balance between efficiency and resource utilization, and adopts an efficient attention mechanism, optimized layer normalization, and carefully designed positional encoding.

[0097] To enhance training stability, the input of each attention block uses Root Mean Square Layer Normalization (RMS-Norm), and a spatiotemporal attention mechanism is introduced to effectively extract contextual information and focus on the associated features of previous and next frames in the frequency-height map sequence to be predicted, thereby better capturing temporal changes. Furthermore, each attention block uses Rotary Position Embedding (RoPE) to encode the spatial and temporal position information of tokens to capture the internal spatial structure of each token in the frequency-height map sequence to be predicted, as well as the sequential relationship between frames. Specifically, a two-dimensional RoPE is used to encode the image of a single frequency-height map in the frequency-height map sequence to be predicted, preserving the two-dimensional spatial structure within the frame; the image patch of each frame is flattened into a one-dimensional token sequence, and a learnable marker token is inserted between different frequency-height maps to clearly distinguish each independent frame. This design not only retains the spatial position information within each frame, but also clarifies the sequential relationship between frames through tokens, enabling the model to effectively process frame sequence data with different aspect ratios and resolutions.

[0098] The aforementioned methods for generating low- and high-resolution frequency maps not only utilize a Transformer-based network architecture but also draw on the design principles of the Diffusion Transformer (DIT) model. The main difference between the Transformer and the DIT is that the low-resolution frequency map generation employs a codec to compress the frequency maps to better adapt to the characteristics of low-frequency data, while the high-resolution frequency map generation does not employ this compression strategy, aiming to better preserve high-frequency details. Furthermore, the Adaptive Instance Normalization (AdaIN) technique is employed in both low- and high-resolution frequency map generation. This technique integrates information from the diffusion step into a conditional signal and injects it into the spatial attention layer. This allows the generation of these frequency maps to be flexibly controlled based on the diffusion process, thereby more precisely recovering detailed features in the frequency maps. Furthermore, a cross-attention mechanism is incorporated to fully utilize the time-frequency planning control signal generated during time series prediction. This mechanism ensures the effective integration of global planning information during the generation process, achieving temporal and spectral coherence between high-frequency image frames. This design, while taking into account the characteristics of both low-frequency and high-frequency data, enables precise restoration and continuous generation of inter-frame time-frequency relationships under the guidance of global planning information.

[0099] Furthermore, to achieve a smooth transition and detail completion from low-resolution to high-resolution frequency maps, the present invention employs a residual shift (ResShift) correction method during the high-resolution frequency map generation process. Specifically, the low-resolution frequency map is used as the starting image and the high-resolution frequency map as the target image. A diffusion model is then used to gradually predict and correct the residuals between low and high frequencies. Specifically, the low-resolution frequency map is first processed during the high-resolution frequency map generation process. In a multi-step iterative diffusion process, detailed residuals are predicted based on the current diffusion progress, and high-frequency information is gradually integrated, so that the generated image gradually approximates the detailed features of the target high-resolution frequency map. During this process, the model structure remains unchanged; only the loss function and diffusion step design are adjusted to ensure a smooth transition and effective restoration of low- and high-frequency information. This approach leverages the stability of the low-resolution frequency map while restoring high-frequency details through residual correction, effectively improving the overall quality and resolution of the generated high-resolution frequency map while maintaining both efficiency and real-time performance.

[0100] like Figure 6 The figure shows the workflow for fusing the generated low-resolution and high-resolution frequency maps. Each number in the figure corresponds to a specific frequency. Number 1 represents 5 MHz, number 2 represents 5.0025 MHz, number 3 represents 5.005 MHz, number 4 represents 5.0075 MHz, number 5 represents 5.01 MHz, number 6 represents 5.0125 MHz, number 7 represents 5.015 MHz, and number 8 represents 5.0175 MHz.

[0101] In the fusion step of the fusion image, the high-spectral-resolution frequency-height image generated in the process of generating the high-spectral-resolution frequency-height image is split into multiple parts according to frequency, and then the corresponding areas (shaded areas) in the image are replaced by the content of the low-spectral-resolution frequency-height image and the preset mask template, thereby realizing the organic fusion of high-frequency details and original low-frequency information, and significantly improving the overall quality and detail performance of the frequency-height image.

[0102] Example 2:

[0103] like Figure 7 Figure 1 shows a schematic diagram of the structure of an ionospheric frequency-height image enhancement system based on time-frequency collaboration, provided by the present invention. The system includes an acquisition module, a preprocessing module, a time series prediction module, a low-frequency spectrum frequency-height image generation module, a high-frequency spectrum frequency-height image generation module, a frequency-height image fusion and reconstruction module, and a pretraining module. This system can implement the ionospheric frequency-height image enhancement method based on time-frequency collaboration described in Example 1. The aforementioned system can effectively solve the problem of interpolating frames in high-frequency spectrum image sequences with long acquisition intervals.

[0104] In practice, the system first uses the acquisition module to acquire a first frequency-height image sequence (in this embodiment, a high-spectral-resolution image sequence with an acquisition interval of 15 minutes and a spectral resolution of 0.025 MHz) with an acquisition interval of no less than a time threshold and a spectral resolution of no less than a spectral resolution threshold. Alternatively, a second frequency-height image sequence (in this embodiment, a low-spectral-resolution image sequence with an acquisition interval of 5 minutes and a spectral resolution of 0.05 MHz) with an acquisition interval of no less than a time threshold and a spectral resolution of no less than a spectral resolution threshold can be acquired. The preprocessing module then preprocesses these frequency-height image sequences to obtain low-spectral image features and mask templates. The time series prediction module then processes the low-spectral image features to generate latent codes and a time-frequency planning control signal. Under the guidance of the time-frequency planning control signal, the low-spectral-resolution frequency-height image generation module generates low-spectral interpolated images of the first frequency-height image sequence. Under the guidance of the time-frequency planning control signal, the high-spectral-resolution frequency-height image generation module processes the low-spectral interpolated images or the second frequency-height image sequence using a residual migration restoration strategy, restoring high-frequency details in the low-spectral interpolated images or the second frequency-height image sequence. The frequency-height map fusion reconstruction module performs super-resolution reconstruction on the low-spectrum interpolation image or the second frequency-height map sequence based on the high-spectral resolution frequency-height map and the mask template to obtain the target frequency-height map sequence.

[0105] In order to improve the processing accuracy of each module in the present invention, the system also trains the preprocessing module, time series prediction module, low spectrum resolution frequency height map generation module and high spectrum resolution frequency height map generation module through a pre-training module.

[0106] Specifically, the present invention adopts a two-stage progressive training strategy to improve training stability and final generation performance by gradually increasing the difficulty of the task. The main advantage of the progressive training strategy is that it helps to alleviate the training instability problem of the Transformer model under large-scale parameter settings. Figure 8 and Figure 9 As shown in FIG, it is a training principle diagram for pre-training and joint training of each module in this embodiment.

[0107] Training consists of an initial training phase and a joint training phase. The initial training phase involves training the preprocessing module, the time series prediction module, the low-spectral-resolution frequency-height map generation module, the high-spectral-resolution frequency-height map generation module, and the encoder's vector quantized variational autoencoder (VQ-VAE). The VQ-VAE encodes the input frequency-height map data into a discrete latent space, providing auxiliary conditions and prior information for the multiple frequency-height map generation modules mentioned above. The joint training phase involves joint training of the time series prediction module and the low-spectral-resolution frequency-height map generation module, as well as joint training of the time series prediction module and the high-spectral-resolution frequency-height map generation module.

[0108] In the initial training stage, the present invention pre-trains each submodule independently so that the parameters of each module reach a reasonable initialization state.

[0109] The training process for the preprocessing module aims to minimize the reconstruction error and quantization error. The loss function consists of three parts: reconstruction loss function, codebook loss function and commitment loss function.

[0110] The reconstruction loss function is , represents the original input frequency-height map (in this embodiment, represents the first frequency-height map sequence or the second frequency-height map sequence), Represents the frequency-height map sequence reconstructed by the preprocessing module, and the original reconstruction loss is calculated using the mean square error (MSE) to measure the frequency-height map of the original input. and the reconstructed output from the encoder and decoder in the preprocessing module The difference between.

[0111] The codebook loss function is , Indicates that the preprocessing module is for input The generated continuous latent representation, represents the most recent codebook vector, Indicates stopping the gradient operation and encouraging the feature vector output by the encoder to be close to the selected quantized embedding vector.

[0112] The commitment loss function is , to prevent the feature vector output by the encoder from deviating too far from the codebook.

[0113] The total loss function is expressed as follows, where is the balance coefficient:

[0114] .

[0115] When the curve of the total loss function of the preprocessing module converges, the pre-training of the preprocessing module is completed.

[0116] The task of the time series prediction module is to independently train a network so that it can accurately capture the time-frequency planning information between frames in the input frequency-frequency map sequence, thereby reflecting the temporal changes of each frame in the video. The time series prediction module uses a prediction function For the masked input (in this example, the low-frequency image features) Make a prediction and generate real planning information (in this embodiment, the real time-frequency planning information corresponding to the low-frequency image features) To this end, the time series prediction module uses reconstruction loss (i.e., MSE loss) to measure the difference between the predicted result and the true value. The reconstruction loss function is:

[0117] ;

[0118] M is a binary function representing whether it is a mask. This reconstruction loss function encourages the time-frequency planning information output by the time series prediction module to be as close as possible to the actual time-frequency planning information, thereby improving the accuracy of the time series prediction module in capturing dynamic changes between frames and temporal structure, and providing an effective conditional signal for the subsequent generation of frequency-height maps.

[0119] When the curve of the reconstruction loss function of the time series prediction module converges, the pre-training of the time series prediction module is completed.

[0120] The main task of the low-resolution frequency-pitch map generation module is to process the low-spectral data using a network structure with codec compression and optimize it through mask diffusion loss to generate high-quality low-resolution frequency-pitch maps. Specifically, the low-resolution frequency-pitch map generation module repairs and refines the input low-spectral data. The input to the low-resolution frequency-pitch map generation module includes the noisy low-spectral data (in this embodiment, the first noisy frequency-pitch map sequence). , the time step in the current diffusion process , condition signal , the conditional signal is usually the time-frequency planning control signal output by the time series prediction module (here, in order to maintain the consistency of the architecture during subsequent joint training, this parameter is retained ). Specifically, the low-spectral resolution frequency-height map generation module compresses the low-spectral data through an encoder, allowing it to be processed efficiently in a lower-dimensional space while retaining the necessary spatial detail information. The generative network gradually recovers and optimizes the detailed features of the low-spectral image during the diffusion process. Its training objective is achieved through mask diffusion loss, and the diffusion loss function is:

[0121] ;

[0122] M It is a binary mask used to mask areas that are known or do not need to be repaired; Represents the time step of the low-frequency part in the current diffusion process The target residual under ; represents the low-spectral-resolution frequency-height map generation module, whose parameters are .

[0123] When the curve of the diffusion loss function of the low spectral resolution frequency height map generation module converges, the pre-training of the low spectral resolution frequency height map generation module is completed.

[0124] The main task of the high spectral resolution frequency map generation module is to restore high-frequency details. The core of the module is to use the ResShift repair strategy to perform noise modeling and correction on the difference between the low spectral resolution frequency map and the high spectral resolution frequency map, thereby completing the high-frequency details. Specifically, the high spectral resolution frequency map generation module first calculates the residual between the low spectral resolution frequency map and the high spectral resolution frequency map, then applies noise perturbation to the residual, and predicts and corrects the residual through the ResShift model within a relatively small number of iterations (20 steps). In the high spectral resolution frequency map generation module, the input includes the residual information with added noise (in this embodiment, the residual between the low spectral interpolation image with added noise and the target frequency map sequence or the residual between the second frequency map sequence and the target frequency map sequence) , the time step in the current diffusion process and conditional signals (For example, the time-frequency planning control signal). The output of this module is the residual correction result generated by the ResShift mechanism. The difference between the residual correction result and the target residual is measured by the following diffusion loss function:

[0125] ;

[0126] M It is a binary mask used to mask areas that are known or do not need to be repaired; Indicates that the high frequency spectrum part is in the current diffusion step The target residual under ; It represents the high-resolution frequency-height map generation module using the ResShift repair strategy, and its parameters are In each iteration, the predicted residual is superimposed with the low-spectral-resolution frequency-height map, thereby gradually approaching the detailed features of the target frequency-height map.

[0127] When the curve of the diffusion loss function of the high spectral resolution frequency height map generation module converges, the pre-training of the high spectral resolution frequency height map generation module is completed.

[0128] During the joint training phase, the present invention utilizes the pre-trained time series prediction module to jointly train both the low- and high-resolution frequency-height map generation modules, leveraging the time-frequency planning information corresponding to the time-frequency planning control signal. Each subtask combines conditional information fusion with a non-classified guidance strategy and a cross-attention mechanism, thereby improving the overall quality and detail of the generated frequency-height map sequences.

[0129] In the joint training of the time series prediction module and the low-spectral resolution frequency-height map generation module, the time-frequency planning control signal output by the time series prediction module is used as a conditional signal to guide the low-spectral resolution frequency-height map generation module to perform detail recovery and interpolation in the mask area, ensuring that the generated low-spectral resolution frequency-height map is temporally coherent.

[0130] The input conditions include the time-frequency planning control signal generated by the time series prediction module , low spectral resolution data with noise perturbation and the time step in the current diffusion process , Represents the prediction function of the time series prediction module. Represents the masked low-frequency image features.

[0131] The goal of the joint training of the time series prediction module and the low-spectral resolution frequency-height map generation module is to use a unified mask diffusion loss function to perform end-to-end training on the low-spectral resolution frequency-height map generation module. The mask diffusion loss function is:

[0132] ;

[0133] is a binary mask, represents the time step of the diffusion process of the low-frequency spectrum The target residual (or noise) under .

[0134] When the curve of the mask diffusion loss function converges, the joint training of the time series prediction module and the low-spectral resolution frequency-height map generation module is completed.

[0135] In the joint training of the time series prediction module and the high-spectral resolution frequency-pitch map generation module, the time-frequency planning signal output by the time series prediction module is used as a conditional signal to guide the high-spectral resolution frequency-pitch map generation module (using the ResShift repair strategy) to perform noise modeling and correction on the difference between the low-spectral resolution frequency-pitch map and the high-spectral resolution frequency-pitch map, thereby achieving the restoration and enhancement of high-frequency details within a relatively small number of iterations (20 steps). The input conditions include the time-frequency planning control signal generated by the time series prediction module , Represents the prediction function of the time series prediction module; the residual between the low-frequency interpolation image obtained after noise perturbation and the target frequency-high image sequence or the residual between the second frequency-high image sequence and the target frequency-high image sequence (or directly use the high-frequency spectrum data to perform difference calculation); and the time step of the current diffusion process . Represents the masked low-frequency image features.

[0136] The high-resolution frequency-frequency map generation module is trained end-to-end using a unified mask diffusion loss. The core of this training is to use the ResShift repair strategy to model and correct the noise in the difference between the low-resolution frequency-frequency map and the high-resolution frequency-frequency map. The loss function is:

[0137] ;

[0138] It is a binary mask used to mask areas that are known or do not need to be repaired; represents the time step of the high-frequency part in the current diffusion process The target residual under , that is, the ideal difference between the low spectral resolution frequency height map and the high spectral resolution frequency height map; It represents the high-resolution frequency-height map generation module using the ResShift repair strategy, and its parameters are The high-resolution frequency-height map generation module predicts and corrects the residual between low and high frequencies within a relatively small number of iterations (20 steps), and gradually corrects the low-resolution frequency-height map to an image containing more high-frequency details. To further enhance the robustness of the high-resolution frequency-height map generation module and support Classifier-Free Guidance, a certain probability can be set during the above training process to randomly Replace with unconditional signal .

[0139] When the curve of the loss function of the joint training of the time series prediction module and the high spectral resolution frequency height map generation module converges, the joint training of the time series prediction module and the high spectral resolution frequency height map generation module is completed.

[0140] In a preferred embodiment of the present invention, the dataset is obtained from the Hainan National Space Weather Science Observation and Research Station (coordinates: 19.5°N, 109.1°E, magnetic 11°N), which is part of the China Meridian Project. The dataset includes 469,213 ionospheric frequency and height maps generated between 2010 and 2023. Of these, 387,668 were acquired with a 15-minute interval and a spectral resolution of 0.025 MHz, and 44,358 were acquired with a 5-minute interval and a spectral resolution of 0.05 MHz. These maps were generated by the DPS-4D Digisonde ionospheric altimeter.

[0141] In addition, when training VQ-VAE, the learning rate is set to 10 -4, the batch size is set to 32, the codebook size is fixed to 512, the reconstruction loss uses MSE, and the balance coefficient Set to 0.25 and the number of training rounds is fixed to 20.

[0142] In the training of the time series prediction module, the time series prediction module adopts the Transformer structure, through the prediction function For masked input Make predictions and generate time-frequency planning control signals so that the system can capture dynamic changes and long-term dependencies between frames. Set the learning rate to 10 -4 , the batch size is set to 16 and the training rounds are set to 30.

[0143] In the training of the low-resolution frequency-frequency image generation module, a generative network with codec compression is used to generate the noise-perturbed low-frequency spectrum data. Perform gradual denoising and reconstruction to restore the basic structure and temporal coherence of the image. Set the learning rate to , the number of diffusion steps is set to 50, the mask ratio is set to 40%, the batch size is set to 16, and the training rounds are set to 50.

[0144] In the training of the high-resolution spectrum-frequency map generation module, the residual between the low-resolution spectrum-frequency map and the target high-resolution spectrum-frequency map is calculated, and noise perturbation is applied to the residual, which is then input into the ResShift module. In this embodiment, the learning rate is set to , the number of diffusion steps is set to 20, the mask ratio is set to 40%, the batch size is set to 16, and the training rounds are set to 50.

[0145] In the joint training of the time series prediction module and the low-spectral resolution frequency map generation module, the learning rate is set to , the number of diffusion steps is set to 50, the mask ratio is set to 40%, the batch size is set to 16, and the training rounds are set to 50.

[0146] In the joint training of the time series prediction module and the high spectral resolution frequency map generation module, the learning rate is set to , the number of diffusion steps is set to 20, the mask ratio is set to 40%, the batch size is set to 16, and the training rounds are set to 50.

[0147] The above specific settings are only preferred embodiments of the ionospheric frequency height map enhancement system based on time-frequency coordination provided in the present invention. The present invention is not limited to the above parameter settings, and the above settings can be appropriately adjusted according to specific circumstances. These are all within the scope of protection of the present invention.

[0148] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above examples is only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A method for enhancing ionospheric frequency-height images based on time-frequency coordination, characterized in that: The following steps are involved: Step 1: Acquire a frequency-height map sequence, and determine whether it is a first frequency-height map sequence or a second frequency-height map sequence based on a time threshold and a spectrum resolution threshold; In step 1, the frequency-height map sequence whose acquisition interval is not less than the time threshold and whose spectral resolution is less than the spectral resolution threshold is used as the first frequency-height map sequence, and the frequency-height map sequence whose acquisition interval is less than the time threshold and whose spectral resolution is not less than the spectral resolution threshold is used as the second frequency-height map sequence; Step 2: Using a pre-trained encoder, pre-process the first frequency-height map sequence to generate a first low-frequency spectrum image feature, a first latent code, and a first mask template corresponding to the first frequency-height map sequence; or pre-process the second frequency-height map sequence to generate a second low-frequency spectrum image feature, a second latent code, and a second mask template corresponding to the second frequency-height map sequence; Step 3: Processing the first low-frequency spectrum image features using the improved context modeling method to obtain a first time-frequency planning control signal corresponding to the first frequency-height map sequence; or processing the second low-frequency spectrum image features using the improved context modeling method to obtain a second time-frequency planning control signal corresponding to the second frequency-height map sequence; Step 4: Under the guidance of the first time-frequency planning control signal, the first frequency-height image sequence is processed to obtain a low-frequency spectrum interpolation image; Step 4 specifically includes: Under the guidance of the first time-frequency planning control signal, a noise frame is applied to the first frequency-height image sequence, and the first frequency-height image sequence with the noise frame applied is reconstructed by a stepwise denoising method to generate a low-frequency spectrum interpolation image; Step 5: Under the guidance of the time-frequency planning control signal, the low-spectrum interpolation image or the second frequency-height map sequence is processed based on the improved residual migration restoration strategy to obtain a high-spectrum resolution frequency-height map sequence; Step 6: Based on the high-spectral-resolution frequency-height map sequence and the mask template, super-resolution reconstruction is performed on the low-spectral-interpolation image or the second frequency-height map sequence to obtain a target frequency-height map sequence.

2. The ionospheric frequency map enhancement method based on time-frequency coordination according to claim 1, characterized in that: Step 2 specifically includes: Downsampling the first frequency-height image sequence to generate a low-frequency spectrum image and a first mask template, processing the low-frequency spectrum image using a pre-trained encoder to generate a first low-frequency spectrum image feature, and compressing the first low-frequency spectrum image feature into a low-dimensional continuous latent space using the pre-trained encoder to obtain a first latent code; Alternatively, the second high-frequency map sequence is upsampled to generate a second mask template, the second high-frequency map sequence is processed using a pre-trained encoder to generate a second low-frequency spectrum image feature, and the second low-frequency spectrum image feature is compressed into a low-dimensional continuous latent space using the pre-trained encoder to obtain a second latent code.

3. The ionospheric frequency map enhancement method based on time-frequency coordination according to claim 1, characterized in that: Step 3 specifically includes: Context modeling is performed on the first low-spectrum image features and the second low-spectrum image features respectively to capture the inter-frame dynamic changes and dependencies of the first low-spectrum image features and the second low-spectrum image features. The first time-frequency planning control signal is composed of the inter-frame dynamic changes and dependencies corresponding to the first low-spectrum image features, and the second time-frequency planning control signal is composed of the inter-frame dynamic changes and dependencies corresponding to the second low-spectrum image features.

4. The ionospheric frequency map enhancement method based on time-frequency coordination according to claim 1, characterized in that: Step 4 also specifically includes: Adaptive instance normalization technology is used to integrate the diffusion step information in the reconstruction process into a conditional signal. The first frequency-height map sequence is restored based on the conditional signal so that the low-frequency interpolation image contains the detailed features of the first frequency-height map sequence. The cross-attention mechanism is used to fuse the global planning information so that each frame of the low-frequency interpolation image remains coherent in time and spectrum.

5. The ionospheric frequency map enhancement method based on time-frequency coordination according to claim 1, characterized in that: Step 5 specifically includes: Under the guidance of the first time-frequency planning control signal, the residual between the low-frequency interpolation image and the target frequency-height map sequence is gradually predicted through a diffusion model. The predicted residual is gradually corrected through a multi-step iterative diffusion process to obtain the high-frequency details in the low-frequency interpolation image. Based on the low-frequency interpolation image and the high-frequency details, a high-frequency-height map sequence with high spectral resolution is generated. Alternatively, under the guidance of the second time-frequency planning control signal, the residuals between the second frequency height map sequence and the target frequency height map sequence are gradually predicted through a diffusion model, and the predicted residuals are gradually corrected through a multi-step iterative diffusion process to obtain high-frequency details in the second frequency height map sequence, and a high-spectral resolution frequency height map sequence is generated based on the second frequency height map sequence and the high-frequency details.

6. An ionospheric frequency-height map enhancement system based on time-frequency coordination, which implements the ionospheric frequency-height map enhancement method based on time-frequency coordination described in claim 1, characterized in that: The system includes: an acquisition module, configured to acquire a frequency-height map sequence and determine whether the sequence is a first frequency-height map sequence or a second frequency-height map sequence based on a time threshold and a spectrum resolution threshold; a preprocessing module, connected to the acquisition module, configured to preprocess the first frequency-height map sequence using a pretrained encoder to generate a first low-frequency spectrum image feature, a first latent code, and a first mask template corresponding to the first frequency-height map sequence; or to preprocess the second frequency-height map sequence to generate a second low-frequency spectrum image feature, a second latent code, and a second mask template corresponding to the second frequency-height map sequence; A time series prediction module, connected to the preprocessing module, is used to process the first low-frequency spectrum image features using an improved context modeling method to obtain a first time-frequency planning control signal corresponding to the first frequency-height map sequence; or to process the second low-frequency spectrum image features using an improved context modeling method to obtain a second time-frequency planning control signal corresponding to the second frequency-height map sequence; a low-spectrum-resolution frequency-height map generation module, connected to the time series prediction module, for processing the first frequency-height map sequence under the guidance of the first time-frequency planning control signal to obtain a low-spectrum interpolation image; a high-spectral-resolution frequency-height map generation module, connected to the time series prediction module and the high-spectral-resolution frequency-height map generation module, for processing the low-spectral-resolution interpolation image or the second frequency-height map sequence based on an improved residual migration and repair strategy under the guidance of the time-frequency planning control signal to obtain a high-spectral-resolution frequency-height map sequence; The frequency-height map fusion reconstruction module is connected to the high-spectral-resolution frequency-height map generation module, and is used to perform super-resolution reconstruction on the low-spectral interpolation image or the second frequency-height map sequence based on the high-spectral-resolution frequency-height map sequence and the mask template to obtain the target frequency-height map sequence.

7. The ionospheric frequency image enhancement system based on time-frequency coordination according to claim 6, characterized in that: The system also includes a pre-training module; The pre-training module is connected to the pre-processing module, the time series prediction module, the low spectrum resolution frequency height map generation module and the high spectrum resolution frequency height map generation module, and is used to pre-train the pre-processing module, the time series prediction module, the low spectrum resolution frequency height map generation module and the high spectrum resolution frequency height map generation module, and to jointly train the time series prediction module and the low spectrum resolution frequency height map generation module, and to jointly train the time series prediction module and the high spectrum resolution frequency height map generation module.

8. The ionospheric frequency image enhancement system based on time-frequency coordination according to claim 7, characterized in that: The pre-training module pre-trains the pre-processing module including: The preprocessing module is set to 8 times the spatial compression rate and 16 channel potential dimensions, and the learning rate is set to 10 -4 , batch size is 32, codebook size is 512; The reconstruction loss function of the preprocessing module is set to ,in, Indicates the first frequency-height graph sequence or the second frequency-height graph sequence, Represents the frequency-height map sequence reconstructed and output by the preprocessing module; The codebook loss function of the preprocessing module is set to ,in, represents the potential representation generated by the preprocessing module for the first high-frequency map sequence or the second high-frequency map sequence, represents the most recent codebook vector, Indicates stopping the gradient operation; The commitment loss function of the preprocessing module is set as ; Total loss function of the preprocessing module Expressed as ,in, represents the balance coefficient; When the curve of the total loss function of the preprocessing module converges, the pre-training of the preprocessing module is completed; The pre-training module pre-trains the time series prediction module including: The reconstruction loss function of the time series prediction module is set to ,in, Represents a binary function, representing whether it is a mask, Represents the prediction function used by the time series prediction module, represents the low-frequency image features after mask processing, Represents the true time-frequency planning information corresponding to the low-frequency image features. The reconstruction loss is used to encourage the difference between the output of the time series prediction module and the true time-frequency planning information to be no greater than a preset threshold, thereby improving the accuracy of the time series prediction module in capturing dynamic changes and dependencies between frames. When the curve of the reconstruction loss function of the time series prediction module converges, the pre-training of the time series prediction module is completed; The pre-training module pre-trains the low-spectral-resolution frequency-height map generation module, including: The diffusion loss function of the low spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, Indicates conditional signal, Represents the time step of the low-frequency part in the current diffusion process The target residual under represents the first frequency-height graph sequence with added noise, represents the low-spectral-resolution frequency-height map generation module, whose parameters are ; When the diffusion loss function curve of the low-spectral-resolution-frequency-height map generation module converges, the pre-training of the low-spectral-resolution-frequency-height map generation module is completed; The pre-training module pre-trains the high-resolution spectrum image generation module, including: The diffusion loss function of the high spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, Indicates conditional signal, represents the time step of the high-frequency part in the current diffusion process The target residual under Represents the residual between the low-frequency spectrum interpolation image with added noise and the target frequency-high image sequence or the residual between the second frequency-high image sequence and the target frequency-high image sequence, It represents the high-resolution spectrum-height map generation module, and its parameters are ; When the diffusion loss function curve of the high-spectral-resolution frequency-height map generation module converges, the pre-training of the high-spectral-resolution frequency-height map generation module is completed; The pre-training module jointly trains the time series prediction module and the low-spectral resolution frequency-height map generation module, including: The mask diffusion loss function for the joint training of the time series prediction module and the low-spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, Represents the time step of the low-frequency part in the current diffusion process The target residual under represents the low-spectral-resolution frequency-height map generation module, whose parameters are , represents low spectral resolution data with added noise, Represents the time-frequency planning control signal output by the time series prediction module, , Represents the prediction function used by the time series prediction module, Represents the low-frequency image features after mask processing; When the curve of the mask diffusion loss function converges, the joint training of the time series prediction module and the low-spectral resolution frequency-height map generation module is completed; The pre-training module jointly trains the time series prediction module and the high-spectral resolution frequency map generation module, including: The loss function for the joint training of the time series prediction module and the high-spectral resolution frequency map generation module is set to ,in, represents a binary mask, represents the time step in the current diffusion process, represents the time step of the high-frequency part in the current diffusion process The target residual under Represents the residual between the low-frequency spectrum interpolation image with added noise and the target frequency-high image sequence or the residual between the second frequency-high image sequence and the target frequency-high image sequence, It represents the high-resolution spectrum-height map generation module, and its parameters are , Represents the time-frequency planning control signal output by the time series prediction module, , Represents the prediction function used by the time series prediction module, Represents the low-frequency image features after mask processing; When the curve of the loss function of the joint training of the time series prediction module and the high spectral resolution frequency height map generation module converges, the joint training of the time series prediction module and the high spectral resolution frequency height map generation module is completed.

Images