A method and system for automatically detecting equatorial ionospheric irregularities

Abstract

The application relates to an equatorial ionospheric irregularity automatic detection method and system, and belongs to the field of artificial intelligence. The method comprises the following steps: downloading GOLD satellite data, pre-processing and labeling, and constructing an original data set; inputting the original data set into a deep learning network model for training to obtain a trained deep learning network model; calculating the performance index of the trained deep learning network model, and outputting an equatorial ionospheric irregularity identification result. The application solves the problem of low accuracy in identifying and detecting equatorial ionospheric irregularities caused by the influence of factors such as slant range, background airglow intensity, instrument characteristics and threshold.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an automatic detection method and system for irregularities in the equatorial ionosphere. Background Technology

[0002] Equatorial ionospheric irregularities are primarily manifested as equatorial ionization anomalies and equatorial plasma bubbles. These are the most important ionospheric structures in low latitudes, and as key phenomena, they affect satellite communication, navigation performance, and other space-related activities. Accurate detection of these phenomena is crucial for space weather monitoring and early warning. The global-scale limb and disk observation satellite (GOLD), launched in 2018, provides a continuous and intuitive tool for monitoring equatorial ionization anomalies and equatorial plasma bubbles. Airglow images from the GOLD satellite can continuously track the diurnal variation of equatorial ionization anomalies and the formation and evolution of equatorial plasma bubbles, providing visualizations of the bimodal bright bands of equatorial ionization anomalies and the dark bands of equatorial plasma bubbles.

[0003] Existing technologies propose using a rolling barrel detrending technique to extract depletion regions associated with equatorial plasma bubbles from GOLD images. With deep learning demonstrating significant potential in Earth observation image analysis, its application to this data exploration is highly effective. However, current methods for automatically identifying and detecting equatorial ionospheric irregularities using deep learning on the GOLD satellite face several challenges: first, the observation results are affected by slant range, background airglow intensity, and instrument characteristics, including spatial resolution, scan period, and sensor noise, impacting the visibility of depletion regions; second, thresholding methods are difficult to adapt to different seasons, weather conditions, or geomagnetic storm conditions. Summary of the Invention

[0004] This application provides an automatic detection method and system for equatorial ionospheric irregularities, which can solve the problem of low accuracy in identifying and detecting equatorial ionospheric irregularities due to the influence of factors such as slant distance, background gas glow intensity, instrument characteristics and threshold.

[0005] To achieve the above objectives, according to a first aspect of this application, an automatic detection method for equatorial ionospheric irregularities is provided, the method comprising the following steps:

[0006] S1. Download GOLD satellite data and preprocess it to obtain preprocessed GOLD satellite data; S2. Label the preprocessed GOLD satellite data and construct the original dataset; S3. Construct a deep learning network model, input the original dataset into the deep learning network model for training, and evaluate it through preset performance indicators to finally obtain the trained deep learning network model. S4. Input the GOLD image into the trained deep learning network model, and the model outputs the automatic recognition result of the GOLD image after processing.

[0007] Optionally, S1 further includes the following steps: Download the raw data based on the changes in the satellite's temporal resolution; Remove the NAN values ​​from the original data and filter to obtain non-empty data; The non-empty data is divided into Southern Hemisphere channel data and Northern Hemisphere channel data. Based on wavelength threshold analysis, the wavelength data that is closest to the wavelength threshold in the Southern Hemisphere channel data and Northern Hemisphere channel data is found to obtain the Southern Hemisphere wavelength data and Northern Hemisphere wavelength data.

[0008] Optionally, S1 further includes the following steps: Select the scanning time points of the Southern Hemisphere and Northern Hemisphere channel data, and normalize the radiance images of the Southern Hemisphere and Northern Hemisphere channel data using the following formula, then remove invalid and outlier values ​​to obtain the normalized Southern Hemisphere data and Northern Hemisphere normalized data:

[0009] in, The input image is after normalization. This is the original radiance image. The image with minimum radiance. Image of maximum radiance; A latitude and longitude grid is created based on normalized data from the Southern Hemisphere and the Northern Hemisphere. The normalized data from the Southern Hemisphere and the Northern Hemisphere are then gridded using numerical transformation based on the latitude and longitude grid. Missing values ​​in the normalized data from the Southern Hemisphere and the Northern Hemisphere are then filled in using interpolation to obtain a stitched image of the target area.

[0010] Optionally, S2 further includes the following steps: using the Labelme tool to label the background, equatorial plasma bubble, and equatorial ionization anomaly of the stitched image of the target region, fixing the label categories and generating a MASK mask, and finally constructing the original dataset.

[0011] Furthermore, the deep learning network model includes a feature encoding module, a feature fusion module, and a feature decoding module; the feature encoding module receives the original dataset and outputs multi-scale features of the irregular body; the feature fusion module receives the multi-scale features of the irregular body and outputs multi-scale fused features; the feature decoding module receives the multi-scale fused features and outputs a distribution map of the equatorial ionospheric irregular body. The feature encoding module adopts a CNN-BiLSTM hybrid network structure, which consists of a feature input layer, four cascaded multi-scale convolutional blocks, two stacked bidirectional LSTM layers, and a feature dimensionality reduction layer. The feature fusion module adopts a cross-scale attention fusion structure, which consists of a feature alignment layer, four scale feature attention units, a cross-scale feature splicing layer, and a fusion feature enhancement layer. The feature decoding module adopts a U-shaped decoding structure with deconvolution upsampling, consisting of a fused feature input layer, three cascaded deconvolution decoding blocks, a skip connection feature fusion layer, a feature reconstruction layer, and a prediction output layer.

[0012] Optionally, step S3 further includes the following steps: The original dataset is input into the deep learning network model for training. In the feature encoding module, convolution operations are performed sequentially on the feature maps of the original dataset:

[0013] in, For the first Layer feature map, It is a non-linear activation function. For the first Layer feature map, For the first Layer convolution kernel weights, For the first Layer convolution kernel bias; extract multi-scale features of irregular volumes from all coding layers through feature encoding of several layers; In the feature fusion module, the multi-scale features of irregularities from all coding layers are fused to obtain multi-scale fused features. :

[0014] in, The number of feature maps participating in feature fusion. For the first Layer features, For the first Weight coefficients of layer features; In the feature decoding module, a deconvolution operation is used to restore the low-resolution 64×64 feature map to a high-resolution 512×512 feature map:

[0015] in, This is a transpose convolution operation. It is a deconvolution filter; Probability is calculated through upsampling:

[0016] in, For pixels The probability of belonging to an irregularity in the equatorial ionosphere. This is the activation function for the output layer.

[0017] Optionally, step S3 further includes the following steps: Deep learning network models are trained and optimized using a composite loss function:

[0018]

[0019]

[0020] in, Let cross-entropy be the loss function. For the first The actual label value of each pixel. For the model to the first The predicted probability value of each pixel. For the pixels participating in the loss function calculation, The Dice loss function, Sum all elements of the probability graph P to predict it. Summing all elements of the true label map G, For composite loss function, These are the weight coefficients of the cross-entropy loss function. These are the weighting coefficients of the Dice loss function.

[0021] Optionally, step S3 further includes the following steps: The network parameters are updated based on the backpropagation algorithm and gradient descent strategy:

[0022] in, For the first The set of network parameters at the next iteration For the first The set of network parameters at the next iteration For learning rate, This represents the gradient of the composite loss function with respect to the current network parameters. After each round of training, the model converges based on the change in the loss on the validation set. If the composite loss function is lower than the convergence threshold for a series of preset rounds, training is stopped, and a well-trained deep learning network model is obtained. The trained deep learning network model is used to automatically detect ionospheric irregularities. Cross-union ratio (CUB), Dice coefficient, recall, precision, and F1 score are used as performance metrics. The calculation of each performance metric is as follows:

[0023]

[0024]

[0025]

[0026]

[0027] in, For intersection, union, and comparison, This is the Dice coefficient. To achieve recall, To achieve precision, For F1 score, This is the set of pixels representing the equatorial ionospheric irregularities predicted by the model. This is a set of pixels representing real equatorial ionospheric irregularities obtained through manual annotation. To predict the number of pixels that are actually irregular shapes. This refers to the number of pixels that are predicted to form the background but are actually irregular shapes. To predict the number of pixels that are irregular shapes but actually form the background; When the crossover ratio, Dice coefficient, recall, precision, and F1 score all meet the preset evaluation criteria, the accuracy of the trained deep learning network model is considered to have reached the standard for automatic detection, and the equatorial ionospheric irregularity identification result is output, which is the distribution map of the equatorial ionospheric irregularity.

[0028] To achieve the above objectives, according to a second aspect of this application, an automatic detection system for equatorial ionospheric irregularities is also provided, comprising a data preprocessing module, a dataset creation module, a network training module, and an image recognition output module; the data preprocessing module downloads and preprocesses GOLD satellite data to obtain preprocessed GOLD satellite data; the dataset creation module annotates the preprocessed GOLD satellite data to construct an original dataset; the network training module constructs a deep learning network model, inputs the original dataset into the deep learning network model for training, and evaluates it through preset performance indicators to finally obtain a trained deep learning network model; the image recognition output module inputs GOLD images into the trained deep learning network model, and the model outputs the automatic recognition result of the GOLD image after processing.

[0029] To achieve the above objectives, according to a third aspect of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including an automatic detection method program for equatorial ionospheric irregularities, the automatic detection method program for equatorial ionospheric irregularities being executed by a processor to implement the steps of an automatic detection method for equatorial ionospheric irregularities as described in the first aspect.

[0030] In this application, the deep learning network model employs a cascaded structure of a feature encoding module, a feature fusion module, and a feature decoding module. The feature encoding module provides high-dimensional representation of temporal, spatial, and statistical features from the original observation data, effectively extracting key information related to equatorial ionospheric irregularities. The feature fusion module jointly models and adaptively weights features from different scales and dimensions, enhancing the model's ability to represent complex scenes. The feature decoding module, based on the fused features, achieves refined discrimination and output, thereby effectively mitigating the interference of factors such as slope range variations, background airglow intensity fluctuations, differences in instrument system characteristics, and manual threshold settings on the identification results. This enables rapid, automatic, and stable identification of equatorial ionospheric irregularities in large-scale observation data. Preprocessing, usability checks, filtering, and labeling of the original data effectively increase the processing efficiency of subsequent datasets, improving the accuracy of subsequent data rendering and significantly reducing the data loading process and complexity, which is beneficial for subsequent deep learning training on different categories. Performance is effectively evaluated by setting evaluation metrics. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0032] Figure 1 This is a flowchart of the automatic detection method for equatorial ionospheric irregularities provided in Embodiment 1 of this application; Figure 2 This is a flowchart of the dataset preprocessing and network training process according to Embodiment 1 of this application; Figure 3 This is a structural diagram of the automatic detection system for equatorial ionospheric irregularities provided in Embodiment 3 of this application; Figure 4 This is a schematic diagram illustrating the changes in the ionospheric irregularity provided in Embodiment 4 of this application; Figure 5 This is an image showing the automatic recognition result under the deep learning network model provided in Embodiment 4 of this application; Figure 6 This is a schematic diagram showing the changes in ionospheric irregularities across the entire region according to Embodiment 5 of this application; Figure 7 This is an image showing the automatic recognition result under the deep learning network model provided in Embodiment 5 of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Example 1 like Figure 1 As shown, this embodiment provides an automatic detection method for equatorial ionospheric irregularities, the method comprising the following steps: S1. Download GOLD satellite data and preprocess it to obtain preprocessed GOLD satellite data.

[0035] like Figure 2 As shown, raw data is downloaded according to the changes in satellite temporal resolution; in a specific embodiment, L1C-NI1 data is downloaded as raw data.

[0036] Remove the NaN values ​​from the original data to filter out non-empty data; in a specific embodiment, based on the REFERENCE_POINT_LON of the L1C-NI1 data... The maximum value is used as the longitude, REFERENCE_POINT_LAT The maximum value is used as the latitude, RADIANCE The sum of these values ​​is used as the radiation value.

[0037] The non-empty data is divided into Southern Hemisphere channel data and Northern Hemisphere channel data. Based on wavelength threshold analysis, the wavelength data closest to the wavelength threshold in both the Southern and Northern Hemisphere channel data is found, thus obtaining the Southern and Northern Hemisphere wavelength data. In a specific embodiment, the wavelength value closest to a specified wavelength of 135.6 nm is found in both the Southern and Northern Hemisphere channel data. The target wavelength is then analyzed using the CHOSEN_WAVELENGTH function to find the Southern and Northern Hemisphere wavelength data.

[0038] Select the scanning time points of the Southern Hemisphere and Northern Hemisphere channel data, and normalize the radiance images of the Southern Hemisphere and Northern Hemisphere channel data using the following formula, then remove invalid and outlier values ​​to obtain the normalized Southern Hemisphere data and Northern Hemisphere normalized data:

[0039] in, The input image is after normalization. This is the original radiance image. The image with minimum radiance. Image with maximum radiance.

[0040] A latitude and longitude grid is created based on normalized data from the Southern and Northern Hemispheres. The normalized data is then gridded using numerical transformation based on this grid. Missing values ​​in the Southern and Northern Hemisphere normalized data are filled in using interpolation, resulting in a stitched image of the target region. In one specific embodiment, a latitude and longitude grid is created using the `xarray.DataArray` function, the data is interpolated, and the stitched image of the target region is drawn using the `meshgrid` and `transform` functions.

[0041] S2. Label the preprocessed GOLD satellite data to construct the original dataset.

[0042] The background, equatorial plasma bubble, and equatorial ionization anomaly of the stitched image of the target region were labeled using the Labelme tool. The label categories were fixed and a MASK mask was generated to finally construct the original dataset.

[0043] S3. Construct a deep learning network model. Input the original dataset into the deep learning network model for training, and evaluate it through preset performance indicators to finally obtain the trained deep learning network model.

[0044] The original dataset is input into the deep learning network model for training. In the feature encoding module, convolution operations are performed sequentially on the feature maps of the original dataset:

[0045] in, For the first Layer feature map, It is a non-linear activation function. For the first Layer feature map, For the first Layer convolution kernel weights, For the first Layer convolution kernel bias; extract multi-scale features of irregular volumes from all coding layers through feature encoding of several layers; In the feature fusion module, the multi-scale features of irregularities from all coding layers are fused to obtain multi-scale fused features. :

[0046] in, The number of feature maps participating in feature fusion. For the first Layer features, For the first Weight coefficients of layer features; In the feature decoding module, a deconvolution operation is used to restore the low-resolution 64×64 feature map to a high-resolution 512×512 feature map:

[0047] in, This is a transpose convolution operation. It is a deconvolution filter; Probability is calculated through upsampling:

[0048] in, For pixels The probability of belonging to an irregularity in the equatorial ionosphere. This is the activation function for the output layer.

[0049] Deep learning network models are trained and optimized using a composite loss function:

[0050]

[0051]

[0052] in, Let cross-entropy be the loss function. For the first The actual label value of each pixel. For the model to the first The predicted probability value of each pixel. For the pixels participating in the loss function calculation, The Dice loss function, Sum all elements of the probability graph P to predict it. Summing all elements of the true label map G, For composite loss function, These are the weight coefficients of the cross-entropy loss function. These are the weighting coefficients of the Dice loss function.

[0053] The network parameters are updated based on the backpropagation algorithm and gradient descent strategy:

[0054] in, For the first The set of network parameters at the next iteration For the first The set of network parameters at the next iteration For learning rate, This represents the gradient of the composite loss function with respect to the current network parameters. After each round of training, the model converges based on the changes in the loss on the validation set. If the composite loss function for consecutive preset rounds is lower than the convergence threshold, training is stopped, and a well-trained deep learning network model is obtained.

[0055] The trained deep learning network model is used to automatically detect ionospheric irregularities. Cross-union ratio (CUB), Dice coefficient, recall, precision, and F1 score are used as performance metrics. The calculation of each performance metric is as follows:

[0056]

[0057]

[0058]

[0059]

[0060] in, For intersection, union, and comparison, This is the Dice coefficient. To achieve recall, To achieve precision, For F1 score, This is the set of pixels representing the equatorial ionospheric irregularities predicted by the model. This is a set of pixels representing real equatorial ionospheric irregularities obtained through manual annotation. To predict the number of pixels that are actually irregular shapes. This refers to the number of pixels that are predicted to form the background but are actually irregular shapes. To predict the number of pixels that are irregular shapes but actually form the background; When the Intersection over Union (IoU), Dice coefficient, recall, precision, and F1 score simultaneously meet the preset evaluation criteria, the accuracy of the trained deep learning network model is considered to have reached the standard for automatic detection, and the equatorial ionospheric irregularity identification result is output, which is an equatorial ionospheric irregularity distribution map. In a specific embodiment, IoU ≥ 0.75, Dice coefficient ≥ 0.85, recall ≥ 0.88, precision ≥ 0.85, and F1 score ≥ 0.86 are used as the evaluation criteria set for the deep learning network model. A recall ≥ 0.88 can effectively reduce missed detections of irregularities, and a precision ≥ 0.85 can reduce misclassification of background as irregularities. These indicators are used to evaluate whether the trained deep learning network model meets the accuracy requirements for automatic detection.

[0061] S4. Input the GOLD image into the trained deep learning network model, and the model outputs the automatic recognition result of the GOLD image after processing.

[0062] Example 2 This embodiment is based on embodiment 1.

[0063] The deep learning network model includes a feature encoding module, a feature fusion module, and a feature decoding module. The feature encoding module receives the original dataset and outputs multi-scale features of the irregular body. The feature fusion module receives the multi-scale features of the irregular body and outputs multi-scale fused features. The feature decoding module receives the multi-scale fused features and outputs a distribution map of the equatorial ionospheric irregular body.

[0064] Furthermore, the feature encoding module adopts a CNN-BiLSTM hybrid network structure, consisting of a feature input layer, four cascaded multi-scale convolutional blocks, two stacked bidirectional LSTM layers, and a feature dimensionality reduction layer. In a specific embodiment, the bidirectional LSTM structure captures the spatiotemporal evolution features of ionospheric irregularities, and a Dropout layer (dropout=0.2) is used to suppress overfitting.

[0065] The feature fusion module employs a cross-scale attention fusion structure (CSAF), consisting of a feature alignment layer, four scale feature attention units, a cross-scale feature concatenation layer, and a fusion feature enhancement layer. In a specific embodiment, the fusion feature enhancement layer comprises a 1×1 convolutional layer, a batch normalization layer, and a GELU activation function, which performs channel fusion and information enhancement on the concatenated features.

[0066] The feature decoding module adopts a U-shaped decoding structure with deconvolution upsampling, consisting of a fused feature input layer, three cascaded deconvolution decoding blocks, a skip-connected feature fusion layer, a feature reconstruction layer, and a prediction output layer. In a specific embodiment, the deconvolution decoding block consists of: a transposed convolutional layer (deconvolution), a batch normalization layer, a LeakyReLU activation function, and a convolutional thinning layer; the transposed convolutional layer has a stride of 2 to perform feature upsampling, and the convolutional thinning layer uses... Convolutional kernels perform detailed optimizations on the features after dimensionality increase.

[0067] Example 3 like Figure 3 As shown, this embodiment provides an automatic detection system for equatorial ionospheric irregularities, including a data preprocessing module, a dataset creation module, a network training module, and an image recognition output module. The data preprocessing module downloads and preprocesses GOLD satellite data to obtain preprocessed GOLD satellite data. The dataset creation module annotates the preprocessed GOLD satellite data to construct an original dataset. The network training module constructs a deep learning network model, inputs the original dataset into the deep learning network model for training, and evaluates it through preset performance indicators to finally obtain a trained deep learning network model. The image recognition output module inputs GOLD images into the trained deep learning network model, and the model outputs the automatic recognition result of the GOLD images after processing.

[0068] Example 4 This embodiment provides a computer-readable storage medium including a program for an automatic detection method of equatorial ionospheric irregularities. When the program is executed by a processor, it implements the steps of an automatic detection method for equatorial ionospheric irregularities as described in Embodiment 1.

[0069] Example 5 like Figure 4 As shown, this embodiment uses GOLD satellite data from a specific year for training. The GOLD satellite data in this dataset is provided by the GOLD website. The main data includes RADIANCE data (one frame every 15-30 minutes) scanned at night across the northern and southern hemispheres of the region, generated by a deep learning network model. The northern and southern hemisphere data from the GOLD satellite are merged together for rendering, enabling complete stitching of the scanned data and large-scale observation, allowing observation of changes in ionospheric irregularities across the entire region. The figure directly demonstrates that the LOSS value continuously decreases and stabilizes over a certain period under deep learning training, illustrating the feasibility of using deep learning training for automatically detecting and identifying equatorial ionospheric irregularities.

[0070] To verify the accuracy of the deep learning-based automatic identification and detection method for equatorial ionospheric irregularities on the GOLD satellite, three images were randomly selected as input for identification and detection. The pre-constructed deep learning model then performed automatic identification and detection. Figure 5 The image shows the automatic recognition results under the deep learning network model, and the original image is preserved for comparison. It can be seen that the automatic detection and recognition of irregularities in the equatorial ionosphere has been successfully achieved.

[0071] Example 6 like Figure 6 As shown, this embodiment uses GOLD data from a certain year to create a dataset for training. The GOLD satellite data in this dataset is provided by the GOLD official website. The main data includes data from nighttime scans of the Northern and Southern Hemispheres of the region. The data (one frame every 15-30 minutes) was generated by a deep learning network model. Data from the Northern and Southern Hemispheres of the GOLD satellite were merged and plotted together, enabling complete stitching of scanned data and large-scale observation, allowing observation of changes in ionospheric irregularities across the entire region. The figure directly shows that the LOSS value continuously decreases and stabilizes over a certain period under deep learning training, demonstrating that the performance of deep learning training is feasible for automatically detecting and identifying equatorial ionospheric irregularities.

[0072] To verify the accuracy of the deep learning-based automatic identification and detection method for equatorial ionospheric irregularities on the GOLD satellite, three images were randomly selected as input for identification and detection. The pre-constructed deep learning model then performed automatic identification and detection. Figure 7 The image shows the automatic recognition results under the deep learning network model, and the original image is preserved for comparison. It can be seen that the automatic detection and recognition of irregularities in the equatorial ionosphere has been successfully achieved.

[0073] The technical solution of this application has the following advantages or beneficial effects: The deep learning network model employs a cascaded structure of feature encoding, feature fusion, and feature decoding modules. The feature encoding module provides high-dimensional representation of temporal, spatial, and statistical features from the original observation data, effectively extracting key information related to equatorial ionospheric irregularities. The feature fusion module jointly models and adaptively weights features from different scales and dimensions, enhancing the model's ability to represent complex scenes. The feature decoding module, based on the fused features, achieves refined discrimination and output, thereby effectively mitigating the interference of factors such as slope range variations, background airglow intensity fluctuations, differences in instrument system characteristics, and manual threshold settings on the identification results. This enables rapid, automatic, and stable identification of equatorial ionospheric irregularities in large-scale observation data. Systematic data preprocessing of the raw data before model training, along with data availability checks, effective data filtering, and precise labeling, constructs high-quality training and validation datasets. This not only significantly improves the efficiency of subsequent dataset construction and retrieval but also effectively reduces redundancy and complexity during data loading, providing a reliable guarantee for the accuracy of subsequent rendering and analysis results. Simultaneously, it facilitates the deep learning model's ability to fully learn different irregular body features under multi-class sample conditions, enhancing the model's generalization ability and robustness. By setting up a multi-dimensional performance evaluation system that includes recognition accuracy, recall, precision, and comprehensive evaluation indicators, the recognition performance of the model under different data scales and observation conditions is comprehensively and quantitatively evaluated, thereby objectively verifying the significant advantages of the method in automatic recognition efficiency, recognition accuracy, and stability.

[0074] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0075] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0076] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An automatic detection method for irregularities in the equatorial ionosphere, characterized in that, The method includes the following steps: S1. Acquire GOLD satellite data and preprocess it to obtain preprocessed GOLD satellite data; S2. Label the preprocessed GOLD satellite data and construct the original dataset; S3. Input the original dataset into the deep learning network model for training, and evaluate it through preset performance indicators to finally obtain the trained deep learning network model. S4. Input the GOLD image into the trained deep learning network model, and the model outputs the automatic recognition result of the GOLD image after processing.

2. The automatic detection method for equatorial ionospheric irregularities according to claim 1, characterized in that, S1 further includes the following steps: Download the raw data based on the changes in the satellite's temporal resolution; Remove the NAN values ​​from the original data and filter to obtain non-empty data; The non-empty data is divided into Southern Hemisphere channel data and Northern Hemisphere channel data. Based on wavelength threshold analysis, the wavelength data that is closest to the wavelength threshold in the Southern Hemisphere channel data and Northern Hemisphere channel data is found to obtain the Southern Hemisphere wavelength data and Northern Hemisphere wavelength data.

3. The automatic detection method for equatorial ionospheric irregularities according to claim 2, characterized in that, S1 further includes the following steps: Select the scanning time points of the Southern Hemisphere and Northern Hemisphere channel data, and normalize the radiance images of the Southern Hemisphere and Northern Hemisphere channel data using the following formula, then remove invalid and outlier values ​​to obtain the normalized Southern Hemisphere data and Northern Hemisphere normalized data: in, The input image is after normalization. This is the original radiance image. The image with minimum radiance. Image of maximum radiance; A latitude and longitude grid is created based on normalized data from the Southern Hemisphere and the Northern Hemisphere. The normalized data from the Southern Hemisphere and the Northern Hemisphere are then gridded using numerical transformation based on the latitude and longitude grid. Missing values ​​in the normalized data from the Southern Hemisphere and the Northern Hemisphere are then filled in using interpolation to obtain a stitched image of the target area.

4. The automatic detection method for equatorial ionospheric irregularities according to claim 1, characterized in that, The S2 also includes the following steps: using the Labelme tool to label the background, equatorial plasma bubble and equatorial ionization anomaly of the target region stitched image, fixing the label categories and generating a MASK mask, and finally constructing the original dataset.

5. The automatic detection method for equatorial ionospheric irregularities according to claim 1, characterized in that, The deep learning network model includes a feature encoding module, a feature fusion module, and a feature decoding module; the feature encoding module receives the original dataset and outputs multi-scale features of the irregular body; the features... The fusion module receives multi-scale features of an irregular body and outputs multi-scale fused features. The feature decoding module receives multi-scale fused features and outputs a distribution map of irregularities in the equatorial ionosphere. The feature encoding module adopts a CNN-BiLSTM hybrid network structure, which consists of a feature input layer, four cascaded multi-scale convolutional blocks, two stacked bidirectional LSTM layers, and a feature dimensionality reduction layer. The feature fusion module adopts a cross-scale attention fusion structure, which consists of a feature alignment layer, four scale feature attention units, a cross-scale feature splicing layer, and a fusion feature enhancement layer. The feature decoding module adopts a U-shaped decoding structure with deconvolution upsampling, consisting of a fused feature input layer, three cascaded deconvolution decoding blocks, a skip connection feature fusion layer, a feature reconstruction layer, and a prediction output layer.

6. The automatic detection method for equatorial ionospheric irregularities according to claim 1, characterized in that, S3 also includes the following steps: The original dataset is input into the deep learning network model for training. In the feature encoding module, convolution operations are performed sequentially on the feature maps of the original dataset: in, For the first Layer feature map, It is a non-linear activation function. For the first Layer feature map, For the first Layer convolution kernel weights, For the first Layer convolution kernel bias; extract multi-scale features of irregular volumes from all coding layers through feature encoding of several layers; In the feature fusion module, the multi-scale features of irregularities from all coding layers are fused to obtain multi-scale fused features. : in, The number of feature maps participating in feature fusion. For the first Layer features, For the first Weight coefficients of layer features; In the feature decoding module, a deconvolution operation is used to restore the low-resolution 64×64 feature map to a high-resolution 512×512 feature map: in, This is a transpose convolution operation. It is a deconvolution filter; Probability is calculated through upsampling: in, For pixels The probability of belonging to an irregularity in the equatorial ionosphere. This is the activation function for the output layer.

7. The automatic detection method for equatorial ionospheric irregularities according to claim 6, characterized in that, S3 also includes the following steps: Deep learning network models are trained and optimized using a composite loss function: in, Let cross-entropy be the loss function. For the first The actual label value of each pixel. For the model to the first The predicted probability value of each pixel. For the pixels participating in the loss function calculation, The Dice loss function, Sum all elements of the probability graph P to predict it. Summing all elements of the true label map G, For composite loss function, These are the weight coefficients of the cross-entropy loss function. These are the weighting coefficients of the Dice loss function.

8. The automatic detection method for equatorial ionospheric irregularities according to claim 7, characterized in that, S3 also includes the following steps: The network parameters are updated based on the backpropagation algorithm and gradient descent strategy: in, For the first The set of network parameters at the next iteration For the first The set of network parameters at the next iteration For learning rate, This represents the gradient of the composite loss function with respect to the current network parameters. After each round of training, the model converges based on the change in the loss on the validation set. If the composite loss function is lower than the convergence threshold for a series of preset rounds, training is stopped, and a well-trained deep learning network model is obtained. The trained deep learning network model is used to automatically detect ionospheric irregularities. Cross-union ratio (CUB), Dice coefficient, recall, precision, and F1 score are used as performance metrics. The calculation of each performance metric is as follows: in, For intersection, union, and comparison, This is the Dice coefficient. To achieve recall, To achieve precision, For F1 score, This is the set of pixels representing the equatorial ionospheric irregularities predicted by the model. This is a set of pixels representing real equatorial ionospheric irregularities obtained through manual annotation. To predict the number of pixels that are actually irregular shapes. This refers to the number of pixels that are predicted to form the background but are actually irregular shapes. To predict the number of pixels that are irregular shapes but actually form the background; When the crossover ratio, Dice coefficient, recall, precision, and F1 score all meet the preset evaluation criteria, the accuracy of the trained deep learning network model is considered to have reached the standard for automatic detection, and the equatorial ionospheric irregularity identification result is output, which is the distribution map of the equatorial ionospheric irregularity.

9. An automatic detection system for equatorial ionospheric irregularities, characterized in that, It includes a data preprocessing module, a dataset creation module, a network training module, and an image recognition output module; The data preprocessing module downloads and preprocesses GOLD satellite data to obtain preprocessed GOLD satellite data. The dataset creation module annotates the preprocessed GOLD satellite data to construct the original dataset; The network training module constructs a deep learning network model, inputs the original dataset into the deep learning network model for training, and evaluates it through preset performance metrics, ultimately obtaining a trained deep learning network model. The image recognition output module inputs the GOLD image into a trained deep learning network model, which then processes the image and outputs the automatic recognition result of the GOLD image.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes an automatic detection method program for equatorial ionospheric irregularities. When the automatic detection method program for equatorial ionospheric irregularities is executed by a processor, it implements the steps of an automatic detection method for equatorial ionospheric irregularities as described in any one of claims 1 to 8.

Images